technicalmarketing.ai Free AI audit

Paper · November 2026

Technical marketing for professional practices: how dental, medical, legal and accounting firms are found when AI answers.

102,609 words · about 7.4 hours · 14 chapters · 346 sources

Abstract

Dental practices, medical clinics, law firms and accounting firms sell services that a buyer cannot judge in advance, under rules that limit what the seller may say, through firms too small to keep a marketing department. The way those firms are found is changing from a ranked list a person scans to an answer a machine writes.

This paper argues that the discipline suited to that change is technical marketing: treating how a practice is found, chosen and reached, the rules it works within, and the measurement of all three as one engineered system. It draws on the economics of information and of services, on the professions' own rules, and on our study of what three AI assistants answered when asked to recommend local businesses. In that study the assistants named 2,975 businesses across 199 questions, and only 4.1% were named by all three, or 3.5% for the ten kinds of business asked about in every city1. Each question was asked once, through each assistant's API1.

The paper sets out the theory, presents the data, applies the framework to each of the four professions, and ends with a compliance architecture, a measurement plan, a first-year program and the limits of what we have shown. It describes rules and research in general terms and gives no legal, medical or tax advice.

What this paper argues

  1. Choosing a professional is a problem of verification under uncertainty, so marketing for a profession is the production and upkeep of signals that a buyer, or an assistant acting for one, can check.
  2. AI assistants do not agree on whom to recommend. In our study 79.7% of the businesses named were named by only one of three assistants1, and the sites cited most were directories, review platforms and list sites, not local businesses' own sites1.
  3. We argue that first contact is where a practice that has been found and chosen can still lose the patient or client, and it is where the professions' rules bind hardest.
  4. The four professions' rules reduce to a short list of design requirements that one system can meet once.
  5. A small practice can measure enough to manage its marketing if it counts outcomes and accepts the limits of small numbers.

How to read this paper

~
Our estimate, built from the nearest cited figures. Not a measurement.
†
A vendor’s own figure, from its customers or its own survey. Not an independent sample.
NF
Not found: we looked for a published measure and found none.
( )
A figure followed by its working in brackets, or by “calc.”, is our arithmetic on cited figures.
Rules
We report what each rule says as read on the date shown. This is not legal advice, and counsel decides how a rule applies to you.
Sources
A small number points to the source in References, which says where we hold a source only through its abstract or a secondary account.

Chapter 1

The problem: professional services in an age of answers

Three assistants, three different lists of professionals

In our study, ChatGPT, Gemini and Perplexity were asked the same local questions and rarely named the same business. For the 160 questions asked in all 16 cities, which cover ten kinds of business, the three assistants named 2,414 businesses, and 84 of them (3.5%) were named by all three1. Pooled across all 199 questions, which include 13 further kinds of business asked about in three cities each, the figure is 4.1%: 122 of the 2,975 businesses named1. The professions this paper studies sit on both sides of those averages. Of the 266 dentists the assistants named across the 16 cities, 3 (1.1%) were named by all three1. Of 272 accounting businesses, none was1. Of 240 family law attorneys, 13 (5.4%) were1.

A business counts as named by all three assistants when each of them, asked the same question, named it among the up to eight names we read from its answer; names were matched across assistants by spelling1. Each question was asked once, on 2026-11-10, through each assistant's programming interface with web search on, and answers collected that way are a baseline, not what every signed-in user sees1. Chapter 4 presents the design and argues, as an argument and not a result, that these choices tend to lower measured agreement, which would make 3.5% closer to a floor than to a central estimate. Our study did not measure how far an assistant disagrees with itself, and Chapter 4 sets that limit beside an outside study that asked the same prompts repeatedly.

The number does not show that any assistant is wrong, that a named business is good, or that a cited page caused a name to appear. Our study did not ask about physicians: the nearest health rows are chiropractors, physical therapists and med spas, and we do not stretch them to cover medical practices. And three assistants can differ because each selects, correctly by its own sources, from a large pool of competent practices, because their sources differ, or because each varies from run to run. The study cannot tell these apart, so low agreement describes the channel and says nothing about the quality of the practices in it.

The number changes the question a practice owner asks. A printed directory was one list, and a search result page, though it varies by searcher and place, can be checked by typing a query into one search box, so the owner's question was whether the practice appeared and where. When three assistants read different sources and write different lists, that question has no single answer. Of the websites the assistants cited, 88% were cited by only one of them1, so the readers differ in what they read as well as in what they conclude. The question becomes whether a patient, a client, or the assistant acting for either can find the practice, verify its facts and trust what it finds, whichever reader they use.

Why dental, medical, law firm and CPA marketing share one problem

Four features set a professional practice apart from a business that sells a product. None is unique to the professions, but we think that together they make the marketing problem different in kind. The paper treats them as premises.

The buyer cannot judge the quality

A patient choosing a dentist, or a client choosing a family law attorney, can seldom judge the quality of the work before buying it and often cannot judge it afterward. Darby and Karni (1973) named the property. They distinguished search qualities, which are known before purchase, experience qualities, which are known costlessly only after purchase, and credence qualities, which in their words are "expensive to judge even after purchase"2. Day and Barksdale (1992) made the point for professional services, which they described as "typically intangible and customized", so that "evaluation prior to purchase is difficult"3.

Buyers who cannot judge quality look for substitutes: a credential, a review, a referral, how quickly someone answers the telephone. These are signals, which can be checked more or less cheaply. In a 2012 national survey, among respondents who had sought online physician ratings in the past year, 35% said they had selected a physician on the strength of good ratings and 37% said they had avoided one with bad ratings4. The survey concerns physicians, is over a decade old, and describes a subset of respondents, so we cite it to show that substitutes are used, not to size their use today. Chapter 2 develops the economics of search, credence goods, asymmetric information and signaling. Our point from it is that when quality cannot be seen, marketing is mainly the production and upkeep of signals that a buyer, or an assistant reading for the buyer, can check.

Speech is regulated

Each of the four professions limits what a practice may say about itself, and the limits come from several kinds of source. In Bates v. State Bar of Arizona (1977), the Supreme Court held that a state could not prevent a lawyer from publishing a truthful advertisement of the availability and terms of routine legal services5. It added that advertising that is false, deceptive or misleading "of course is subject to restraint"5. ABA Model Rule 7.1 says a communication is false or misleading if it contains a material misrepresentation of fact or law, or omits a fact needed to keep the statement as a whole from being materially misleading6.

The ADA Principles of Ethics say that no dentist shall advertise in a manner that is false or misleading in any material respect7, and the AICPA Code bars a member in public practice from seeking clients by advertising that is false, misleading or deceptive8. The AMA's ethics opinion on advertising says there are no restrictions on physician advertising except those that can be specifically justified to protect the public from deceptive practices9.

These texts differ in force. The ABA Model Rules are models, and what binds a lawyer in a given state is for that lawyer's own counsel to say. The ADA, AMA and AICPA texts are professional ethics codes, not statutes. Health care adds a rule of a different kind. The HIPAA Privacy Rule defines marketing, in general, as a communication about a product or service that encourages recipients to buy or use it, subject to exceptions the definition lists10.

HHS's guidance gives the consequence in plain words: generally, a communication that is marketing can occur only if the covered entity first obtains the individual's authorization11. Such rules can reach past the words on a page to the tools a practice uses. HHS's online tracking bulletin gives an example: a patient makes an appointment through a covered health clinic's website, and the site's third-party tracking technologies send the appointment and the IP address to a vendor12. The bulletin says the vendor is then a business associate and a business associate agreement is required (a term defined later in this chapter)12. Chapters 6 to 9 set out each profession's rules, and Chapter 10 reduces them to design requirements. This paper describes the rules and does not say whether a given page, message or tool meets them.

The firm is small

The practices this paper addresses are small or mid-sized: an owner or a few partners make the decisions, and there is no marketing department. For dentistry, the ADA Health Policy Institute reports that as of 2025 one in three U.S. dentists works in a single location with no other dentist, 39% work in a single-location practice with at least one other dentist, and 12% work in group practices with 100 or more locations13. Roughly 7 in 10 dentists work at a single location (one in three + 39%). The shares are of dentists, not of practices.

For physicians, the AMA's 2024 Physician Practice Benchmark Survey found 47.4% of physicians in practices of 10 or fewer physicians, against 61.4% in 201214. The AMA says that in 2024 it switched the survey's vendor and expanded the sample from 3,500 to 5,000 physicians14, so the comparison with 2012 carries a caveat. For law and accounting, the nearest count is of offices, not firms. The Bureau of Labor Statistics counts private-sector establishments by their number of employees, and Table 1 gives its first-quarter 2024 counts for all four professions.

Offices ofAll officesFewer than 5 employees5 to 9 employeesSource
Lawyers187,335141,51823,98915,16,17
Certified public accountants62,82844,8009,33316,17,18
Dentists135,23354,22546,16816,17,19
Physicians237,573138,19739,89616,17,20

Table 1. Private-sector establishments by number of employees, United States, first quarter 2024, from the Bureau of Labor Statistics.

The table counts establishments, meaning single locations, not firms. A firm with several offices appears once for each office, and an owner with no employees does not appear at all. We tried the Census Bureau's tables by firm size and the American Bar Association's count of solo lawyers, and neither could be read, so the firm-level question is open. Employees are not professionals either: an office of nine may be several lawyers and their staff, or one lawyer and eight others. Fewer than 10 employees describes 88.3% of offices of lawyers, 86.2% of CPA offices, 75.0% of physician offices and 74.2% of dentist offices, and fewer than 5 describes 75.5%, 71.3%, 58.2% and 40.1% of them (calc., Table 1).

The premise is therefore supported at the level of the office in all four professions. Counted by physician, medicine is the weaker case, since fewer than half of physicians now work in practices of 10 or fewer physicians; the two measures count different things, physicians in a practice and employees at an office, so we do not set them side by side. The premise is also moving: in 2023, 72.5% of U.S. dentists were private practice owners, down from 84.7% in 200521. In 2024, 42.2% of physicians were in a practice wholly owned by physicians, 18 percentage points fewer than in 201214.

We state the premise about the firms the paper addresses, not about every firm in these professions. We infer that a firm of this size seldom has anyone whose job is marketing, from its headcount; this paper does not measure it. The consequence is that whatever a small firm builds, it must also operate, so a design it cannot run is worth little.

Demand arrives as a first contact

Most professional work begins with a first contact: a telephone call, a form, a booking request or a chat. We argue that this contact is where marketing's output is either converted or lost, and that it is where several of the rules bind hardest, because a person describing a health condition or a legal problem is already sharing information. In a 2025 survey by AMN Healthcare, research specialists called 1,391 physician offices in 15 large metro areas to schedule a new-patient appointment22. The average wait across six specialties was 31 days22. The survey is one company's report and covers large metros only.

A secret-shopper study of 500 law firms in 2024 by the vendor Clio† reported that 40% answered a telephone call, down from 56% in 2019, and that 33% responded to an emailed inquiry23. It placed one call to each firm, and we treat the figures as indicative of that test, not as a rate for all law firms.

For dental practices, the figures come from vendors who sell phone services and from test calls. In the data of Patient Prism† (phone services), covering 11,552,668 patient calls in 2025 to 8,280 dental locations, 31 of every 100 calls hung up before reaching an agent24. In a 2015 trade column, the head of a dental practice-training company reported that 35% of the company's test calls to dental practices in four countries went straight to voicemail during regular business hours25. For health care as a whole, Invoca† (call tracking) reports that 54% of inbound calls across six health care fields were answered by a person26. These count different things: a hang-up before an agent, a call that went to voicemail, a call not answered by a person. We therefore do not average them.

A measurement of missed calls at accounting firms is NF; the nearest evidence is the law firm test and the health care data above. In the law firm test, 60 of every 100 firms did not answer the call (100 less the 40% that did), and in Invoca†'s health care data 46 of every 100 calls were not answered by a person (100 less the 54% that were). The two measure different things, so the estimate is a rough bracket: ~4 to 6 of every 10 first calls to a small accounting firm are not answered by a person (from those two shares; likely high for a firm with a full-time receptionist). Lead conversion at accounting firms is NF; the nearest is health care, where Invoca† reports that 43% of calls answered by a person were leads and 45% of those leads converted on the call26.

People also search for tools for the first contact. "Medical answering service" averages 5,400 searches a month in the United States27, though Chapter 5 explains why search volume shows interest, not purchase.

The four features act on one another. Because quality cannot be judged, buyers lean on signals. Because speech is regulated, a practice cannot simply claim quality and must show it in ways its rules permit. Because the firm is small, the people who build the signals also answer the telephone and see patients. And because demand arrives as a first contact, the last step of any marketing system is a person or a machine speaking with someone who may be unwell, in legal trouble or worried about taxes. We think a plan that treats these as separate problems leaves the interactions to chance, which is the case for treating them as one problem.

From lists to answers: AI search and what changes when an assistant writes the recommendation

In the list era, position carried much of the weight. Ghose, Goldfarb and Han found that ranking effects were higher on mobile phones, where links at the top of the screen were especially likely to be clicked28. A search marketing vendor's analysis of Google's organic results reported that the first result averaged a click-through rate of 27.6%, ten times as likely to be clicked as the tenth29. A practice's task was to earn a high position on a page that the buyer then scanned, comparing listings and judging the signals alone.

An answer changes the structure: the assistant reads sources, chooses among them and writes a recommendation, often with citations. In a Pew Research Center study of participants' Google searches, 18% generated an AI summary, and 88% of those summaries cited three or more sources30. In our study the 597 answers cited 1,455 different websites, for 3,530 citations in all, or 5.9 per answer (3,530 / 597)1. The buyer who once read the listing, the reviews and the practice's own site may now meet them already condensed by a machine, and the machine decides which signals count. We think that describes the mechanism, as Chapter 2 argues; the figures show only that answers appear in search and are built from several sources.

Two bodies of work address what a source can do about it. The paper that introduced generative engine optimization, the label for methods that raise a source's visibility in generative engine responses, reports that its methods can boost visibility by up to 40%31. It also reports that the efficacy of its strategies varies across domains31 and that content creators have "little to no control over" how generative engines display their content31.

Google says that a page must be indexed and eligible to appear in Search with a snippet to be a supporting link in AI Overviews or AI Mode, that there are "no additional technical requirements"32. It adds that existing SEO fundamentals "continue to be worthwhile"32. The two do not conflict: one reports gains from methods tested in a research setting, and the other describes what Google's own AI features require. Our study tested neither: it asked three assistants a local question and counted what they named, and it did not query Google's AI features. How much a small professional practice gains from anything beyond sound fundamentals is, we think, an open question.

How many people use assistants this way depends on the instrument. Chatterji and co-authors report that by July 2025 ChatGPT's consumer product had been adopted by around 10% of the world's adult population33. They also report that the topics "Practical Guidance," "Seeking Information," and "Writing" together account for nearly 80% of ChatGPT conversations33. That counts users of one product worldwide and sorts messages into broad topics, so it does not isolate people choosing a dentist or a lawyer. A Pew Research Center survey of 5,119 US adults in February 2026 found that 49% ever use AI chatbots and that 42% ever use them to search for information34. "Ever" does not say how often, and "search for information" is Pew's wording for a use that is not limited to finding a professional.

A 2025 consumer survey by the local search vendor BrightLocal† reported that 6% of consumers used ChatGPT or other AI tools to find local reviews, down from 9% in 202435. Its 2026 survey of 1,002 US adults reports that use of ChatGPT and other generative AI tools for local recommendations rose from 6% in 2025 to 45%36. These are two different tasks, finding reviews and getting recommendations, in separately drawn panels. The 2026 report does not print the question, so a rise of more than seven times in a year may partly reflect a change in wording. Both count stated use, not calls to a practice, so we do not combine the figures into a trend.

The question that matters most to a practice is what happens to visits when an answer appears. A clickstream analysis by SparkToro† and Datos, reported by a vendor, found that 58.5% of American Google searches in 2024 resulted in zero clicks37. The figure covers all searches, not only those with an AI summary, and a search without a click is not necessarily a failed one, so it does not tell us what AI summaries change.

Pew Research Center measured that directly. In its analysis of American adults' browsing data, users who met an AI summary clicked a traditional search result link in 8% of visits, against 15% of visits without a summary30. They also ended the browsing session after 26% of search pages with an AI summary, against 16% of pages with only traditional results30. The dataset held 68,879 unique Google searches, of which 12,593 produced an AI summary30. The rates are for the dataset's searches taken together, which are general browsing, so we take Pew's result as the direction of the effect, not as a rate for any practice.

The nearest figure for the professions comes from a vendor. Whitespark† hand-ran 540 queries in Houston, Phoenix and Denver across six industries, including dentists, medical clinics and personal injury lawyers, and reported AI Overviews on 15% of local-intent queries against 92% of informational-intent queries38. Its authors chose the queries and the snapshot is from May 2025, so we use it only for the contrast between kinds of query.

Practices are also adopting AI themselves, and the figures again depend on the instrument. The Census Bureau's Business Trends and Outlook Survey put overall business AI usage between 17% and 20% from December 2025 to May 202639. In a mid-2026 panel of responding dentists, the ADA Health Policy Institute found 43.3% using AI for at least one type of task40. In the ABA's 2024 Legal Technology Survey, 30.2% of attorneys, among 512 respondents answering the online research volume, said their offices were using AI-based technology tools41. So did 17.7% of solo practitioners41. These count businesses, responding dentists and surveyed attorneys, in different years and with different wording, and the professional surveys rest on self-report, so we do not set them side by side as a trend. The paper does not say that most practices have adopted AI.

The thesis that professionals are increasingly found through written answers therefore rests on indirect evidence, not on a measurement of how many patients or clients choose a practice this way. In our study the assistants answered local questions by naming 2,975 specific businesses, citing most often directories, review platforms, list sites, publishers and a forum1.

A measurement of AI referrals on the website of a dental, medical, legal or accounting practice is NF; the nearest evidence is the cross-industry shares that follow. Conductor† (search software) reports that AI referral traffic was 1.08% of all website traffic across the 10 industries it analyzed, on enterprise sites42. It was .64% for its Health Care industry43 and .48% for Financials44. Microsoft Clarity† found AI referrals under 1% of traffic on more than 1,200 publisher and news sites45. AI referrals bring a practice's website ~1% of its visits or less, in 2025 data (from Conductor†'s and Clarity†'s shares; likely low as a count of patients who choose through an assistant)42,45. An answer that names a practice and leads to a call with no click leaves no referral.

A Pew, KFF or Gallup figure for AI used to find legal information or local businesses is NF; the nearest numbers count stated use: Bain, a consulting firm, reports that of 1,500 consumers surveyed in September 2025, 56% said they mostly or always use search engines as their default search tool and 16% said they mostly or always use chatbots46. Chapter 14 takes this up as the first objection.

Technical marketing: a definition, a through-line and a working image

Our thesis is that professional services are largely credence goods, sold under rules about what the seller may say, by firms too small to carry a marketing department. The way they are found is moving, on the indirect evidence above, from a ranked list a person scans to an answer a machine writes. In that setting we argue for technical marketing over louder promotion, because the work that matters is making a practice findable, checkable, reachable and countable.

We define technical marketing as the design, construction and measurement of a firm's acquisition, intake, compliance and reporting as a single system, in which each stage has an owner, an input, an output and a count, so that a change at one stage can be traced to its effect on the next. Its parts are positioning and strategy, acquisition (search, listings, referrals and reviews), conversion (pages, forms and booking), intake (the call, the message and the hand-off to the office team), tracking and attribution, and retention, together with the systems that connect them. Search engine optimization is one channel inside it. Chapter 3 develops the definition and says what it is not: a tool, a channel or a guarantee.

The paper's through-line is five questions that every chapter able to bear them asks of a practice, in the same order:

  1. Found: can a patient or client, or the assistant acting for them, find the practice and verify what it says?
  2. Chosen: does what they find resolve their uncertainty?
  3. Reached: does the first contact turn into an appointment or a matter?
  4. Within the rules: does all of it stay inside the rules the profession works under?
  5. Measured: can the practice count what happened at each step?

The order follows the path of a prospect, except that the last two are conditions on the whole path, not further steps along it: the rules bind every step, and the count is how a practice learns whether the other steps worked.

In our description, traditional web design treats a website like a billboard and Google like a phonebook. The goal is to be listed and to look good. A phonebook listing gave a name, an address and a number to a person who already knew what they needed, and a billboard cannot answer a question or take a booking.

A technical marketing system turns Google into a funnel, a sequence of stages through which a stranger becomes an appointment, each of which can be counted. It turns the website into a working rep for the front office: it answers the question the visitor arrived with, books, qualifies the inquiry and hands off to the office team. On this view web design starts with system design, framing and plumbing first and the finish fitted to the practice afterward, the way a house is built. We argue that SEO, web design and technical systems pay when they are built together and measured as one system, a claim the paper does not test.

When the first reader of a practice's website may be an assistant, the site has two audiences with different needs: a machine that reads its facts, and a person who arrives to ask, book or call. We think a brochure site serves the second poorly and the first only by accident. We offer the image as a way of saying what is built and in what order. The evidence in this paper bears on whether the stages exist and can be counted, not on whether the image is apt.

Scope, evidence and how each claim was checked

Scope

The paper covers four professions: dentistry, medicine, law and accounting. For medicine it means physician practices and clinics, and for accounting it means accounting and tax practices. It covers the United States, and practices that are small or mid-sized in the sense above. The rules it cites are federal texts, model rules, professional codes and, where stated, particular states, and none should be read as the rule everywhere. It does not cover hospital systems, large firms or national dental support organizations, although Chapters 6 and 7 note where ownership is changing the picture, and it does not cover creative strategy, brand design or media buying beyond their place in the system. A companion paper, How technical marketing systems make AI profitable, takes up the economics of AI returns in service businesses, including how to attribute an AI agent's own contribution.

Kinds of evidence

Table 2 lists the kinds of evidence the paper uses and how each is treated.

Kind of evidenceWhat it is in this paperHow it is treated
Economics and marketing scienceClassic and peer-reviewed work on search, credence goods, signaling, service quality, advertising measurement and response speedUsed for theory and for the limits of measurement; works cited from secondary summaries are marked and attributed in the sentence
Primary legal and regulatory sourcesStatutes and regulations, agency guidance and enforcement announcements, court decisionsDescribed, never applied; dated; the force of each is stated
Professional codes and opinionsEthics codes and formal opinions of dental, medical, legal and accounting bodiesTreated as model or advisory texts, not as law
Official statisticsBureau of Labor Statistics, Census Bureau and CDC releasesEach counts what its instrument counts; different instruments are not combined
SurveysSurveys by professional bodies, vendors and consultanciesVendor and consultancy surveys are attributed in the sentence and never carry a conclusion alone
Our own dataThe recommendation study, Google Ads keyword data, a check of tool vendors' agreements, our pricesIntroduced as ours; limits carried

Table 2. The kinds of evidence in the paper and how each is treated.

How each claim was checked

Every statement of fact about the world in this paper rests on a claim recorded in an evidence file. Definitions, arguments and judgments are not claims of that kind, and the text labels them. For an outside source, the check was to open the source, find the passage on the page and record it. A claim whose source was a secondary report, or one our check could not read in full, is marked as less certain; the text attributes it inside the sentence and never rests a conclusion on it alone. A quotation is the recorded passage, and arithmetic on cited figures is shown. An example takes its inputs from cited figures, and a link that nobody has measured is solved for, as the count or share needed beside a cited base rate, not guessed. Where the evidence for a section was thin, the section is argued from first principles and says so.

This paper does not measure four quantities itself, so each rests on the published record. The first, how often calls to small practices go unanswered, comes from the call figures above.

The second is the share of inbound calls an AI voice agent finishes without staff. A randomized or field study of an AI agent answering inbound calls or scheduling in a dental, medical, legal or accounting setting is NF. The nearest peer-reviewed study is of outbound pre-procedure calls at one hospital, where an AI voice assistant completed 86.4% of calls in the first phase and 36.6% were fully automated, needing no human intervention47. The nearest vendor figure is a release from Relatient† (scheduling) saying its voice product answered 62% of inbound calls, of which 72% were handled or passed to staff48. That is 45% of inbound calls handled or passed on, with the escalations inside it (72% x 62%). An agent finishes ~3 to 4 of every 10 inbound calls on its own (from the hospital's outbound calls; likely high for open-ended inbound scheduling, because those calls followed a script). The cases in later chapters solve for the share an agent must hold to cover its cost and do not depend on this estimate.

The third is how many new patients or clients a practice gains in a month. For dentistry, Henry Schein One† (software) reports an average of 30 new patients a month for practices of one to seven locations49. For primary care, new patients were 12.5% of office visits in 2019 in the CDC's National Ambulatory Medical Care Survey50. For accounting, a trade survey whose sample the report does not describe found that 63% of firms said 6% or less of their client base is new each year51. For law firms, call volume and inquiry-to-client conversion are NF; the nearest evidence is Clio†'s call test above and its report that the average lawyer bills 38% of an eight-hour day52, so the law cases in later chapters solve for the number of matters needed to cover a purchase.

The fourth is the conversion rate from an AI recommendation to an appointment. We make no estimate, because the nearest neighbors disagree in direction just over a year apart: Adobe reports that AI-referred retail traffic converted 9% worse than other traffic in August 2025 and 90% better in September 202653. Where a chapter needs a figure of that kind it solves for the number needed, beside the cited base rate where one exists.

What our own data means

By our own data we mean four things. The first is the recommendation study: 199 questions put to ChatGPT, Gemini and Perplexity across 16 US cities and 23 kinds of business on 2026-11-101. The second is Google Ads keyword data for November 2026, which gives US monthly searches and average cost per click for phrases we chose.

The third is a check of 28 software tools that small practices commonly use, read from the vendors' own pages, to see whether each will sign a business associate agreement54. A business associate agreement is the written arrangement through which, in HHS's description, a covered entity obtains satisfactory assurances that a vendor handling protected health information will safeguard it55. The fourth is our own services and prices. All four are cited like any other claim and introduced as ours, and the limits of the first are carried wherever a chapter leans on it1.

An interest to declare

We have an interest in the argument. Technical Marketing sells marketing plans56, an AI voice agent56, managed marketing56 and HIPAA-ready programs57, and the paper's thesis recommends the kind of work those services do. A reader should weigh our interpretations with that in mind. The paper tabulates its data so that a reader can read it differently, and says where the evidence runs out. It contains no client cases or testimonials, because we have none to show. It gives no guarantee about rankings, AI answers, leads or revenue, and it does not describe any tool or program as HIPAA compliant or certified. Two of our plans carry a guarantee on a measure agreed after an audit56; the paper's argument does not rest on it. It gives no legal, medical, tax or financial advice. Chapter 14 sets out the objections most likely to apply to a seller's argument.

A map of the fourteen chapters

Chapter 2 sets out the economics of choosing a professional and argues that an AI assistant reads the signals a practice sends for the buyer. Chapter 3 defines technical marketing in full and compares it with creative marketing and with SEO.

Chapter 4 presents our recommendation study as the paper's central empirical chapter, with its method and limits, and reads it for dentists, lawyers and accountants; the study has no row for physicians. Chapter 5 uses search volumes to show what practices look for and the gap between demand for marketing and demand for AI.

Chapters 6 to 9 apply the through-line to one profession each: Chapter 6 to dentistry, Chapter 7 to physician practices and clinics, Chapter 8 to law firms and Chapter 9 to accounting and tax practices. Each closes with a system design for one practice, an illustration and not a case.

Chapter 10 reduces the four rulebooks to a short set of design requirements, shows how one system can meet them and says what the architecture cannot do. Chapter 11 says what a small practice can and cannot know about what its marketing returned. Chapter 12 treats the front desk as the conversion point and examines what an AI receptionist may do in each profession. Chapter 13 turns the argument into a first-year sequence and is the one chapter that addresses the reader as "you".

Chapter 14 states what the paper has not shown, answers the strongest objections from the evidence and lists what we do not know. A practice owner who wants the evidence first can read Chapters 4 and 5 and the chapter for their own profession before Chapter 13.

Chapter 2

The economics of choosing a professional

Why choosing a professional differs from choosing a product

Choosing a dentist, a physician, a lawyer or an accountant differs from choosing a product in three ways that economists have described separately. The buyer has to search among sellers who differ in price, availability and quality. The buyer often cannot judge quality before buying and sometimes cannot judge it afterward. And the seller, who knows more than the buyer, is often the person who tells the buyer what is needed. Each difference has its own literature and its own consequence for how a practice comes to be known and trusted.

The chapter builds to an argument of its own: an AI assistant now performs part of the buyer's search and part of the buyer's checking, reading signals on the buyer's behalf. If that is so, marketing for a profession becomes the production and upkeep of signals that a buyer, or a machine acting for the buyer, can verify.

The text marks where this argument goes beyond the evidence. Several of the classic works are held only as abstracts, citation records, excerpts or a later author's summary; the References list says which, and we quote only the words we read.

Search costs and the economics of information

In February 1959, Chicago dealers' asking prices for an identical Chevrolet model averaged $2,436, with a range from $2,350 to $2,515 and a standard deviation of $4258. The range is $165 ($2,515 - $2,350), or 6.8% of the average ($165 / $2,436). Stigler reported these figures in his 1961 paper on the economics of information. He wrote that "dispersion is ubiquitous even for homogeneous goods" and called price dispersion a manifestation, indeed the measure, of ignorance in the market58. If every buyer knew every price, no dealer could charge more than the lowest; dispersion persists because knowing is costly.

Where prices differ, a buyer gains by looking. A buyer "who wishes to ascertain the most favorable price must canvass various sellers (or buyers)," Stigler wrote, and he termed that canvassing search58. Search costs time and effort, so it stops somewhere. His rule was that the optimum amount of search will be found if "the cost of search is equated to its expected marginal return"58. Anything that lowers the cost of one more look makes the buyer look more, and anything that lets the buyer skip looks lowers the cost of deciding.

Stigler named two devices of the second kind. Reputation, he wrote, "commands a price (or exacts a penalty) because it economizes on search"58. And advertising is "the obvious modern method of identifying buyers and sellers," where "the classified advertisements in particular form a meeting place for potential buyers and sellers," so that the identification of buyers and sellers "reduces drastically the cost of search"58.

Stigler's object was the price of an identical good; the object of search for a professional is wider. It includes the fee, if it can be learned beforehand, whether the practice takes the buyer's insurance, handles the buyer's kind of matter and can see the buyer soon, and, above all, quality, which in these markets has no posted number. We expect dispersion in what a professional offers to be at least as wide as in Stigler's cars, because the service is not homogeneous.

Fee surveys show some of the spread. In the State Bar of Michigan's 2023 survey, 2,191 attorneys reported hourly billing rates with a median of $300, a 25th percentile of $250, a 75th percentile of $400 and a 95th percentile of $62059. In the National Association of Tax Professionals' 2025 fee study, half of the preparers who charge a minimum fee plus a cost based on complexity charge between $150 and $328 for a base Form 1040 with Schedules 1-360. Their average is $23660. The 95th percentile is 2.1 times the median ($620 / $300), and $328 is 2.2 times $150.

Both surveys record what professionals say they charge, one among the lawyers of a single state who answered and one among the preparers the association could reach, and neither says anything about quality. A fee survey for dentists or physicians is NF; the nearest evidence is the lawyer and tax-preparer surveys above. We estimate that fees for the same service spread about two to one between the low and high ends of a typical range (from those two spreads; likely low). The reason is that the Medical Expenditure Panel Survey's 2016 mean expense per office visit ran from $159 to $419 across specialty types, 2.6 to 1 ($419 / $159)61. That is a spread between specialties, not among practices of one kind. This paper does not measure dispersion in quality.

One study of local search shows the mechanism at small scale. Ghose, Goldfarb and Han found that ranking effects are higher on mobile phones, which suggests higher search costs, and that links at the top of the screen are especially likely to be clicked there. They also found that stores close to a user's home are much more likely to be clicked on mobile phones28. The passages we cite do not give their sample or say which businesses it covered, so we use the result for the mechanism and not for any size of effect. It fits Stigler's rule: when each additional look costs more, buyers stop sooner, and a buyer who stops sooner chooses from what is placed first and what is near.

Search, experience and credence qualities

Nelson's 1970 article, Information and Consumer Behavior (citation record only), analyzed how consumers acquire information about product quality and distinguished goods whose quality can be judged by search before purchase from goods whose quality is learned by experience62. In an excerpt, Nelson defines search as any way of evaluating the options subject to two restrictions: the consumer must inspect the option, and the inspection must occur before purchase62. In a 2010 review, Nakayama, Sutcliffe and Wan describe the two classes. Search goods, as they summarize Nelson, are products consumers can "search, inspect, and compare" before purchasing63. For experience goods, they write, "consumers cannot know the quality of products such as cars and air conditioning systems until they purchase and use them for a while"63.

Darby and Karni (1973) added a third class (citation record only)2. An excerpt reports that they distinguished three types of qualities associated with a purchase: search qualities, known before purchase; experience qualities, known costlessly only after purchase; and credence qualities, expensive to judge even after purchase2. Their article, Free Competition and the Optimal Amount of Fraud, analyzed how market conditions and reputation affect the equilibrium amount of fraud2. That is a question about equilibrium in theory and says nothing about the conduct of any profession.

Dulleck and Kerschbamer's 2006 survey of the economics of credence goods asks: "Under which conditions do experts have an incentive to exploit the informational problems associated with markets for diagnosis and treatment?"64. Its examples are doctors, mechanics and computer specialists. We think the question describes a seller who both diagnoses the need and supplies the remedy, so that quality is hard to see and the buyer must take the seller's word on what is needed, and that this holds in each of the four professions.

Zeithaml (1981, excerpt only) applied the three classes to the difference between goods and services65. A 1995 account by Bendixen and Gault (secondary) reports that she posited eleven hypotheses about how consumers evaluate goods and services differently66. The first reads: "Goods will be higher in search qualities; services will be higher in experience and credence qualities"65. Her Figure 1 "arrays goods and services high in search, experience, and credence qualities along a continuum of evaluation"65. The continuum matters because no service is purely one kind. A dental visit contains a location and a price that can be checked beforehand, a manner and a wait that are learned in the chair, and a diagnosis the patient cannot assess.

Two further works (excerpts only) bring the point to professional services. Day and Barksdale (1992) wrote that purchasing professional services "is different from buying physical products and many other services," because such services "are typically intangible and customized," so that "evaluation prior to purchase is difficult"3. Bendixen and Gault (1995) present a consumer evaluation model in which a "pure service" segment has "high degrees of credence qualities," legal advice and medical diagnosis are among its typical products, and "Word of mouth recommendation is key to decision-making"66. These works show that scholars placed legal and medical services at the credence end. A model does not measure any market, and neither work shows how any buyer behaves.

Nothing in the evidence classifies dentistry, medicine, law or accounting, so we place them on this scale ourselves in Table 1, reasoning from the structure of each service. We classify by where most of the value to the buyer lies, because every service mixes all three kinds of quality.

ProfessionSearch qualities: can be checked before buyingExperience qualities: known after the visitCredence qualities: hard to judge even afterward
DentalLocation, hours, insurance accepted, services listed, stated trainingWait, comfort, manner of the staff, clarity of the billWhether the recommended treatment was needed; whether the work is sound
MedicalLocation, insurance accepted, stated credentials, how soon an appointment is offeredWait, whether the physician listens, ease of follow-upWhether the diagnosis was right and the treatment needed
LegalPractice area, location, stated credentials, fee terms where publishedWhether calls are returned, clarity of explanationWhether the strategy was sound, and whether the result was owed to the lawyer's work or to the facts of the case
AccountingServices listed, location, stated credentialsTimeliness, clarity, whether the return was filed or the books balancedWhether the advice was the best available, and what was never found or never claimed

Table 1. Where the main qualities of four professional services fall, by type of quality. The placement is the authors' analysis.

Dentistry has the most search and experience content of the four. Its credence core is the recommendation: that a crown, a filling or an extraction is needed, and that it was done well, are claims the patient mostly cannot test. We expect routine hygiene visits to sit nearer the experience end and restorative recommendations nearer the credence end.

Medicine has the same structure with higher stakes. In one 2012 survey, reported below, insurance acceptance, which can be checked beforehand, was the factor most often rated very important. The diagnosis cannot be checked by the patient, and when treatment works the patient cannot tell whether recovery would have come without it.

We think law is the most credence-heavy of the four. The quality of a lawyer's work is confounded with the facts of the case: a client who loses cannot tell whether another lawyer would have won, and a client who wins cannot tell whether the lawyer was needed. The Supreme Court made a related observation in Bates, in a passage on a matter it said it was not addressing. Advertising claims as to the quality of services "are not susceptible of measurement or verification," it said, and so "may be so likely to be misleading as to warrant restriction"5. We think that recognizes how hard legal quality is to verify. The passage concerns lawyers' advertising claims and is not a holding on quality advertising. We cite it for what the Court said in 1977; later decisions may have changed the law.

Accounting differs by task. Preparing a return or keeping books produces visible artifacts: a filed return, a reconciled ledger, a refund. The credence part is what the client cannot see: whether the positions taken were the best supportable ones, whether a saving was missed and whether a risk was flagged. We place routine compliance work nearer the experience end and advice nearer the credence end.

Quality uncertainty, licensing and guarantees

Akerlof's 1970 article, written about used cars, supplies the next piece. Sellers know more about a car's quality than buyers do: "An asymmetry in available information has developed: for the sellers now have more knowledge about the quality of a car than the buyers"67. Good and bad cars must then sell at the same price, and as a result, he argued, "most cars traded will be the 'lemons,' and good cars may not be traded at all"67.

Akerlof continued: "Numerous institutions arise to counteract the effects of quality uncertainty. One obvious institution is guarantees"67. Another is licensing: "there is the licensing of doctors, lawyers, and barbers," and most skilled labor "carries some certification indicating the attainment of certain levels of proficiency"67. A third is the brand name, which, he argued, indicates quality and also gives the consumer "a means of retaliation if the quality does not meet expectations"67.

We do not claim that markets for professional services unravel as the used-car market does in Akerlof's model: practitioners practice and patients are treated. Our claims are narrower. First, the asymmetry is the same in kind, and in credence services it may be larger in degree, because even after the purchase the buyer cannot confirm the quality. Second, licensing certifies attainment of a level of proficiency. It separates those who meet the level from those who do not but does not by itself rank those who meet it. The buyer still has to choose among licensed practitioners, and that choice is where marketing operates. Third, we think guarantees are weak for credence services, because a guarantee is only as good as the buyer's ability to tell whether it was breached.

Professional rules add a second layer to licensing that makes some signals harder to fake. ABA Model Rule 7.2(c) bars a lawyer from stating or implying certification as a specialist unless the lawyer has been certified by an organization approved by an appropriate authority and the communication clearly names that organization68. Each state adopts the model in its own version. The ADA's code of ethics, which is not law, says dentists who announce specialization should use "specialist in" and must devote enough of their practice to the specialty to maintain expertise in it7. We think the first rule keeps a credential credible by tying it to a certifier outside the practice, and the second ties an announcement to the share of the practice's work that the specialty occupies. Chapters 6 to 10 describe what each profession's rules say.

Nothing in this chapter says that dentists, physicians, lawyers or accountants mistreat the people they serve, and this paper does not measure how often any of them do. The theory says only that when a buyer cannot check, honest and dishonest sellers look alike, and the honest ones need a way to be told apart.

Signals a practice can send and a buyer can read

Spence's 1973 article was about hiring, and it starts from a problem this chapter shares: "In most job markets the employer is not sure of the productive capabilities of an individual at the time he hires him"69. Replace the employer with a patient and the individual with a practice, and the problem is ours. Spence distinguished two kinds of observable thing. He called "observable, unalterable attributes" indices, and he reserved the term signals for "those observable characteristics attached to the individual that are subject to manipulation by him"69. His example was education, "something that the individual can invest in at some cost in terms of time and money"69.

Whether a signal separates better sellers from worse depends on cost. Spence's two-group model assumes that the cost of y units of education is y for one group and y/2 for the other, so the same education costs one group twice what it costs the other69. We think the model shows that a signal carries information only when it is cheaper for the better seller to send. A claim that costs nothing, such as a statement that the practice cares about its patients, carries none, because the practice that does not care can make it as easily. A credential that took years to earn, or a review history that took years to accumulate, is costly to send, and the cost is what makes it believable.

Spence's categories need adapting for a practice. Its credentials are signals in his sense, since the practitioner invested in them. Its reviews are neither indices nor pure signals, because the practice influences them through the service it gives and through asking, but does not write them. Its responsiveness is observable at the first contact and costs staff time. Table 2 lists the signals a practice can send.

SignalSpeaks toHow it can be checkedHow it fails
Credential: degree, license, board or specialty certificationCompetence at a threshold; a search qualityAgainst a register or certifier outside the practiceShows a threshold and not who is best; a credential that cannot be checked is a claim
Price and terms: fees, insurance accepted, what a first visit costsSearch qualityBy asking, or from a published pageOften not published; says little about quality
ReviewsExperience qualities; credence qualities only by inferenceOn a platform outside the practice: count, dates, textSkewed positive in the data we read; subject to rules on bought and suppressed reviews, noted below
Referral from a person or another professionalCredence qualities, through trust in the referrerHardly at all; the buyer trusts the personInvisible to the practice's own measures; limited to the referrer's circle
Responsiveness: answering, offering a time, replying to a messageExperience qualities, and careAt first contact by trying it; by an assistant only if recorded somewhere it readsCosts staff time; varies by hour and by buyer
Consistent facts across listings: name, address, phone, hours, servicesExistence and reliability; a search qualityBy comparing sourcesDecays as hours change, staff leave and listings drift
The practice's own claims: site copy, brochuresWhatever the practice chooses to sayNot by itselfCostless to make; separates nothing unless something outside confirms it

Table 2. Signals a practice can send, in the authors' framework.

One survey, about physicians, shows what buyers weigh. In the September 2012 nationally representative US survey reported by Hanauer and co-authors, whether a physician accepted one's health insurance was rated "very important" most frequently, by 89%, when choosing a primary care doctor70. Word of mouth from family and friends was rated very important by 38% and somewhat important by 47%, and a referral from another physician very important by 34%70. Physician ratings on websites were rated very important by 19% and somewhat important by 40%, and the authors note that rating sites "were endorsed less frequently than other factors"70.

Insurance, a plain fact that can be checked in advance, was rated very important most often, word of mouth and physician referral less often, and rating websites least70. The survey asks about physicians, about stated importance and not observed choices, and it is over a decade old. It is consistent with an ordering of signals by checkability in one profession at one date, and shows nothing more.

Why services are harder to judge than goods, and how service quality is judged

Parasuraman, Zeithaml and Berry opened their 1985 article (abstract and excerpts) by observing that marketers had described and measured quality in tangible goods, whereas quality in services had hardly been defined or studied71. Their model rested on an extensive exploratory investigation of quality in four service businesses71. What follows relies partly on a 2013 review's statement of the model.

As that review by Mauri, Minazzi and Muccio describes it, the model identifies five gaps, of which the customer gap, "the discrepancy between expectations and actual perception of service quality by the customer," is the main one72. In the excerpts, two of the others appear as propositions in the 1985 article: the gap between consumer expectations and management's perception of them, and the gap between service quality specifications and actual delivery, each bear on service quality as the consumer evaluates it71. On this view quality, as a buyer judges it, is the comparison of what was expected with what was perceived.

The review reports that the components of service quality were grouped in ten categories in 1985 and in five in 198872. It defines the five dimensions as tangibles, "aspects of physical facilities, equipment and personnel"; reliability, "capability of the firm to perform the promised service in a careful and accurate manner"; responsiveness, "willingness of the firm to help customers and perform the service promptly"; assurance, "competence and politeness of the personnel, capability to inspire confidence"; and empathy, "personalized assistance that the firm conveys to its customers"72. The 1986 working paper (abstract only) that introduced the scale called SERVQUAL reported a final version of 26 items measuring these dimensions73. An excerpt of the published 1988 article describes a 22-item instrument74.

In the 1986 study, across four services (appliance repair, a bank, long-distance telephone and a credit card), average expectation scores were around 6 on a 7-point scale and average perceptions around 573. The average SERVQUAL score was -173, the gap between the two averages (5 - 6). None of the four services is a professional service, and we do not carry the figure to one. The 2013 review also found that, despite criticisms, the gaps model and SERVQUAL were still the most used instruments for studying service quality in the marketing literature72.

The five dimensions show which parts of quality a buyer can observe at the moment of choice. Tangibles can be seen, in a photograph or on a visit. Responsiveness can be tried: a buyer calls, writes or books, and learns how promptly the practice helps. Assurance and empathy show partly in how the practice presents its people and in how it treats the first contact. Reliability, the careful and accurate performance of the promised service, is largely unobservable until the service has been delivered, and in credence services it may remain so afterward. This mapping is ours. It suggests that a buyer choosing in advance mostly weighs the dimensions that can be observed in advance, and that a reader of online sources, human or machine, finds most of the evidence about reliability in other people's reports.

What the research on online reviews shows, and in which markets

Online reviews are the newest institution of the kind Akerlof described, and the empirical work on them is the best developed part of the evidence, almost all of it from other markets. Table 3 lists the studies cited.

StudyMarket and dataResult
Luca, working paper, 2011, revised 2016Restaurants in one city, using Washington State Department of Revenue data and Yelp's rounding of displayed ratingsA one-star increase in Yelp rating led to a 5 to 9% increase in revenue; the effect was driven by independent restaurants, and ratings did not affect restaurants with chain affiliation75
Anderson and Magruder, 2012328 restaurants in the San Francisco Bay Area and 148,000 Yelp reviews, according to a Berkeley press release76An extra half-star rating made restaurants sell out peak reservations 19 percentage points more frequently, with larger impacts when alternate information was scarcer77; restaurants rated in popular guidebooks or newspaper rankings showed no statistically significant effect, according to the release76
Chevalier and Mayzlin, 2006Books sold at Amazon.com and Barnesandnoble.comAn improvement in a book's reviews led to an increase in relative sales at that site, and for most samples one-star reviews had a greater impact than five-star reviews; reviews were overwhelmingly positive at both sites78
Gao and co-authors, 2012Patients' ratings of physicians on one site, RateMDsMean rating 3.93 on a scale of 1 to 5, with 45.80% of rated physicians receiving a perfect 579
Hanauer and co-authors, 2014A September 2012 survey of the US public about online physician rating sitesUsage of the sites was 23%; among respondents who had sought ratings in the past year, 35% selected a physician on the strength of good ratings and 37% avoided one with bad ratings70
Lin and co-authors, 2020Reviews of 204,751 US dentists on one platform, HealthGradesHigher ratings were associated with female dentists, younger dentists and short patient waits80
Block, Reddy and Corkum, 20231.12 million new patient appointments and 12,882 physician reviews at one large single-specialty practice in the New York and New Jersey areaMean new patient volume decreased 2.3% to 2.6% after a 1-star review, with effects lasting at least 16 weeks81

Table 3. Empirical studies of online reviews held in our evidence, by market and data.

The restaurant studies show ratings moving revenue and sell-outs most where the buyer had least other information. Luca found no effect among chains and suggested that "online consumer reviews substitute for more traditional forms of reputation"75. We think a chain affiliation is a brand name in Akerlof's sense, an existing signal, so a rating adds little where one is present. Anderson and Magruder's no-effect result for restaurants already rated in guidebooks or newspapers, as the press release reports it, fits the same reading76.

That pattern fits Stigler's and Akerlof's reasoning, which is why we think the result bears on professionals: many practices are small and local, and unknown to a buyer who has just moved or just developed a need, which resembles the condition under which the restaurant studies find that reviews matter most. The three studies concern restaurants and books, and none tested the effect of reviews on a professional practice. The one study in Table 3 that did, by Block and co-authors, covers a single practice in one specialty and compares new patient volume before and after a review. It shows an association, not a rate for physicians in general.

We expect transfer to professionals to fail in at least four ways. First, stakes and frequency differ. A restaurant meal is, as a rule, cheap and repeated, and its quality is learned by eating it, whereas the choice of a lawyer is rare and consequential, and a buyer may want more than a star rating before deciding. Second, a review can report only part of quality. Lin and co-authors' association between ratings and short waits fits the view that ratings record what a patient can observe, such as waiting, which is an experience quality. It says nothing about whether patients chose dentists on the strength of ratings, so we think it suggestive only. A patient cannot rate whether a diagnosis was right.

Third, the shape of the data limits what a rating can say. In the data of Chevalier and Mayzlin's 2003 working paper, 53.05% of Amazon.com reviews and 67.22% of Barnesandnoble.com reviews were five-star82, and 45.80% of the rated physicians in the RateMDs analysis had a perfect 579. When ratings cluster near the top, a rating separates sellers poorly at the top, and the information lies in the low ratings. That fits the finding that for most samples one-star reviews had more impact than five-star reviews78, and, by Spence's logic, a signal that nearly every seller sends at no cost cannot tell the sellers apart.

Fourth, the rules differ. Replying to a review raises confidentiality questions that Chapters 6 and 8 take up, and federal rules on consumer reviews bar a business from writing or selling a review that materially misrepresents the reviewer or the reviewer's experience, and from giving compensation or incentives in exchange for reviews of a particular sentiment83,84. They also bar anyone from using an unfounded legal threat, intimidation or a knowingly false public accusation to keep a review from being written or to have it removed85.

For lawyers and accountants, this paper cites surveys and no peer-reviewed study of reviews. A BrightLocal† vendor survey reports that 81% of consumers felt reviews were important or very important in the financial and legal sector in 202386. FindLaw†, a legal marketing company, ran a 2024 survey of 2,000 US adults who had a legal need in the past year87. It reports that 82% of those who contacted an attorney after learning about them online used online reviews in their decision, and that nearly 40% said reviews were their primary source of information87. The base is a subgroup, the people who found their attorney online, and a stated use of reviews is not a measured choice.

A measurement of the effect of reviews on an accounting firm is NF; the nearest evidence is the 5 to 9% revenue effect Luca found for independent restaurants75 and the 2.3% to 2.6% fall in new patients after a 1-star review at one medical practice81. We estimate that a one-star difference moves an accounting firm's new clients by a few percent, at most the first figure and probably nearer the second (from those two; likely high). The reason is that the accounting surveys below name referral as the leading source of new clients and Luca found no effect where another signal was already present.

The main peer-reviewed US survey in our evidence of how the public uses reviews to choose a professional concerns physicians. In September 2012, 65% of US adults were aware of online physician ratings, lower than awareness for cars (87%) and for non-health-care service providers (71%)70. Among respondents who had not sought ratings, 43% reported a lack of trust in the information on the sites4. We cite the survey to show that awareness was uneven and trust was contested, not to give current levels.

Pew's Health Online 2013 survey, also fielded in 2012, found that about one in five internet users had consulted online reviews of drugs or treatments, doctors or other providers, or hospitals, a count that combines the three88. A medical society reports that Press Ganey† (patient experience) found in a 2021 survey of over 1,000 US adults that consumers used three websites and read 5.5 reviews on average before deciding on an appointment89. It also reports that 84% would not see a referred provider whose online rating was under 4 stars, a stated intention89.

Referral and reputation, the historic channel

We think that before search engines and review platforms, referral did the work that Stigler assigned to reputation. It carried quality information from someone who had been served to someone who had not. The evidence on how much it still does is fragmentary: it comes from different professions, years and countries, and from surveys of opinion or stated behavior that count different things. A share of actual new patients or clients by source, taken from practice records, is NF in any of the four professions; the nearest evidence is the stated figures below.

In a September 2010 Harris poll for the ABA Standing Committee on the Delivery of Legal Services, as reported by the ABA Journal, 46% of respondents said they would ask a friend, family member or colleague for a lawyer referral, and 34% said they would contact a lawyer they knew or had used before90. Those are stated intentions from a news report of the poll.

In Clio†'s consumer expectations survey, as relayed in a 2019 summary by an Illinois court commission, 59% of clients sought a referral to an attorney while 57% searched for one on their own91. If both shares are of the same group of clients, they overlap by at least 16 percentage points (59% + 57% - 100%), so at least 16% of them did both. The same Illinois summary reports that only 46% of Millennials were likely to value referrals, against 60% of Boomers91. FindLaw†'s survey reports that 48% of those who learned about the attorney they contacted said they used referrals from family members, friends or trusted coworkers, the most popular offline resource87. Using referrals as a source of information is not the same as finding the lawyer by referral. Clio† and FindLaw† are both vendors, so we rest nothing on them alone.

The dental evidence is of mixed kinds, and the peer-reviewed part comes from outside the United States. In an online survey of adult dental patients in private clinics in Bucharest, Romania, 99.1% said they learned about their dentist or dental office from family or friends, while 20.5% also searched online reviews92. We cite it as a dental example of the pattern, not as an estimate for the United States.

A peer-reviewed US survey of how patients choose a dentist is NF; the nearest evidence is weaker. MouthWatch†, a teledentistry company, surveyed 448 dental practices, not patients. It found that 65.6% named word of mouth and patient referrals as their top source of new patients, down from 77.5% in its 2025 survey, whose sample may differ93. IronMonk†, a marketing agency, asked 1,500 Americans a hypothetical question in 2019. Of them, 43.3% said that, choosing among dentists in a Google search, they would pick the one with the best online reviews, 19.4% the closest and 18.4% the top-ranked result94. Referral and insurance acceptance were not among its options. Patient Prism† (call analysis) reports that 90% of the attributable patient calls came from the Google ecosystem, and that Google Business Profile alone drove 54% of attributable inquiries24. That counts the calls the vendor could trace to a source and does not describe how patients chose.

For physicians, the 2012 US survey above rated word of mouth and physician referral very important more often than rating sites70. Press Ganey†'s surveys point the other way. The society's summary of the 2021 survey reports that patients rely on digital resources 2.2 times more than provider referrals when choosing a provider89, and Press Ganey†'s own 2024 blog says that 83.5% of consumers who have a referral in hand at least occasionally turn to online reviews to validate it95. Both concern health care in general.

For accounting, three surveys point toward referral, and none asks individual taxpayers how they chose a preparer. In a survey that Accounting Today ran with ADP, 79% of accounting firms named unsolicited client referrals, and 79% named referrals from fellow professionals, as a top source of new clients, against under 10% for advertising51. The firms were describing where their clients came from, the report gives no sample size, and a firm could name several sources.

CPA.com's benchmark of client advisory practices found that 64% named external referrals, 67% existing tax clients and 4% advertising as a top source of leads for that service line96. TaxDome†, a vendor of practice software, surveyed more than 350 US business decision-makers. It reports that 58% said they found their current accounting firm through a referral and 17% through online search, and that 85% called responsiveness very or extremely influential in selecting an accountant97. The IRS's 2021 Comprehensive Taxpayer Attitude Survey of 2,099 members of the public found that 88% rated paid tax professionals among the most valuable sources of tax advice and information98. That measures how far taxpayers value professionals, not how they pick one.

Gluckler and Armbruster's 2003 paper on the market for management consulting argues that its defining features are "institutional and transactional uncertainty" and that "market competitiveness is based on experience-based trust and networked reputation"99. Management consulting is not one of our four professions, but on the authors' account it is a professional service sold under similar uncertainty.

The pieces do not all point the same way, and none shows how often a referral decides the outcome. We therefore estimate that the share of new clients who arrive after a referral, in a law, accounting or dental practice, falls in a range from roughly half to two thirds (from the stated shares above; likely high). The shares are 48% who used referrals in the legal survey, 58% who found their accounting firm through a referral and 65.6% of practices naming referral their top dental source87,93,97. The reason is that a person asked about a referral may report one that did not decide the choice, and that Patient Prism†'s call-tracking data attribute 90% of traceable dental calls to Google24.

The instruments count stated intentions, shares who sought a referral, shares who learned of a practice and practices' opinions of their top source, in different populations and years, so we do not average them. Referral and search are complements as well as substitutes, since a person who is referred to a practice may then look it up. We infer this from the overlap in Clio†'s shares, and Press Ganey†'s 83.5% fits it95. It matters because a referred buyer who looks the practice up meets whatever the search shows. A referral is also a signal a practice can earn but cannot observe, because it travels in conversations the practice is not part of. Asking new patients or clients how they came is the most direct way to see the channel.

What an AI assistant does to the buyer's search and verification

So far the theory has described a buyer who does the reading: canvassing listings, opening sites, reading reviews, asking around and weighing the signals. Search engines have long done part of that work. Google's Business Profile help page says that local results are "mainly based on relevance, distance, and popularity," and that prominence is based on information such as how many websites link to a business and how many reviews it has100. It also says there is no way to request or pay for a better local ranking100. That is a company's description of its own ranking as the page stood in November 2026, not an independent measurement of its weights, and it is enough to show that a ranking system was already reading signals for the buyer. It returned a list and left the buyer to read it.

An assistant goes further. It reads sources, chooses among them and writes a recommendation, often with citations. In its 2025 study, Pew Research Center found that 88% of the Google AI summaries cited three or more sources30. In our study the assistants named 2,975 businesses across 199 questions, and the sites they cited most often were directories, review platforms, list sites, publishers and a forum1. The buyer who once read the listings, the reviews and the practice's site may now meet them already condensed. Evidence that this changes behavior is thin and indirect.

In a 2025 Pew Research Center analysis of US adults' browsing data, users who encountered a Google AI summary clicked a traditional search result link in 8% of visits, against 15% of visits without a summary30. The analysis is of general browsing and is not broken out by profession. A hand-run test by Whitespark†, a local-SEO vendor, described in Chapter 1, found AI Overviews on 15% of local-intent queries against 92% of informational ones38. We infer, without measuring it, that searches for a practice meet an AI Overview less often than information searches do.

In Stigler's terms, we think the assistant lowers the buyer's cost of canvassing and changes who canvasses. The buyer's search is delegated, and with it the first reading of the signals: the machine decides which sources count, which facts to extract and what to say. The delegation has an awkward property: the assistant is itself an intermediary whose quality the buyer cannot easily judge, so the credence problem moves up one level, from the practice to the assistant. Our study gives a reason for care. On the 160 questions asked in all 16 cities, the three assistants named 2,414 businesses, and 84 of them (3.5%) were named by all three1. The figure carries the limits that Chapter 4 sets out: each question was asked once, names were matched across assistants by spelling, up to eight names were read per answer, and answers through an API are a baseline and not what every signed-in user sees1. It does not show that any assistant is wrong.

An assistant can use only what exists in sources it can reach. Signals that leave no trace in any source, such as a receptionist's manner, are invisible to it. Signals that exist in several independent sources, such as a consistent name, address, telephone number, credential and review history, can be checked against one another. The 25 most-cited websites drew 29% of all citations, and they are of the kinds named above, not any one local business's site1. Sites cited in only one city, where a local business's own site would fall, drew 39% of citations1. That is consistent with assistants leaning on sources outside the practice, but our study counts names and sites separately and does not show why any site was cited. And 88% of the websites cited were cited by only one of the three assistants, so the three mostly cited different sites1.

The argument has four steps. First, a professional service is largely a credence good, so buyers rely on signals. Second, a signal separates better sellers from worse only if it is costly to send or can be checked by someone outside the seller. Third, part of the reading of signals has moved from the buyer to an assistant, which reads at scale and chooses what to report. Fourth, a reader of that kind can use only signals that exist in readable, consistent and checkable form. It follows that the practice's task is to make the facts that bear on choice exist, agree with one another and be checkable wherever the readers look. Steps one and two are theory we have cited. Step three rests on indirect evidence and is open to dispute, and Chapter 14 lists it among what we do not know. Step four is logic, and the conclusion is ours.

The strongest objection is that step three carries the whole argument. A measurement of AI referrals or conversions on the website of a dental, medical, legal or accounting practice is NF; the nearest figures, given in Chapter 1, count stated use and traffic on other kinds of site. Chapter 1 builds ~1% of visits or less from them, with the basis and the way it likely errs. We accept the objection, and our answer is conditional. To the extent that reading moves to machines, the work described here pays, and the same work serves the buyer who still reads the sources personally, the referred buyer who looks the practice up and the list-based search Google describes. We think the cost of being wrong about step three is small, because the signals that serve an assistant are the signals a person checks.

A second objection is that signals can be gamed. They can, and Spence's logic says why producing more of them is no answer: a signal that every practice can produce at no cost stops separating. What keeps its value is what is costly to fake: real credentials, real reviews from real patients or clients, and real responsiveness. The paper that introduced generative engine optimization reports that its methods can boost visibility by up to 40% in generative engine responses31. Its authors also say that content creators have little to no control over how those engines display their content31. Google says there are no additional technical requirements for its AI features beyond being indexed and eligible to appear with a snippet32. We take no position on tactics beyond this: a practice should expect to compete on facts, and the evidence does not support promises of placement.

Dental, law firm and CPA marketing as the production of verifiable signals

If the argument holds, the product of marketing for a profession is a set of facts and signals with three properties. They are true, they are legible to a person and to a machine, and they can be verified outside the practice. We call a signal verifiable when a reader can check it against a source the practice does not control: a credential against its certifier, a telephone number against several directories, a review history on a platform, a claim of specialization against the body that confers it. Four consequences follow.

First, truth is the precondition and upkeep is the work. Facts decay: hours change, a provider leaves, and a listing keeps an old telephone number. A practice whose facts appear in many places needs a way to change all of them when one changes. Second, responsiveness has to be built, because a page that says the practice responds promptly costs nothing to write. A first contact that works costs staff time or a system, and a buyer who tries it learns something no claim can say. Chapter 12 treats that first contact.

Third, reviews need a program, and the program has to stay inside rules on what may be asked, offered and answered. Fourth, none of this can be managed unless it is counted: whether the signals are present, whether they agree, and whether the buyers who see them act. Chapter 11 sets out what a small practice can count, including a dated check of what each assistant says. Taken together this is engineering in the ordinary sense. It specifies what must be true, builds the connections that keep it true, and measures whether it is. Chapter 3 gives the discipline a name and a definition.

Signals describe quality and do not create it. A practice with strong signals and poor work will harm the buyer who relied on them, and a practice with excellent work and weak signals will be passed over. Work on signals can reduce the second error, in which good work goes unseen. It cannot reduce the first, in which a signal outruns the work, and that belongs to the profession's own standards and its regulators. This paper does not test whether better signals lead to better matches between buyers and practitioners; that is a hypothesis for future research.

Chapter 3

What technical marketing is, and what it is not

A definition of technical marketing, with its parts

Chapter 1 defined technical marketing as the design, construction and measurement of a firm's acquisition, intake, compliance and reporting as a single system. In it each stage has an owner, an input, an output and a count, so that a change at one stage can be traced to its effect on the next. The definition and the comparison below are the paper's framework: they organize the argument and do not report a consensus in the literature. The phrase "technical marketing" is searched about 210 times a month in the United States, in Google Ads keyword data27, too rarely for us to assume a shared meaning.

The definition has parts, each with a job. Positioning and strategy decide whom the practice serves, for which services and in which places, within what capacity and under which rules. Their output is a short list of the questions the rest of the system must be able to answer and the actions it must be able to take. Acquisition puts the practice in front of people who are asking. It covers search, listings, referral and reviews, and it is the part Chapter 2 treated as the production of signals that a buyer or an assistant can check. Search engine optimization is one channel inside acquisition: the work of making a practice's pages and facts eligible to appear in organic search results.

Conversion turns a visit into a request, such as a call, a form or a booking. Intake is what happens to the request: someone or something answers, asks the few things the office needs to route it, offers a time or a next step, and hands the person to the office team with the context intact. Because demand for a professional service arrives as a first contact, we treat conversion and intake as separate stages and count them separately: a page can convert and an intake process can still lose the person. Chapter 12 treats intake and the front desk.

Tracking and attribution record what happened at each stage and, as far as the evidence allows, where the person came from; the measurement literature below shows how hard that is. Retention brings people back and, in some professions, brings their families and colleagues. It is the stage where earlier work pays again, and it is counted like the others. The systems that connect the parts are the data flows, integrations and agreements that move a fact or a person from one stage to the next without losing it or sending it somewhere it should not go. In health, tax and legal settings that last clause carries legal weight, and Chapter 10 treats it.

The five questions of Chapter 1's through-line map onto the parts. Found rests on positioning and acquisition. Chosen rests on what the buyer finds, which means the practice's pages and the signals of Chapter 2. Reached is intake. Within the rules and measured are conditions on every part. Retention extends the path, because a patient or client who is reached is later reached again.

Creative marketing, SEO and technical marketing compared

Creative marketing, search engine optimization and technical marketing are easy to confuse. Table 1 compares them by what each runs on, produces and is judged by. It describes ideal types as we define them. Real agencies and in-house teams mix the three, and nothing in the table says that one is better than another in general.

Creative marketingSearch engine optimizationTechnical marketing
Runs onInsight into people, a message and craft: writing, design, photography, mediaThe signals search engines read: content, links, site structure, listingsA map of the stages, data flows, tracking, integrations and the office's own process
ProducesAttention and a distinct impression: a brand, a campaign, creative workVisibility in organic search results and the visits that followA working path from a question to an appointment or a matter, with a count at every stage
Is judged byWhether the message is noticed, remembered and persuasive; often by reach, recall or tasteRankings and organic traffic, and sometimes leads from searchWhether the stage counts connect to results, such as appointments kept, matters opened and clients retained, and whether each change can be traced
Fails whenThe message is good and nothing behind it can answer, book or be countedVisits arrive and no one answers, or the visits cannot be tied to outcomesThe system is built and no one operates it, or the counts are wrong

Table 1. The three disciplines compared.

The last row applies to us as well: a reader should ask any provider, ours included, about the two ways technical marketing fails. A system nobody operates decays: hours change, staff leave and tracking breaks. A system with wrong counts is worse than none, because it directs attention to the wrong stage with the authority of a report.

Technical marketing needs the other two: creative work for the content of what is said, because a system that routes visitors to a poor message does not help the practice, and SEO for one channel, because search is one of the ways people ask. Its unit of work sets it apart. Creative marketing works on a message and SEO on a ranking; technical marketing works on the whole path from the question to the outcome, and requires a count at every stage so that a change at one place can be followed to the next.

The pipeline view: why improving one stage does not move the result

A practice's marketing can be drawn as a pipeline of stages through which a stranger becomes a patient or client. The through-line names the first three stages, found, chosen and reached, and the definition adds a fourth, retention. Each stage takes an input from the one before and hands on an output, and a count at each stage lets the owner see where the path narrows. The picture is simple; its use is to explain how a practice can pay for an improvement at one stage and see no change in the result it cares about.

We think the result may fail to move for three reasons. The first is capacity: a stage with a fixed capacity caps what passes through it, so extra volume from an earlier stage does not by itself add visits at that stage. Capacity is a real limit, and not a rare one. In the ADA Health Policy Institute's Q4 2025 poll, 12% of dentists said they were too busy to treat all the patients requesting care and 18% said they treated all patients but were overworked, while one third said they were not busy enough101. In its Q3 2026 poll nearly two in five said so102. Capacity does not make demand useless: a practice that is full can fill the gaps that cancellations leave, choose better-paying work, raise its minimum fee, or justify another provider, and each of those needs more inquiries than the practice books.

The second reason is a mismatch of measures. The stage that is cheapest to count, such as rankings or visits, is not always the stage that pays, so an improvement there improves the report and not the result. The third is interaction. A new website that promises a prompt reply raises the expectation the front desk must meet, which is one side of the gap between expectation and perception that Chapter 2 described, and a tracking tool added to count a stage can create a legal problem of its own.

Table 2 shows how a path with a count at every stage behaves when one stage limits it.

If this stage limits the pathWhat more visits from an earlier stage doWhat lifts the result
Contacts: the schedule has gaps and few people askAdd booked and kept appointments, one stage after anotherDemand, which acquisition produces
Handling: calls hang up or reach voicemail, and forms go unansweredAdd contacts that are lost at the door, because more arrive where the door does not openAnswering, booking and follow-up at the first contact
Attendance: booked visits are missedAdd booked visits that are missed at the same rateReminders, confirmations and filling the slot a cancellation opens
Capacity: the schedule is fullAdd no visits at that stage this month; they let the practice fill gaps, choose better-paying work, raise its minimum fee or justify another providerA better use of the capacity the practice has, or more of it

Table 2. Where a path limits and what more demand does there. A structural illustration with no figures; it describes no practice.

A fix pays most at the stage that limits the path, and a report of visits can improve without the result moving when the limit lies further along. Real constraints are softer than a hard cap, since callers wait, give up or try the next practice, but the direction of the effect is the same. A practice cannot know which row it is in without counting the path.

This paper does not measure any stage inside a practice; Chapter 1 gives the dental call figures. The reason to draw the pipeline before improving any part of it is the one the table shows: a fix at the wrong stage costs money, consumes the attention of the people who run the practice, and leaves the constraint where it was.

Marketing attribution and why the return is hard to see

Seeing what a stage returned is harder than it looks, even for firms with the means to run experiments. Lewis and Rao reported on twenty-five large field experiments with major US retailers and brokerages, most reaching millions of customers, which "reveal that measuring the returns to advertising is difficult"103. The median confidence interval on return on investment across their experiments was over 100 percentage points wide103. For the brokerage experiments the median standard error on return was 115%103. Individual-level sales are very volatile relative to the per capita cost of advertising, a coefficient of variation of 10 being common103. As a result "informative advertising experiments can easily require more than 10 million person-weeks, making experiments costly and potentially infeasible for many firms"103.

The observational alternative is no safer. Lewis and Rao concluded that "selection bias, due to the targeted nature of advertising, is a crippling concern for widely employed observational methods"103. Gordon, Zettelmeyer, Bhargava and Chapsky tested the point at Facebook, using 15 US advertising experiments with 500 million user-experiment observations and 1.6 billion ad impressions104. They found that the observational methods often fail to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables104. In 50% of their studies, the estimated percentage increase in purchase outcomes was off by a factor of three across all methods104.

Blake, Nosko and Tadelis report results from eBay. They found, in large field experiments, that returns from paid search are a fraction of non-experimental estimates and that brand keyword ads have no measurable short-term benefits105. When eBay turned off brand-keyword paid search, 99.5% of the forgone click traffic was immediately captured by unpaid search traffic from the platform105. Typical OLS regression methods gave a return on investment of over 4,100% without time and geographic controls and over 1,400% with them105. The experiment found a return of -63%, with a 95% confidence interval of -124% to -3%105. For non-brand keywords, new and infrequent users were positively influenced by the ads105. Frequent users, whose purchasing the ads did not influence, accounted for most of the expense, so that average returns were negative105.

Each result carries its scope: the studies concern retailers, brokerages, Facebook and eBay, firms whose advertising reached very large numbers of people, and none concerns a professional practice. We think they set a lower bound on difficulty: if firms with millions of customers cannot easily see what advertising returned, a practice with far fewer new patients or clients will see a noisier picture.

The eBay result also suggests a specific trap: a person who searches for a firm by name already knows it, and a report that credits the click to the ad credits the ad for a visit that would have happened anyway. A patient who is referred by a friend and then types the practice's name into a search box is in that position. This paper does not measure whether a local practice's branded searches behave like eBay's, and a practice's name is known to far fewer people, so the share that would have arrived anyway may differ by an amount these studies do not show.

For the definition, the lesson is that measurement must be designed. We think counts that a practice can observe directly, such as calls, bookings and where new patients or clients say they came from, are more dependable than a statistical attribution of credit among channels. Chapter 11 treats measurement in practice. The definition asks for a count at every stage and does not claim to isolate the causal return of each channel.

Measured management and the firms that do it

The management literature gives a reason to ask for measurement despite the difficulty, with limits. Brynjolfsson, Hitt and Kim used survey data on the business practices and information technology investments of 179 large publicly traded firms106. They found that firms adopting data-driven decision making have output and productivity 5-6% higher than would be expected given their other investments and information technology usage106. Bloom and Van Reenen collected management-practice data from 732 medium-sized manufacturing firms in the US, France, Germany and the UK107. They found management practices strongly associated with firm-level productivity, profitability, Tobin's Q, sales growth and survival107. They also found significant within-country differences, with a long tail of extremely badly managed firms107. A move from the lower to the upper quartile of management scores was associated with an increase in total factor productivity of between 3.2% and 7.5% under OLS107.

The samples are large publicly traded firms and medium-sized manufacturers, and both studies rest on survey data. We read the results as associations, since the passages we cite report associations and better-run firms may simply measure more. Neither concerns marketing decisions in particular. Brynjolfsson and McElheran, using Census Bureau data on a large representative sample of US manufacturing plants, found that the rate of data-driven decision making nearly tripled, from 11% to 30%, between 2005 and 2010108. They report that the key correlates of adoption are size, the presence of potential complements such as information technology and educated workers, and firm learning108.

Measured management is associated with better performance in firms that could measure, and adoption is associated with having the complements that make measurement possible and the learning that makes it useful. We think a small practice has few of either. It seldom has an analyst or a data team, and it has little time to learn a new reporting tool. The discipline's job is then to supply the complements: the counts, the connections and the routine of reading them, built so that the people who run the practice can operate them. This paper cites no study of data-driven management in dental, medical, legal or accounting practices, so this is a hypothesis about why the discipline fits.

From phonebook to funnel, from billboard to working rep

The paper's working image says that traditional web design treats a website like a billboard and Google like a phonebook, and that a technical marketing system turns Google into a funnel and the website into a working rep. This section is argument, and we offer no statistics about websites, templates or conversion in support of it.

A phonebook listing gave a name, an address and a number to a person who already knew what kind of provider they wanted. Stigler described the function in 1961: classified advertisements form a meeting place for potential buyers and sellers, and identifying buyers and sellers reduces drastically the cost of search58. Everything after the number was dialed happened on the telephone with the front desk, and the listing itself could not answer a question, show an open appointment or tell the office who was calling. For a practice, a listing was a cost of being findable, and the work of choosing and converting the caller fell largely on the person who picked up the telephone.

The brochure site is the billboard moved online. It describes the practice, shows photographs, lists services and names the staff. It can be well written and rank well, but it cannot answer the question this visitor arrived with, such as whether the practice takes a particular insurance or handles a particular kind of matter, offer the visitor a time, or tell the office who is coming and why. We think it also serves a machine reader poorly, because an assistant needs facts in a form it can quote and check, and a page built to impress seldom states them plainly. Chapter 2 gave the reason: a reader of online sources can use only signals that exist in a readable and checkable form.

Treating Google as a funnel asks more of the practice than being listed. A query expresses a need, a results page presents a choice among practices, a visit follows, and then come the practice's own stages. Google's description of local results, as its help page stood in November 2026, is that they are mainly based on relevance, distance and popularity100. It ties prominence to information such as how many websites link to a business and how many reviews it has100. We think this rests partly on what the practice and others publish about it, and a practice can supply and maintain much of that. The funnel view asks of each stage what comes in, what goes out and what is counted.

A working rep is the site conceived as a member of the front-office team. It answers the question the visitor arrived with, in words an assistant could also quote, and it offers a time or takes a request. It qualifies, which here means asking only the few things the office needs to route the inquiry, such as the service wanted, the location and the plan or the kind of matter. It then hands off to the office team with the context attached, so that the person does not start over. In our own service the rep may be an AI voice agent or a site assistant, and a caller or site visitor is always told that they are dealing with an AI assistant. This describes how we build and carries no claim about results.

Web design as system design: framing, plumbing and finish

The second half of the image says that web design starts with system design: framing and plumbing first and the finish fitted to the practice afterward, the way a house is built. Framing is the set of decisions that fix what the system must do: whom it serves, which questions it must answer, which actions count as success, which stages get a count, which rules bind each page and who answers what. Plumbing is the set of connections that make the framing real: listings and profiles that agree with one another, tracking events, forms and calls that reach a person, booking tied to the practice's schedule, and an agreement with every vendor that touches personal information. Finish is the design, copy and photography fitted to the practice.

We think the order matters because of what each layer costs to change. A page can be restyled without touching what lies behind it. Reconnecting a form to a different inbox, changing what a tracking tool may collect, or moving a booking system after a site has been built around the old one reaches into everything that depends on it, and the cost rises after the walls are closed. A house built the other way, with the paint chosen before the wiring, can look finished and still fail the first time someone turns on the lights. The same is true of a site designed before anyone has decided what it must do.

Finish still matters. In the terms of Chapter 2, it is where a buyer reads the dimensions of quality that can be observed in advance: the tangibles of a practice, the look of its people and premises, and the assurance conveyed by how credentials are presented are seen on the site, and a buyer weighs them. Finish comes third, and it is judged on whether it carries the facts and actions the framing specified. A beautiful page whose telephone number is wrong in several directories is a plumbing failure that the finish cannot repair.

Mapped onto the definition, framing is positioning and strategy. Plumbing is the connecting systems, tracking, intake and the listings that acquisition depends on. Finish is where conversion happens on the page. We think SEO, web design and technical systems pay when they are built together and measured as one system; the paper does not test this.

Why technical marketing fits dental, medical, law firm and accounting practices

Three features of the professions make the discipline fit. The first is size. Chapter 1 gave the evidence. As of 2025 one in three US dentists works in a single location with no other dentist13. In the AMA's 2024 survey, 47.4% of physicians worked in practices of 10 or fewer physicians14. By the Bureau of Labor Statistics counts, 88.3% of offices of lawyers and 86.2% of CPA offices had fewer than 10 employees in the first quarter of 2024 (calc., Chapter 1, Table 1). We infer that a firm of this size seldom has anyone whose job is marketing, from headcount; this is not measured here. The consequence is a design rule: the system must be operable by the people who run the practice, with few counts, each with a named owner, and with nothing that requires a department.

Regulated speech is the second feature. The rules reach every communication a practice makes. ABA Model Rule 7.1 is a model each state adopts in its own version. It says a lawyer's communication is false or misleading if it contains a material misrepresentation of fact or law, or omits a fact necessary to keep the statement, taken as a whole, from being materially misleading6. Chapters 6 to 9 set out each profession's rules. For design, a practice that states the same fact in many places needs one place where the approved statement lives and a way to change every copy when it changes. That is plumbing, and a creative or SEO engagement that does not include it leaves the practice to maintain consistency by hand.

The third is value per client, which brings in retention and lifetime value. Gupta, Lehmann and Stuart define the value of a customer as the expected sum of discounted future earnings109. In a valuation of five firms from public data, they found that a 1% improvement in retention, margin or acquisition cost improves firm value by 5%, 1% and 0.1% respectively109. They also report that improving retention by 1% is likely to improve customer and firm value by 3% to 7%, against a margin elasticity of about 1 and an acquisition elasticity of only 0.02 to 0.3109.

Two older statements point the same way. Reichheld and Sasser wrote in 1990, of service businesses, that "Reducing the defection rate just 5% generates 85% more profits in one bank's branch system, 50% more in an insurance brokerage, and 30% more in an auto-service chain"110. Berry, writing on services in 1995, called relationship marketing an old idea with a new focus at the forefront of services marketing practice and research111.

Gupta, Lehmann and Stuart's result is a valuation exercise on five firms, not an experiment. Reichheld and Sasser's is the authors' own statement about three service businesses, and the passage we cite does not say how its figures were obtained.

Retention figures for the four professions are fewer and weaker, and each has its limit. For primary care, a study of 799 patients aged 65 and over of 23 physicians in Oklahoma, with data from 1999 and 2000, found a median patient-physician relationship of 8 years112. For dentistry, Henry Schein One† (software) reports average patient retention of 70% against 94% for its top 10% of practices, without defining retention49. For accounting, a trade survey of firms, whose sample the report does not describe, found that 56% of their clients had been with them for 8 years or more51. The CPA.com and AICPA benchmark of client advisory practices reports a median client turnover of 10%113, and the National Association of Tax Professionals' fee study shows 87% returning clients among responding preparers114.

For law, the repeat share of a firm's clients is NF; the nearest evidence is a vendor survey of stated intentions. We estimate a ceiling on it: ~76 of every 100 recent lawyer-hirers would not rule out rehiring (100 less the 24% who said they were unlikely to rehire; likely high, because stated intentions overstate repeat business)52.

The argument from these results is structural and ours. Where a patient returns for recall visits, a client returns each year, or a family refers the next matter, the cost of winning a person is spread over a relationship. The question in such a practice then includes how many people are kept as well as how to acquire one more, and a system that counts returning patients or clients can see what an acquisition-only report cannot. For a profession whose work is a single matter, the argument is weaker, and the same system serves referral rather than recurrence.

What technical marketing is not

Technical marketing is not a tool. Our own check of 28 software tools that small practices commonly use found that 9 will sign a business associate agreement without requiring a special tier, 15 sign one only on certain plans and 4 offer none, according to the vendors' own pages54. A business associate agreement is the written arrangement through which, in HHS's description, a covered entity obtains satisfactory assurances that a vendor handling protected health information will safeguard it55. A practice that picks a tool first can find that the tool conflicts with a requirement it had not yet written. Whether a practice may use a tool that offers no such agreement, and for what, is for its own counsel or compliance lead. The management studies above point the same way, since adoption correlates with complements and learning, and software alone does not supply the routine of reading it.

It is not a channel either. SEO, paid search, listings, reviews and referral are channels, ways of reaching people. The system is what connects the channels to outcomes and lets the owner ask whether a channel pays. Google states that there is no way to request or pay for a better local ranking100. We infer that for local results the work lies in the facts Google reads, not in a purchase. For the channels that can be bought, the eBay result above is a reason to treat any claim that a purchased channel pays as a question to be tested.

Nor is it a guarantee of an outcome. We promise no ranking, no place in any assistant's answer, no fixed number of leads and no revenue. Two of our plans carry a narrower guarantee, on a measure agreed after an audit and set against the practice's own baseline, with more work as the remedy if it is missed56. Google says there are no additional technical requirements for its AI features beyond being indexed and eligible to appear with a snippet, and that existing SEO fundamentals continue to be worthwhile32. The paper that introduced generative engine optimization reports that the efficacy of its strategies varies across domains and that content creators have little to no control over how engines display their content31. In our study, only 3.5% of the 2,414 businesses named on the 160 questions asked in all 16 cities were named by all three assistants, from questions asked once1. We expect assistants to change their answers over time. A system makes a practice's facts findable, checkable and actionable and makes the result countable; it does not control what a search engine, an assistant or a patient does.

Finally, the discipline does not make a practice compliant, good or busy. Signals describe quality and do not create it. Whether a given page, message or tool meets a rule is for the practice's counsel or compliance lead. We have an interest in the argument, as Chapter 1 declared, so we state the definition in a form a reader can hold any provider to, including us: an owner, an input, an output and a count at every stage, and a system someone actually operates.

Chapter 4

How AI assistants recommend dentists, doctors, lawyers and accountants

The question and the design of the study

Ask three AI assistants the same local question and they mostly name different businesses. In our study, 199 questions put to ChatGPT, Gemini and Perplexity produced 2,975 business names. Of these, 122 (4.1%) were named by all three1. The study bears on the first part of the paper's through-line: whether a patient, a client or the assistant acting for them can find a practice.

On 2026-11-10 we put 199 local questions to the three assistants, across 16 US cities and 23 kinds of business, and collected 597 answers through each assistant's API with web search on1. Every question had the same form, with the city and the kind of business changed: "Who is a good dentist in Atlanta, GA? Name specific businesses."1. The wording asks for names, the behavior we set out to measure. It asks for a good business without saying by what standard, which we think is how a person with no other information would ask.

The sample: which kinds of business, in how many cities

Ten kinds of business were asked about in every one of the 16 cities, which makes 160 questions1. The ten are dentists, chiropractors, med spas, family law attorneys, accountants, financial advisors, real estate agents, roofers, plumbers and HVAC companies1. The other 13 kinds were asked about in three cities each, so each of those rows rests on three questions and nine answers1. The counts reconcile: 13 kinds in 3 cities give 39 questions, and 160 plus 39 is 199.

The split matters because the professional kinds this paper covers fall on both sides of it. Dentists, chiropractors, med spas, family law attorneys, accountants and financial advisors were asked about in all 16 cities. Personal injury lawyers, estate planning attorneys, physical therapists and bookkeepers were asked about in three1. Where we report the three-city rows we flag them as small samples and we do not lean on them. Our study did not ask about physicians. The nearest health rows are chiropractors, physical therapists and med spas, which we do not stretch to cover medical practices.

What the measures mean

A business counts as named when it appears among the names read from an answer, up to eight per answer. It is counted once per question however many assistants name it1. Two names count as the same business when they match by spelling1. "Named by all three" means that all three assistants named the same business in answer to the same question, and "named by only one" means that the other two did not. A website counts as cited when an answer's citations include it, and it is counted once per answer1.

Four features of the method qualify every figure below. Each question was asked once, names were matched across assistants by spelling, up to eight names were read per answer, and answers obtained through an API are a baseline, not what every signed-in user sees1.

How often the three assistants agree on which professional to name

Table 1 sets out agreement across all 199 questions.

MeasureResultSource
Questions asked1991
Answers collected5971
Businesses named, once per question2,9751
Named by all three assistants122 (4.1%)1
Named by exactly two assistantsabout 16.2%, roughly 480 businesses (calc.)1
Named by only one assistant79.7%1
Questions in which no business was named by all three108 of 199 (54%)1
Questions in which at least one business was named by all three91 of 199 (46%; 199 - 108)1
Questions in which all three named the same business first9 of 199 (4.5%; 9 / 199)1
Questions in which at least two agreed on the first name75 of 199 (37.7%; 75 / 199)1
Questions in which no two named the same business first124 of 199 (62.3%; 199 - 75)1
Overlap between any two assistants11.9% to 14.3% of the businesses they named1

Table 1. Agreement among ChatGPT, Gemini and Perplexity across all 199 questions, from our study.

Agreement is the exception. Of the 2,975 businesses named, 79.7% were named by only one assistant and 4.1% by all three1. A practice that appears in one assistant's answer is therefore most likely to be absent from the other two.

Counted by question, the picture is the same. In 108 of the 199 questions (54%) no business at all was named by all three assistants1. The first name is the strictest test of agreement, because each assistant must commit to a single business. All three assistants named the same business first in 9 of the 199 questions, and at least two agreed in 75 (4.5% and 37.7%; calc., Table 1)1.

Taken two at a time, the assistants overlapped on between 11.9% and 14.3% of the businesses they named1. The range is narrow, so no pair of assistants was much closer to agreement than any other, and we think that is a reason to treat the three as separate channels. The assistants also write lists of different lengths: ChatGPT named 4.8 businesses per answer, Gemini 6 and Perplexity 7.81. A pairwise overlap therefore mixes agreement with list length, and we do not read small differences between pairs.

The headline figure depends on the mix of questions

The 4.1% figure pools two groups of questions of different depth. For the ten kinds asked about in all 16 cities, the assistants named 2,414 businesses, and 84 of them (3.5%) were named by all three1. The 13 kinds asked about in three cities account for the other 561 names (2,975 - 2,414) and 38 businesses named by all three (122 - 84)1. That is 6.8% for the small-sample rows against 3.5% for the large ones, so the three-city rows pull the pooled figure up from 3.5% to 4.1%1. When this paper needs one number for how often three assistants agree on a local business, it uses the 3.5% from the 160 questions asked everywhere1.

Agreement also moves with the city. For the ten kinds asked about everywhere, the share of named businesses that all three assistants named ranged from 1.9% in Charlotte to 7.9% in Seattle1. That is a spread of 6.0 percentage points, about fourfold, inside one design and one set of questions. It serves as a rough yardstick for the next section: a difference between two professions that is smaller than the difference between two cities should not carry much weight.

Agreement by profession: dentists, accountants, lawyers and the rest

Table 2 gives the figures for each profession in the study, with roofers, plumbers, HVAC companies and real estate agents added for comparison.

Kind of businessCitiesBusinesses namedPer questionNamed by all threeNamed by one onlySource
Dentists1626616.63 (1.1%)88%1
Chiropractors1622814.310 (4.4%)80.3%1
Med spas1622814.313 (5.7%)79.4%1
Physical therapists3, small sample51171 (2%)78.4%1
Family law attorneys162401513 (5.4%)75.8%1
Personal injury lawyers3, small sample4615.33 (6.5%)76.1%1
Estate planning attorneys3, small sample4715.71 (2.1%)83%1
Accountants16272170 (0%)90.4%1
Bookkeepers3, small sample5016.73 (6%)84%1
Financial advisors1628117.65 (1.8%)87.5%1
Real estate agents (comparison)1625916.24 (1.5%)83%1
Roofers (comparison)1621613.515 (6.9%)70.4%1
Plumbers (comparison)162081311 (5.3%)70.7%1
HVAC companies (comparison)1621613.510 (4.6%)72.7%1

Table 2. Agreement by kind of business, from our study. Rows marked "small sample" rest on three questions each.

Dental and health rows

Among the health rows, dentists have the lowest agreement. The assistants named 266 dentists across the 16 cities, and 3 of them (1.1%) were named by all three, while 88% were named by only one assistant1. Chiropractors reached 4.4% and med spas 5.7%1. Physical therapists, asked about in three cities only, show 1 of 51 (2%), which we report for completeness and do not interpret1. The dentist row says that, in this run, a patient who asked three assistants for a good dentist in a practice's city would rarely see the same practice in all three answers.

Family law attorneys, the one legal kind asked about in all 16 cities, reached 13 of 240 (5.4%), with 75.8% named by one assistant only1. The two legal kinds asked about in three cities show 6.5% for personal injury lawyers and 2.1% for estate planning attorneys1. The gap between them, 4.4 percentage points (6.5% - 2.1%), is the difference between 3 businesses named by all three and 1. With three questions behind each row, the two legal figures cannot be told apart from each other or from the family law figure.

Accounting and finance rows

The accountant row is the starkest in the study. Across 16 cities the assistants named 272 accounting businesses, 17 per question, and none was named by all three. Of these, 90.4% were named by one assistant only, the highest such share in Table 21. Financial advisors look similar, with 5 of 281 (1.8%) named by all three and 87.5% by one only1. Bookkeepers, asked about in three cities, show 3 of 50 (6%), which says little1.

Professions against trades

The six professional kinds asked about everywhere are dentists, chiropractors, med spas, family law attorneys, accountants and financial advisors. Pooled, 44 of 1,515 names (2.9%) were named by all three (calc., Table 2). For the three trades (roofers, plumbers and HVAC companies) the pooled figure is 36 of 640 (5.6%; calc., Table 2). The share named by one assistant only runs from 75.8% to 90.4% across the six professional kinds and from 70.4% to 72.7% across the three trades (Table 2). The grouping is ours, and the comparison is descriptive.

The division is not clean. In Table 2, med spas (5.7%) and family law attorneys (5.4%) sit with or above plumbers (5.3%) and HVAC companies (4.6%). Real estate agents (1.5%), which we did not count among the professional kinds, sit with dentists and accountants (Table 2).

What does line up is list length. Four 16-city kinds have 16 or more names per question: dentists, accountants, financial advisors and real estate agents. In Table 2 they have the four lowest all-three shares, 0% to 1.8%, and one-assistant-only shares from 83% to 90.4%. The six kinds with 15 or fewer names per question have all-three shares from 4.4% to 6.9% and one-only shares from 70.4% to 80.3% in Table 2. The two groups do not overlap on either measure. This is a pattern across ten rows, with no test behind it. It may be partly mechanical, because the more distinct names the assistants produce between them, the less likely it is that all three name the same one. It may also reflect a larger pool of plausible practices in a city, or more divergent sources. The data cannot separate these.

How much a row can bear

A row in Table 2 is one question per city, asked once. The city-to-city spread of 6.0 percentage points (7.9% - 1.9%)1 is larger than every gap between neighboring 16-city rows, so we do not rank neighboring professions against each other. And in a three-city row of 46 names, one more business named by all three assistants would move the share by 2.2 percentage points (1 / 46), so those rows can move a long way on a single name. What the 16-city rows support is a statement about all of them together: agreement is low for every profession measured. Accountants, dentists and financial advisors had the lowest figures among the professional kinds in this run, but by the yardstick above the gaps are too small to build on.

What the assistants cite: directories, review platforms and Reddit

The assistants also cite websites, and cited sources are something a practice can inspect. The 597 answers cited 1,455 different websites, for 3,530 citations in all1, which is 5.9 citations per answer (3,530 / 597). As with names, the assistants did not converge: 88% of the websites cited were cited by only one of the three assistants1.

Assistant (model)Names per answerSites cited per answerAnswers citing RedditMost-cited sites, answers of 199Source
ChatGPT (gpt-6-luna)4.83.52%birdeye.com 32; bestprosintown.com 30; expertise.com 19; yably.com 19; contractorsranked.com 18; bestinhood.com 141
Gemini (gemini-3.8-flash)66.542%reddit.com 83; consumeraffairs.com 27; threebestrated.com 24; forbes.com 21; bestpickreports.com 14; angi.com 131
Perplexity (sonar)7.87.70%bbb.org 58; bestinhood.com 54; threebestrated.com 52; consumeraffairs.com 38; thumbtack.com 36; yahoo.com 351

Table 3. What each assistant named and cited, from our study.

ChatGPT gave the shortest lists and cited the fewest sites, fewer than the businesses it named (3.5 against 4.8 per answer), where Gemini and Perplexity cited about one site per name (6.5 against 6, and 7.7 against 7.8)1. Perplexity cited 2.2 times as many sites per answer as ChatGPT (7.7 / 3.5) while naming 1.6 times as many businesses (7.8 / 4.8). ChatGPT may have drawn several names from one page, but the figures do not show it. The six most-cited sites of each assistant overlap very little. The site bestinhood.com appears in ChatGPT's list and Perplexity's, and threebestrated.com and consumeraffairs.com appear in Gemini's and Perplexity's, but no site appears in all three1. The 18 places in the three lists are held by 15 different sites (18 - 3 sites listed twice).

Reddit is one assistant's habit

Reddit gives the clearest single-site result. Gemini cited Reddit in 42% of its answers, ChatGPT in 2% and Perplexity in 0%1. In a pooled ranking Reddit comes first, with 87 answers1. Gemini alone cited it in 83 of its 199 answers1, so 83 of the 87 (95%) are Gemini's. A pooled ranking of sources hides a large difference: for one assistant a forum is a regular source for local recommendations, and for the other two it is almost never one. In this run, whether a practice's reputation in forum threads reaches a patient can depend on which assistant the patient asks.

SiteAnswers citing it (of 597)Share of answers (calc.)Source
reddit.com8714.6%1
threebestrated.com7612.7%1
consumeraffairs.com7111.9%1
bestinhood.com6811.4%1
bestprosintown.com6410.7%1
bbb.org6310.6%1
expertise.com559.2%1
birdeye.com457.5%1
zocdoc.com447.4%1
forbes.com427.0%1

Table 4. The ten most-cited websites across all three assistants and all 23 kinds of business.

National sites, local sites and the practice's own site

The most-cited sites are not local. The 25 most-cited websites drew 29% of all citations, and each was cited in at least 9 of the 16 cities. They are directories, review platforms, list sites, publishers and a forum, not any one local business's site1. That is about 1,020 of the 3,530 citations (29% x 3,530)1. At the other end, websites cited in only one city drew 39% of citations1, about 1,380 citations, and the remaining 32% went to sites cited in more than one city that are not among the 25.

A local business's own website would fall among the single-city sites, but so would a local publisher, a regional directory or a municipal page, and the figures do not say what share are practices' own sites. A second limit applies to the whole of Table 4: the site counts pool all 23 kinds of business, including trades, so they cannot show which sites matter for dentists as opposed to roofers. We have not broken them down by kind.

How the results fit the research on generative engines

The paper that introduced generative engine optimization reports that its methods can raise a source's visibility in generative engine responses by up to 40%31. It also reports that the effectiveness of its strategies varies across domains31. Its authors describe generative engines as black boxes that leave content creators with little to no control over when and how their content is displayed31. Our study tests none of those methods. It measured what three assistants answered to a local question, not whether any intervention changed a source's visibility. The figure of up to 40%31 is a best case in the authors' own setting, which we do not carry over to local professional recommendations.

Our study bears on the black-box point in one respect. If creators have little control over how content is displayed, a practice cannot assume that its position in one engine carries over to another. The finding that 79.7% of names and 88% of cited sites appeared in one assistant only fits that1. It does not establish the point, because the same figures would arise if each assistant's answers varied widely from run to run, and each question was asked once1. An outside study, discussed below, reports such variation for other prompts.

Google's own guidance says that a page must be indexed and eligible to be shown in Google Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements32. It adds that no specific optimization is required and that existing SEO fundamentals remain worthwhile32. That guidance concerns Google's own AI features. Our study did not query AI Overviews or AI Mode. It queried three assistants through their APIs with web search on, so the guidance and our figures describe different systems and neither confirms the other.

The GEO paper and Google's guidance do not conflict: one reports gains from specific methods in a research setting, and the other says Google's AI features require nothing beyond sound fundamentals. We think it is an open question how much an optimization beyond fundamentals buys, in a given assistant, for a small professional practice. Chapter 5 covers the vocabulary around the idea, including that the phrase "generative engine optimization" is searched about 4,400 times a month in the US27.

Reading the results for dental marketing, law firm marketing and CPA marketing

This section is interpretation, and nothing in it is a finding of the study. Each point names the finding it rests on, so that a reader who disagrees can see where.

Each assistant is its own channel. This rests on the 79.7% of names and 88% of cited sites that appeared in one assistant only, and on pairwise overlaps no higher than 14.3%1. A practice named by one assistant has learned something about that assistant and little about the other two. For any single question, checking three assistants is a different exercise from checking one assistant three times.

Third-party profiles carry weight. This rests on the 25 most-cited sites drawing 29% of citations and being directories, review platforms, list sites, publishers and a forum1. It suggests that the facts a practice states on those sites (name, address, phone, hours, services, the people who work there) are among the sources the assistants read most. Keeping them accurate and consistent is work a practice controls. The study does not show that a cited profile causes a name to appear, since it counts names and sites separately. The single-city sites, 39% of citations1, are where the practice's own pages would fall, so the practice's own site is in the mix too and has to be readable on its own.

One check proves little. This rests on each question having been asked once and on the 6.0-point spread between cities1. A practice owner who types a query into one assistant on one day and sees the practice named, or missing, has one draw from a process we have not shown to be stable. The minimum useful check is dated, repeated, covers all three assistants and uses many questions, and Chapter 11 sets one out.

Being named is not a measure of quality, and low agreement is not evidence that any assistant is wrong. Three assistants can differ because each selects, correctly by its own sources, from a large pool of competent practices, because their sources differ, or because their answers vary from run to run; the study cannot tell these apart. Agreement would not show correctness either, since three assistants that cite the same list site would agree and could all be wrong. The profession's advertising rules, covered in Chapters 6 to 9, and the practice's own counsel or compliance lead decide how a practice may describe an assistant's recommendation in its own advertising.

The consequence we draw is an argument, not a result. If the assistants read different sources and write different lists, then being found is a matter of being legible to several readers: consistent facts across the places they read, signals a reader can verify, and pages that answer the specific questions people ask. That is engineering work, and it is the work this paper calls technical marketing. We do not promise a place in any assistant's answer, and nothing in these figures supports such a promise.

Limits of the study, and what a stronger study would do

One draw on one day

Each question was asked once, on 2026-11-101. We did not measure how far an assistant disagrees with itself when asked the same question twice, so part of the disagreement between assistants may be ordinary run-to-run variation, which tends to lower measured overlap. We expect assistants to change their answers over time, so the figures describe one date.

An outside study bears on this limit. SparkToro† and Gumshoe.ai†, in a study published on 27 January 2026, had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in all115. SparkToro† and Gumshoe.ai† reported a less than 1 in 100 chance that ChatGPT or Google's AI, asked a prompt 100 times, would give the same list of brands in any two responses115. The same study reported that the same list in the same order came up less than 1 in 1,000 times, even in spaces with limited options such as Los Angeles Volvo dealers115. The study is not peer reviewed. Its author discloses a possible conflict of interest, since his research partner works at Gumshoe.ai†. Its prompts were not about dentists, doctors, lawyers or accountants, and its volunteers pasted answers by hand, where we used each assistant's programming interface.

We think it supports the direction of our argument but gives no rate for our questions. Of the names on the 160 questions asked everywhere, 96.5% (calc.) were named by fewer than all three assistants1, and some unknown part of that is run-to-run variation. We estimate that the measured agreement of 3.5%1 is a floor and the true agreement higher (from SparkToro†'s large variation for brand lists; the measured figure likely errs low). That variation may be smaller for a question with web search on and a narrow local pool of practices than for SparkToro†'s prompts. A repeat-asking study of AI recommendations for local businesses or professional practices is NF; the nearest is SparkToro†'s.

Matching and truncation

Matching by spelling can split one business into two when the assistants write its name differently, and it can merge two businesses that share a name1. We think the first error is likelier within a single city, so the effect on agreement is probably downward, by an amount we cannot state. Up to eight names were read per answer, and Perplexity averaged 7.8 names per answer1. That average sits just under the limit, so the limit may have shortened some of Perplexity's lists. Truncation tends to lower measured overlap for the assistant with the longest lists, and the figures do not tell us how many answers were cut.

Run-to-run variation, matching by spelling and truncation all tend to lower measured agreement, so the agreement we report is closer to a floor than to a central estimate. We do not claim that the floor is near the true rate.

The API baseline and the choice of models

A person using a consumer app, signed in, with a location and a history, may be shown something different from the API baseline, and we did not test that1. Each assistant was represented by one model1, named in Table 3, and a different model from the same company could answer differently. The model names are part of the record because they will date.

Sample and wording

The study covers 16 cities and 23 kinds of business, and we make no claim that the 16 cities are a random sample of US cities1. Each kind of business was asked about in one wording. We expect that patients and clients also ask for a specialty, an insurance plan, an emergency appointment or a language, and those questions may behave differently.

What the figures do not measure

Names and cited sites are counted separately, so the figures do not trace a name to the page behind it. They say nothing about whether the businesses named exist, are licensed, are open, accept new patients or clients, or fit the question, about whether a cited page supports what the answer says, or about whether any answer led to an appointment or a matter. Nothing here measures the quality of care or advice a practice gives.

What a stronger study would do

A stronger study would begin by measuring each assistant against itself: the same question many times, on many days, to separate real disagreement from variation, as SparkToro† and Gumshoe.ai† did for brand prompts. It would capture answers from the consumer apps, signed in and signed out, as well as from the APIs. It would match names by address and telephone number and read every name given. It would use several wordings for each kind of business, add physicians, and draw the cities at random. It would check named businesses against an independent register of practitioners, classify every cited site by type, report them by profession and link each name to the page that supplied it. These are proposals for work we have not done.

Chapter 5

What practices and their clients search for

The data and what it can and cannot show

This chapter reads the demand for services to practices from one source: Google Ads keyword data on US monthly searches, collected in November 202627. The data show how often people typed particular phrases, and the tables set the dental, medical, legal and accounting phrases side by side. The main finding is the gap set out in Table 5: in legal and accounting, the leading AI phrase is searched about five times as often as the leading marketing phrase, in medical somewhat more often, and in dentistry less often.

The phrases are the ones we chose to look up: phrases an owner, a practice manager or a vendor would type, such as "dental marketing" or "legal answering service". None is a phrase a patient or client would type to find a practice, so the demand described here is demand for services to practices. What patients and clients ask when they look for a practice is the subject of Chapter 4, where the questions took that form. A list we assembled is not a census: a phrase missing from our tables may be searched far more than one that appears in them, and every comparison below is between phrases we looked up.

Rounded bands and shared figures

Each volume is an average monthly figure, and the figures sit on a short ladder of values. From 590 to 2,400 the steps are 590, 720, 880, 1,000, 1,300, 1,600, 1,900 and 2,400. The figures for "medical marketing", "dental advertising", "dental ai", "dental seo", "dental marketing", "healthcare marketing", "ai for law firms" and "ai for lawyers" sit on those steps27. Each step is between 14% and 30% above the one before (1,000 / 880 is the smallest, 1,300 / 1,000 the largest). We treat each figure as a rounded band and two phrases one step apart as indistinguishable.

Close variants share a figure. "Dental marketing" and "dentist marketing" both show 1,300 a month at an average cost per click of $47.3027. "Attorney seo" and "lawyer seo" both show 2,90027, and "ai accounting", "ai for accountants" and "ai in accounting" each show 6,60027. We think the keyword tool reports one figure for phrases it treats as the same search. We never add the volumes of variants together, which would count one pool of searches two or three times. The tables list variants on one row, except where volumes match but costs per click differ: "dental seo" and "dental seo services" both show 1,000 a month, at $44.79 and $48.6727, so they have separate rows.

The second figure in each table is cost per click: the average amount an advertiser pays Google each time someone clicks its advertisement for the phrase, as the keyword data reports it. Competition among advertisers sets the price, so it signals what a visitor from that phrase is worth to the advertisers bidding. Where the keyword data returned no cost-per-click figure, the tables say "no figure".

Interest is not purchase

Search volume measures interest, not purchase: a phrase may be typed by an owner ready to hire, a student, a journalist, a vendor watching a competitor or someone who meant something else, and the data cannot separate them. The data cover one month, so they show level, not direction, and we cannot say whether any phrase is rising or falling. And we sell an AI voice agent56, marketing plans56 and HIPAA-ready programs57, and a reader should weigh our reading of the numbers with that in mind.

The phrase for the discipline this paper describes is small: "technical marketing" averages about 210 searches a month and "what is technical marketing" about 5027. We do not argue from demand for the name. The argument is that the parts of the system (answering the telephone, being found by an assistant, tracking what happened) are each searched on their own, and that the system joins them.

Dental marketing, dental SEO, dental AI and the dental answering service

Table 1 lists the dental phrases we looked up, by type.

PhraseTypeSearches a monthCost per clickSource
dental marketing; dentist marketingMarketing1,300$47.3027
dental seoSEO1,000$44.7927
dental seo servicesSEO1,000$48.6727
dental aiAI880$27.6927
dental advertisingMarketing720$41.2827
dental marketing companyMarketing720$66.1127
local seo for dentistsSEO720no figure27
dental marketing agencyMarketing590$50.2527
dental website designWebsite590$62.3927
dental practice managementPractice software320$24.1227
dental practice marketingMarketing210$50.1127
ai dental receptionistAI, front desk170$152.0827
dental office marketingMarketing170$106.8527
ai for dentistsAI140$28.6427
dental marketing ideasMarketing140$16.9027
dental answering serviceAnswering service110$181.6827
how to get more dental patientsPatient acquisition50$29.2527
dental patient acquisitionPatient acquisition10$35.7327

Table 1. Dental search phrases: US searches a month and average cost per click, Google Ads keyword data, November 2026.

The largest dental phrases are about marketing: "dental marketing" and its variant "dentist marketing" show 1,300 searches a month as one figure, and "dental seo" shows 1,00027. The largest AI phrase, "dental ai", shows 88027, which is 0.68 of the leading marketing phrase (880 / 1,300). Dentistry is therefore the one profession in our set where the leading marketing phrase outruns the leading AI phrase, the exception to the pattern in the other three.

The AI and front-desk phrases are small in volume. "Ai dental receptionist" averages 170 searches a month, "ai for dentists" 140 and "dental answering service" 11027. The last carries the highest cost per click in the dental list, $181.68, followed by "ai dental receptionist" at $152.0827. We think the dental list says that the specific front-desk phrases are small but worth a lot to the advertisers who pay for them, and that dentists, or the vendors selling to them, ask for marketing by name.

Healthcare marketing, medical practice marketing and HIPAA compliant AI

Table 2 lists the medical and healthcare phrases. Phrases with "medical" and phrases with "healthcare" both appear, and we do not merge them.

PhraseTypeSearches a monthCost per clickSource
hipaa compliant aiAI, compliance6,600$41.8827
healthcare marketing agencyMarketing5,400$25.1727
medical answering serviceAnswering service5,400$35.9227
medical marketing agencyMarketing4,400$42.1627
healthcare marketingMarketing1,600$25.4827
healthcare crmPractice software720$42.9927
healthcare advertisingMarketing590$27.0527
medical marketingMarketing590$22.2127
patient engagementPatient communication590$12.1027
medical seoSEO390$24.6827
medical spa marketingMarketing390$77.6127
medical website designWebsite390$69.7627
healthcare seoSEO320$64.4827
medical practice marketingMarketing320$87.7627
medical practice seoSEO170$27.8127
doctor marketingMarketing170no figure27
ai medical receptionistAI, front desk140$80.7627
physician marketingMarketing90no figure27
hipaa marketing rulesCompliance70$1.5727
hipaa compliant marketingCompliance40$26.6527

Table 2. Medical and healthcare search phrases: US searches a month and average cost per click, Google Ads keyword data, November 2026.

The medical list is led by a phrase about privacy: "hipaa compliant ai" averages 6,600 searches a month, with "healthcare marketing agency" and "medical answering service" tied behind it at 5,400 each27. The plain marketing phrases are smaller, from 1,600 for "healthcare marketing" to 320 for "medical practice marketing"27. "Medical answering service" is 3.4 times "healthcare marketing" (5,400 / 1,600) and "hipaa compliant ai" is 4.1 times it (6,600 / 1,600). Against the agency phrases the picture is closer: "healthcare marketing agency" equals "medical answering service" at 5,400, and "hipaa compliant ai" is 1.2 times it. We give both comparisons because the choice of marketing phrase changes the ratio.

Compliance phrases show a gap of their own. "Hipaa compliant marketing" averages 40 searches a month and "hipaa marketing rules" 70, against 6,600 for "hipaa compliant ai"27. That is 165 times (6,600 / 40). "Hipaa marketing rules" also carries the lowest cost per click in our data, $1.57, while "medical practice marketing" costs $87.76 and "medical spa marketing" $77.6127. We infer that a phrase that asks what a rule says draws people whom advertisers do not expect to buy, and a phrase that names a service draws people whom they do.

We use the phrase "hipaa compliant ai" because it is what people type. We do not describe any tool as compliant, because compliance depends on how a tool is set up and on a signed agreement. HHS describes a business associate as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information55. The HIPAA Rules permit a covered entity to disclose protected health information to a business associate if it obtains satisfactory assurances, in a contract or other written arrangement, that the business associate will appropriately safeguard the information55.

Someone searching for the phrase meets a market in which that agreement is not uniform. Of 28 software tools that small practices commonly use, 9 will sign a business associate agreement without requiring a special tier, 15 sign one only on certain plans and 4 offer none. These counts come from the vendors' own pages54. Whether a particular set-up meets the rules is for the practice's own counsel or compliance lead to decide.

Table 3 lists the legal phrases.

PhraseTypeSearches a monthCost per clickSource
legal aiAI18,100$31.6227
law firm seoSEO3,600$84.5327
attorney seo; lawyer seoSEO2,900$108.7827
ai for lawyersAI2,400$58.2127
ai for law firmsAI1,900$33.8727
law firm website designWebsite1,600$96.0627
law firm marketingMarketing1,300$75.6827
law firm marketing agencyMarketing880$135.9727
legal seoSEO880$99.4027
attorney marketing; lawyer marketingMarketing720$134.6627
legal answering serviceAnswering service720$214.5127
legal marketingMarketing480$81.1427
law firm crmPractice software390$108.0427
personal injury marketingMarketing170$105.5727
legal intake softwareIntake140$106.4427
law firm lead generationMarketing90$137.2027
law firm intakeIntake70$92.4127
attorney advertising rulesCompliance70no figure27

Table 3. Legal search phrases: US searches a month and average cost per click, Google Ads keyword data, November 2026.

"Legal ai" is the largest profession-specific phrase in our data, at 18,100 searches a month27. It outruns the leading legal marketing phrases by a wide margin: "law firm seo" averages 3,600, "attorney seo" and "lawyer seo" share a figure of 2,900, and "law firm marketing" averages 1,30027. "Legal ai" is 5.0 times "law firm seo" (18,100 / 3,600) and 13.9 times "law firm marketing" (18,100 / 1,300).

Several legal phrases sit near the top of the cost-per-click ranking. "Legal answering service" costs $214.51 a click, the highest figure in our data27. Among the phrases in Table 3, "law firm lead generation" follows at $137.20, "law firm marketing agency" at $135.97 and "attorney marketing" and "lawyer marketing" at $134.6627. The phrases about intake, the point where a call becomes a matter, are small: "legal intake software" averages 140 searches a month and "law firm intake" 7027. Chapter 8 treats intake as the conversion problem for law firms, and we think the small volume of the intake phrases suggests that firms do not yet name the problem that way when they search.

Accounting firm marketing, CPA marketing and AI for accountants

Table 4 lists the accounting phrases.

PhraseTypeSearches a monthCost per clickSource
ai accounting; ai for accountants; ai in accountingAI6,600$38.9527
cpa marketingMarketing1,300$12.3127
accountant marketing; accounting marketing; marketing for accountantsMarketing260$24.9727
accounting firm marketingMarketing170$42.1027
seo for accountantsSEO170$10.0827
accounting practice managementPractice software90$47.2027
ai for cpasAI90$49.3327
accounting firm website designWebsite70$13.2227
bookkeeping marketingMarketing70$27.6327
accounting lead generationMarketing70no figure27
cpa firm marketingMarketing50$52.6827
how to get accounting clientsClient acquisition50$7.8527
cpa seoSEO40no figure27
tax preparer marketingMarketing30$39.1927
accounting firm seoSEO10no figure27
how to get cpa clientsClient acquisition10$4.7827

Table 4. Accounting search phrases: US searches a month and average cost per click, Google Ads keyword data, November 2026.

One phrase dominates the accounting list: "ai accounting", "ai for accountants" and "ai in accounting" each show 6,600 searches a month at $38.95 a click, one figure for the three variants27.

The largest marketing phrase by volume, "cpa marketing", averages 1,300 at $12.31, and the "accountant marketing" variants share a figure of 260 at $24.9727. The larger figure is not a clean count of interest in marketing a CPA firm. CPA is also an advertising abbreviation, for cost per action. Our long-tail data shows searches that ask about that meaning: "what is cpa marketing" and "cpa meaning marketing" each average 720 a month, and "cpa marketing network" 48027. The keyword data do not show what share of the 1,300 searches for "cpa marketing" has nothing to do with accountants, though the long-tail phrases just named suggest that it is not zero27. The AI phrase is 5.1 times "cpa marketing" (6,600 / 1,300) and about 25 times the accountant-marketing variants (6,600 / 260). Only the second ratio is free of that ambiguity.

The phrases that ask how to win clients are tiny: "how to get accounting clients" averages 50 and "how to get cpa clients" 1027. Chapter 9 argues that an accounting firm markets for fit, not volume, because its scarce resource is professional time. Small demand for marketing phrases would be consistent with that argument, but the search data cannot test it: a firm that does not seek volume and a firm that does not know what to type look the same in a keyword count.

Table 5 sets each profession's highest marketing or SEO phrase beside its highest AI phrase and its answering-service phrase, with the ratio of the two volumes (calc.).

ProfessionTop marketing or SEO phrase, a monthTop AI phrase, a monthAI divided by marketingAnswering-service phrase, a monthSource
Dentaldental marketing, 1,300dental ai, 8800.68dental answering service, 11027
Medicalhealthcare marketing agency, 5,400hipaa compliant ai, 6,6001.22medical answering service, 5,40027
Legallaw firm seo, 3,600legal ai, 18,1005.03legal answering service, 72027
Accountingcpa marketing, 1,300 (ambiguous); accountant marketing, 260ai accounting, 6,6005.08, or 25.38 against 260none looked up27

Table 5. The leading marketing and AI phrases by profession, from the figures in Tables 1 to 4.

The pattern is common, not universal. In accounting the ratio against "cpa marketing" is ambiguous, but against the unambiguous accountant-marketing figure the AI phrase is about 25 times larger (calc., Table 5). In medical the AI phrase is 1.2 times the leading marketing phrase27, the least certain comparison, because "hipaa compliant ai" names a privacy rule and not a profession, and we did not look up a phrase such as "medical ai". Dentistry is the exception. The answering-service phrases differ even more widely between professions: "medical answering service" averages 5,400 searches a month, 49 times "dental answering service" at 110 and 7.5 times "legal answering service" at 72027.

On the figures we have, the gap is larger when no profession is attached: "ai receptionist" averages 74,000 searches a month, "ai agent" 49,500 and "ai voice agent" 33,10027. The two largest marketing-discipline phrases in our general list are "marketing funnel" at 9,900 and "martech" at 6,60027. "Ai receptionist" is 7.5 times "marketing funnel" (74,000 / 9,900). Adding a profession collapses the volume: "ai dental receptionist" averages 170 and "ai medical receptionist" 14027, so the generic phrase is about 435 and 529 times the specific ones. The generic phrase is the largest in our data, but the data do not show how much of that demand comes from practices, and the profession-specific forms carry only a small fraction of it.

What the gap might mean, and what else could explain it

One reading is that an owner who sees a problem at the front desk (missed calls, slow answers, a full inbox) searches for the tool that seems to fix it, an AI receptionist or an answering service. Such an owner does not search for "marketing", because marketing names a different problem. If that is right, many owners think they are buying a tool. The argument of this paper is that the tool is one stage of a system. The call it answers was produced by a search, the booking it makes has to be counted, and what it says is bound by the profession's rules.

The same figures fit other explanations, and the data cannot choose among them. The searchers may not be owners: "legal ai" may be typed by law students, developers, journalists or lawyers at large firms. Many searches may come from people learning what the technology is, which is not buying. Marketing demand may sit under phrases we did not look up, such as an agency's name or a request for a website. Interest in AI may also reflect news attention, not intent. A survey of owners about what they believe they are buying would narrow the question, as would a practice's own count of which phrases bring visitors to its site and which of those visitors call. Chapter 11 returns to counting.

What a click costs: cost per click by profession

Table 6 puts the cost per click of comparable phrases side by side.

ProfessionMarketing phraseSEO phraseAnswering-service phraseAI phraseSource
Dentaldental marketing, $47.30dental seo, $44.79dental answering service, $181.68dental ai, $27.6927
Medicalmedical marketing, $22.21medical seo, $24.68medical answering service, $35.92hipaa compliant ai, $41.8827
Legallegal marketing, $81.14legal seo, $99.40legal answering service, $214.51legal ai, $31.6227
Accountingaccounting marketing, $24.97seo for accountants, $10.08none looked upai accounting, $38.9527

Table 6. Average cost per click for comparable phrases, by profession, Google Ads keyword data, November 2026.

Read down the marketing and SEO columns and the professions separate. "Legal marketing" costs 3.7 times "medical marketing" ($81.14 / $22.21) and 1.7 times "dental marketing" ($81.14 / $47.30)27. "Legal seo" costs 4.0 times "medical seo" ($99.40 / $24.68) and 2.2 times "dental seo" ($99.40 / $44.79)27. The AI column spans a narrower range, from $27.69 to $41.8827, than the marketing and SEO columns, from $10.08 to $99.4027. The AI phrases are therefore priced more alike across the four professions, although the medical entry is a compliance phrase. One reading is that the bidders on marketing phrases are agencies pricing in what a practice client is worth to them, which differs by profession. The bidders on AI phrases, on this reading, sell the same product to every profession. We cannot test that reading from this data.

The answering-service row shows a different split. "Legal answering service" costs $214.51 a click and "medical answering service" $35.92, so the legal phrase is 6.0 times the price ($214.51 / $35.92)27, while the medical phrase has 7.5 times the volume (5,400 against 720, from Table 5). In this pair the phrase with more searches has the lower price per click. Two phrases are too few to draw a rule from.

What cost per click does not measure

Cost per click should not be read as the value of a client. First, we expect the advertisers are mostly vendors: agencies, answering services and software companies selling to practices. A click on "dental marketing" is valued at what a vendor expects to earn from a dental practice as a customer, not at what a practice expects to earn from a patient.

Second, the price is per click, not per client. An inquiry costs the price of a click divided by the share of clicks that become inquiries, and this paper does not measure that share. For the $47.30 phrase "dental marketing"27, 1 click in 10 becoming an inquiry would mean $473 an inquiry (10 x $47.30), 1 in 20 would mean $946 (20 x $47.30) and 1 in 50 would mean $2,365 (50 x $47.30). Each is before any inquiry becomes a patient. The three rates are a range for the one link nobody has measured here.

Third, a cost per click is an average over auctions, and for a phrase searched 10 or 30 times a month it is a weak guide. We would rest nothing on a figure such as the $35.73 shown for "dental patient acquisition"27. Fourth, cost per click says nothing about retention or lifetime value, which determine what a client is worth to a professional practice. Chapter 11 returns to lifetime value as the frame, and "customer lifetime value" is itself searched about 3,600 times a month27.

The vocabulary of AI search: generative engine optimization, answer engine optimization and AI SEO

Four labels name roughly one idea, making a source visible in an AI assistant's answer. Table 7 gives their volumes beside two conventional SEO phrases. The labels overlap in meaning but are not variants, since their figures differ, and we do not add them.

PhraseSearches a monthCost per clickSource
ai seo6,600$41.4527
generative engine optimization4,400$35.0427
answer engine optimization2,400$21.3127
ai search optimization1,300$52.2327
law firm seo (for comparison)3,600$84.5327
dental seo (for comparison)1,000$44.7927

Table 7. Search phrases for AI search optimization, US searches a month and average cost per click, Google Ads keyword data, November 2026.

"Ai seo" is searched 1.5 times as often as "generative engine optimization" (6,600 / 4,400), which is itself 1.8 times "answer engine optimization" (4,400 / 2,400)27. Interest in the idea is real, and the name has not settled. The term "generative engine optimization" comes from a research paper whose authors report that their methods can raise a source's visibility in generative engine responses by up to 40%31. They also report that the effectiveness of the strategies varies across domains31. Google, for its own AI features, says that no specific optimization is required and that existing SEO fundamentals remain worthwhile32. Chapter 4 weighs the two statements. What the search data add is that many people are looking for the subject, under several names, and that a label's popularity says nothing about whether buying a service under that label works.

A practice or a writer who wants a page to be found for the idea should use the labels in circulation, each named once and in context, and explain how they relate, because the searches are spread across several names.

Marketing attribution, conversion rate optimization and speed to lead: the discipline's own terms

Table 8 lists phrases that describe the work technical marketing does (measuring, converting and responding), searched from a few dozen to several thousand times a month. They are not profession-specific, so they show general interest in the craft, not interest among practices.

PhraseSearches a monthCost per clickSource
marketing funnel9,900$19.0427
martech6,600$19.8027
customer lifetime value3,600$12.9827
marketing analytics2,900$21.5527
conversion rate optimization2,400$25.5027
marketing automation2,400$51.1027
call tracking1,600$49.1727
speed to lead880$18.5327
marketing attribution720$46.4127
multi touch attribution720$161.5327
lead response time50$68.0227
marketing for professional services210no figure27
technical marketing210no figure27

Table 8. Search phrases for marketing measurement and conversion, US searches a month and average cost per click, Google Ads keyword data, November 2026.

One idea can carry two names with very different volumes: "speed to lead" averages 880 searches a month and "lead response time" 50, which is 17.6 times (880 / 50)27. Chapter 12 examines the idea behind both names, the speed of a first response at the front desk. The sector-specific phrases are also tiny next to the general ones. "Marketing for professional services" averages 210 searches a month, against 9,900 for "marketing funnel"27. The general vocabulary exists, and the data show few searches that apply it to the professions by name. The work this paper argues for is that application, profession by profession and under each profession's rules.

Building for the compound, specific phrase

The data point to one practical rule: build for the compound phrase that names both the profession and the job. At the front desk, "ai receptionist" averages 74,000 searches a month at $43.66 a click, and "ai answering service" 4,400 at $58.4827. The specific forms are far smaller in volume and far higher in price: "ai dental receptionist" costs $152.08 a click, 3.5 times the generic phrase ($152.08 / $43.66), and "dental answering service" costs $181.68, 3.1 times "ai answering service" ($181.68 / $58.48)27.

We draw arguments from that. A page aimed at a phrase with 74,000 searches a month27 competes with every vendor in the country, and we think a small practice or its marketer is unlikely to win that contest. A page aimed at "dental answering service" meets a smaller field and a buyer whose problem is specific. And the assistants studied in Chapter 4 were asked specific questions, a kind of business in a city1, so a page that answers one specific question precisely is the kind of material such a question may draw on. The same logic is behind the rule we set for our own pages: the opening words answer the query the page targets, in a form an assistant could quote.

A phrase searched 110 times a month27 is a small prize, and whether it pays depends on how many visitors become patients and on what a patient is worth over time. We make no promise about rankings, about what an assistant will say or about revenue. Chapters 6 to 9 apply the framework to each profession, with the rules that govern what each practice may say.

Chapter 6

Dental marketing as a system

The dental market: dentists, ownership and practice economics

The American Dental Association's Health Policy Institute (HPI) reports over 200,000 dentists in the United States, a ratio of 60.0 per 100,000 people in 2025 that ranges by state from 40.6 in Alabama to 100.9 in the District of Columbia13.

Counts of dentists and dental establishments come from different instruments. Table 1 shows what each one counts and sets no figure against another, because they differ in definition, source and year.

SourceYearFigureWhat it countsCitation
ADA Health Policy Institute2025over 200,000dentists13
Bureau of Labor Statistics, Occupational Outlook Handbook2025161,400dentist jobs116
ADA, relaying Census County Business Patterns (a relayed figure)2023135,665dental practice establishments21
Bureau of Labor Statistics, Quarterly Census of Employment and Wages2024136,040private-sector establishments in offices of dentists (NAICS 6212), with 1,039,896 employees on annual average117

Table 1. Four counts of U.S. dentists and dental establishments, each from a different instrument.

The third row is the ADA's relay of a Census figure that we could not read at its source, so we attribute it to the ADA and lean on nothing in it21.

We think ownership is the structural fact that matters most for marketing, because who owns a practice decides who does its marketing. As of 2023, 72.5% of U.S. dentists were private practice owners, down from 84.7% in 200521. As of 2025, one in three dentists worked in a single location with no other dentist, 39% in a single-location practice with at least one other dentist, and 12% in group practices with 100 or more locations13. Those are shares of dentists, not of practices. The share of dentists affiliated with a dental support organization (DSO) rose from 7% in 2015 to 17% in 2025, and nearly one third of dentists less than five years out of dental school were affiliated in 202513.

We write this chapter for the dentist who owns a practice or two and has no marketing department: an owner who decides about a website, a phone line and a review reply between seeing patients. This paper does not measure how many DSO-affiliated practices run their own marketing and how many receive it from a parent organization. A larger organization would use the chapter's design as a checklist for what each location must have.

What a practice earns, and why that bears on marketing

The HPI's Survey of Dental Practice puts average gross billings per dentist in private practice in 2025 at $965,660 for general practitioners and $1,213,040 for specialists, and average net income at $215,320 and $346,52021. A general dentist's net income is 22.3% of billings ($215,320 / $965,660), which leaves 77.7% for everything else. We take the remainder, ~77.7% of billings, as dental expenses (from the HPI's billings and net income; likely high). Billings are not collections, and we read only the survey's headline numbers, not its definitions of billings and net income. The HPI also reports that expenses per dentist are rising more than revenue per dentist, which it says is driving the downward trend in general practice dentists' net incomes13. Their average annual income fell from $274,198 in 2010 to $215,320 in 2025 after adjusting for inflation13. The Bureau of Labor Statistics' median wage for dentists was $176,110 in May 2025, a different measure from the HPI's net income for dentists in private practice116.

Our argument from this is that a practice whose costs are rising faster than its revenue has less room for marketing that is only spending, so a marketing program has to show what it returned. And billings per dentist do not say what one new patient is worth to a given practice, which depends on its own mix of services and on how long patients stay. For scale, Henry Schein One† (practice software) reports in its 2026 benchmark an average of 30 new patients a month for dental practices with one to seven locations, against 81 for its top 10%, and does not describe its sample49. AHRQ's Medical Expenditure Panel Survey puts average dental spending at $887 a year for each person with any dental visit in 2021, from all payers and as households reported it118. That is spending, not a practice's revenue.

Demand is large. The Centers for Disease Control and Prevention's FastStats page, citing the early release of the 2025 National Health Interview Survey, reports that 66.7% of U.S. adults aged 18 and older had a dental exam or cleaning in the past year119. The other 33.3% did not (100% - 66.7%). The figure is an early-release survey estimate; for adults aged 18 to 64, an earlier report found the share with a dental visit in the past 12 months fell from 65.5% in 2019 to 62.7% in 2020, a different age range120. None of this says how many of those visits were to the dentist the patient saw the year before, so we draw no conclusion about how many patients change dentists.

Found: dental marketing, dental SEO and what the assistants named

In our study the three assistants were asked "Who is a good dentist in [city]? Name specific businesses." in each of 16 cities, and they named 266 dentists, 16.6 per question1. Of those 266, 3 (1.1%) were named by all three assistants and 88% were named by only one1. The study has limits: each question was asked once, names were matched across assistants by spelling, up to eight names were read per answer, and answers through an API are a baseline, not what every signed-in user sees1.

Among the health kinds asked about in all 16 cities, dentists had the lowest share named by all three assistants1. Chiropractors reached 4.4% and med spas 5.7%1. The gap between dentists and chiropractors is 3.3 percentage points (4.4% - 1.1%), which is smaller than the 6.0-point spread between the best and worst city for the ten kinds of business asked about in every city1, so we do not rank dentistry below its neighbors. The row supports a narrower claim: in this run, a patient who asked three assistants for a good dentist in the practice's city would rarely have seen the same practice in all three answers. Chapter 4 gives the full table and the reasons to read it cautiously.

The study is less specific about the sources the assistants cited. Across all 23 kinds of business, the 25 most-cited websites drew 29% of citations and were directories, review platforms, list sites, publishers and a forum1. The ten most-cited sites include zocdoc.com, cited in 44 of the 597 answers1. Those counts pool trades with professions and are not broken down by kind of business, so we cannot say which sites the assistants cited for dentists. A practice does control the facts it states in the places an assistant reads, and whether they agree.

The search side of dental SEO

Dental SEO is the work of making those facts, and the practice's own pages, readable by search engines and by the assistants built on them. Google states that a page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements, and that existing SEO fundamentals remain worthwhile32. Whether anything beyond the fundamentals pays for a small practice is open, as Chapter 4 argued.

The search data in Chapter 5 describe what owners and vendors type, not what patients type. Within that limit, dentistry is the one profession in our set where the leading marketing phrase outruns the leading AI phrase among the phrases we looked up: "dental marketing" averages 1,300 US searches a month and "dental ai" 88027. "Dental seo" averages 1,000, "ai dental receptionist" 170 and "dental answering service" 110, the last at $181.68 a click27. Dentistry therefore does not follow the pattern Chapter 5 reports for the other three professions, where the leading AI phrase outruns the leading marketing phrase.

Profiles, insurance and new-patient status as signals

A peer-reviewed US survey of how patients choose a dentist is NF; the nearest dental survey this paper cites is of adult patients in private clinics in Bucharest, Romania, where 99.1% said they learned about their dentist from family or friends92. A 2019 survey of 1,500 Americans commissioned by the agency IronMonk† asked which dentist they would pick from a Google search: 43.3% chose the one with the best online reviews, 19.4% the closest and 18.4% the top-ranked result94. The choice was hypothetical, among five options that left out insurance acceptance and referral, so the 43.3% most likely overstates the share who choose mainly on reviews94.

In a 2026 survey of 448 dental practices by MouthWatch† (teledentistry), 65.6% named word of mouth and patient referrals as their top source of new patients, down from 77.5% in 2025, which records what practices say, not where patients came from93. Patient Prism† (call tracking for dental offices) reports that 90% of the patient calls it could attribute to a source came from the Google ecosystem, and 54% of attributable inquiries from Google Business Profile alone24. None of these observes a US patient's actual choice, so what follows is argument.

A patient's first filters are practical: is the office near, does it take my plan, is it taking new patients, is there an opening this week. A profile that states those facts correctly, in every place an assistant reads, answers the filters before the patient or the assistant has to ask. A profile that states them wrongly sends the patient to a phone call that ends in a refusal.

Chosen: reviews, the reply trap and the ADA's rules on claims

Chapter 2 reviews the research on how reviews move choice. The studies of effects on sales or bookings that this paper cites concern restaurants, books and one physician practice. The one dentist study it cites analyzed the online reviews of 204,751 US dentists on a single site, not patient choice80. The physician study is the nearest to a booking effect. At one large single-specialty practice in the New York and New Jersey area, mean new-patient volume fell 2.3% to 2.6% after a 1-star review81; that is an association, not a trial. Anderson and Magruder found that an extra half-star rating caused restaurants to sell out 19 percentage points more frequently, with larger impacts when alternate information was more scarce77. A patient choosing a dentist has little other information about quality, so we expect reviews to carry weight.

Replying to a review can be a disclosure

A dentist who is a covered entity and wants to answer a review faces a rule that a restaurant does not. In an announcement dated October 2, 2019, HHS's Office for Civil Rights (OCR) said Elite Dental Associates of Dallas agreed to pay $10,000 and adopt a corrective action plan to settle potential violations of the HIPAA Privacy Rule121. According to the resolution agreement, a patient complained that the practice, replying to her Yelp review, disclosed her last name, details of her treatment plan, and insurance and cost information122. Table 2 sets out that case and three others that arose the same way, three of them resolution agreements122,123,124 and one a penalty.

PracticeDate on HHS's pageOutcomeWhat was allegedCitation
Elite Dental Associates, Dallasannounced October 2, 2019$10,000 and a corrective action plana reply to a Yelp review disclosed a patient's last name, treatment plan details, insurance and cost information; OCR found other patients' information disclosed in review replies without valid authorizations121,122
New Vision Dental, Californiapage reviewed December 14, 2022$23,000 and a corrective action planreplies to Yelp posts habitually disclosed patient information, sometimes full names where patients used only monikers, with detailed visit and insurance information123,125
Manasa Health Center, New Jersey (a psychiatric provider, not a dental practice)page reviewed June 5, 2023$30,000 and a corrective action planinformation of four patients disclosed in replies to negative Google reviews124,126
Dr. U. Phillip Igbinadolor, D.M.D. & Associates, North Carolinapage reviewed March 28, 2022$50,000 civil money penalty imposed by OCRa patient's information disclosed on a webpage in response to a negative online review; the practice did not respond to OCR's data request or subpoena and waived its right to a hearing127

Table 2. Four OCR actions over replies to online reviews. Three were resolution agreements; the fourth was a penalty imposed after the practice did not respond to OCR.

Breadth and time matter for system design. In the Elite case the complaint concerned one reply posted on June 4, 2016, and OCR's review of the practice's Yelp page found other patients' information disclosed in other replies122. About three years and four months passed between that reply and the October 2, 2019 announcement (calc. from the dates in the agreements and on HHS's pages)121,122. About five years passed between the New Vision complaint (November 29, 2017) and the December 2022 page123,125. One complaint about one reply can open an inquiry into the practice's other replies, and a published resolution can take years.

OCR's list of covered conduct for Elite also included a failure to implement policies and procedures on protected health information, including releasing it on social media or public platforms122. The agreement cites 45 C.F.R. § 164.530(i). We think the practical content is a written policy for who may answer a review, in what words and with what approval, in place of an ad hoc reply by whoever holds the login. The cases do not say what a safe reply contains, and we do not supply one. Whether a reply that mentions no care at all can still confirm that the reviewer is a patient is a question for the practice's counsel, and the OCR documents cited here do not address it.

What the penalty tiers say, with their dates

HIPAA civil money penalties are set by tiers of culpability128. Table 3 gives the ranges in HHS's inflation-adjusted table published January 28, 2026 (the 2025 adjusted amounts), together with the annual limits that HHS said in a notice published April 30, 2019 that it would apply, before inflation adjustment.

Culpability tier2025 adjusted range per violationAnnual limit under the 2019 notice, before inflation adjustment
Did not know and would not have known by reasonable diligence$145 to $73,011129$25,000128
Reasonable cause, not willful neglect$1,461 to $73,011129$100,000128
Willful neglect, corrected within 30 days$14,602 to $73,011129$250,000128
Willful neglect, not corrected$73,011 to $2,190,294129$1,500,000128

Table 3. HIPAA civil money penalty tiers: the 2025 adjusted ranges published January 28, 2026, and the 2019 enforcement-discretion annual limits.

The $2,190,294 in the last row is also the calendar-year cap printed in HHS's table129. Before 2019 the Enforcement Rule applied an annual limit of $1.5 million to each of the four tiers128. The 2019 notice states lower annual limits for the three lower tiers128. We did not check whether a later notice or rulemaking has replaced the 2019 notice, so the current figures need confirming. Each payment in Table 2 is below $73,011, the upper end of the per-violation range in each of the first three tiers, and we read the payments as a record of what OCR resolved, not a forecast of what it will.

What the ADA's code says about advertising and specialist claims

The ADA Principles of Ethics say that although any dentist may advertise, no dentist shall advertise or solicit patients in any form of communication in a manner that is false or misleading in any material respect7. A dentist who announces specialization should use "specialist in" and must devote a sufficient portion of the practice to the announced specialty to maintain expertise in it7. An advisory opinion adds that statements to be avoided include those likely to create an unjustified expectation about results7. They also include objective representations that the advertised services are superior in quality to other dentists' if the representation is not subject to reasonable substantiation7.

The ADA's Principles of Ethics are an ethics code and not itself law7. The binding rules come from the states. Table 4 sets out what the texts we read in November 2026 say in five of them, with one to three provisions from each state. It does not survey the whole of any state's law, and the current text for the state where a practice operates needs confirming.

State and provisionWhat the text saysCitation
California, Business and Professions Code 651(b)(3)(C)"Before" and "after" views must be comparable in presentation and must carry a statement that the same results may not occur for all patients130
California, Business and Professions Code 651(b)(6)A claim of professional superiority or of performing services in a superior manner is false or misleading unless it is relevant to the service and can be substantiated with objective scientific evidence130
California, Business and Professions Code 1680Advertising professional superiority, or a guarantee of any dental service or of a painless operation, is unprofessional conduct by a dentist131
Texas, 22 TAC 108.54(a)A dentist may use "specialty" or "specialist" only in recognized specialty areas that a board certifies and that the Commission on Dental Accreditation of the ADA accredits132
Texas, 22 TAC 108.57(c)(8)A testimonial from a person who is not a patient of record, that is false or not readily verifiable, or that omits disclosures about the identity and credentials of the person making it is an example of false, misleading or deceptive advertising133
Texas, 22 TAC 108.57(d)Photographs of actual patients of record need the patient's written consent, and the advertising must say "Actual results may vary"133
Florida, Statutes 466.019(3)An advertisement may not contain laudatory statements about the dentist or group of dentists, or statements relating to the quality of dental services compared with other available dental services134
Florida, Statutes 466.019(1)"Advertisement" includes websites and social media134
New York, 8 NYCRR 29.5(b)Claiming professional superiority or special professional abilities is unprofessional conduct, except that a dentist who completed Regents-approved specialty training may advertise the specialty; "practice limited to" is deemed a claim of special abilities135
New York, 8 NYCRR 29.1(b)(12)(iv)Testimonials are permitted only if conditions are met, including the patient's express written authorization of the portrayal136
Washington, WAC 246-817-420(3)A general dentist may render specialty services but may not advertise or represent themselves as a specialist in the listed areas137

Table 4. Provisions on dental advertising in five states, from the texts read in November 2026.

Two applications are ours. A page that calls a practice the best in its city makes a superiority claim of the kind the ADA opinion describes. Under the texts in Table 4, California treats an unsubstantiated superiority claim as false or misleading, Florida bars laudatory statements and comparisons of quality, and New York and Washington restrict claims of special ability and of specialist status. Whether a given page falls within any of them is for the practice's counsel. And an assistant naming a practice is not a measure of quality (Chapter 4), so a practice that advertises that an assistant recommended it should expect the same question.

The FTC's rule on reviews

Reviews are also covered by consumer-protection law. The Federal Trade Commission's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials (16 CFR part 465) took effect on October 21, 2024138. It makes it an unfair or deceptive act for a business to write, create or sell a review that materially misrepresents that the reviewer exists, used the service or had the experience described83. It is equally unfair or deceptive to provide compensation or other incentives in exchange for reviews expressing a particular sentiment, positive or negative84. It also reaches the use of unfounded legal threats, intimidation or knowingly false public accusations to prevent a review or have it removed85.

In warning letters of December 22, 2025, the FTC said violations can bring civil penalties of up to $53,088 per violation139. The provisions described above are addressed to a business or to anyone, and not to health practices in particular83,85. For a dental practice, any program that asks patients for reviews should note that a reward depending on the sentiment of the review is within what the rule names84.

Reached: scheduling, reminders and the dental answering service

First contact for a dental practice arrives as a call, a form or a booking request, and each can fail. Several sources measure how often dental calls reach a person. They count different things, so we give each with its owner and definition and do not average them.

Patient Prism† analyzed 11,552,668 patient calls in 2025 to 8,280 dental practice and DSO locations. In its data, of every 100 calls, 69 connected with the front desk and 31 hung up before reaching an agent, counting existing and new patients together24. Patient Prism† also reports that 890,621 patients who called and walked away without booking never got a callback, and that the average share of walk-aways who received a follow-up was 38%, against more than 80% for its best operators24. PatientNews† (dental marketing) says its average client practice answers 75% of calls, against 94% for its top practices140. In a 2015 trade column, Jay Geier, whose company sells front-desk training, reported that 35% of test calls to more than 12,600 dental practices in the United States, Canada, Australia and the United Kingdom went straight to voicemail during regular business hours25.

One survey is restricted to a single corner of dentistry: a secret-shopper phone survey of Arizona Medicaid pediatric dental practices reached 159 of 185 cataloged practices (86%) on behalf of a Medicaid-enrolled child and 149 (81%) on behalf of a commercially insured child141. Together these sources show that a share of dental calls does not reach a person, but none gives a rate for dental offices as a whole. Table 8 turns Patient Prism†'s hang-up rate into calls and bookings a month.

Reminders and no-shows

The evidence on reminders is general to health care. A 2013 Cochrane review found moderate-quality evidence from seven trials (5,841 participants) that mobile text message reminders improved attendance at healthcare appointments compared with no reminders, with a risk ratio of 1.14 (95% confidence interval 1.03 to 1.26)142. Across the pooled trials, attendance was 67.8% with no reminders, 78.6% with text reminders and 80.3% with phone call reminders142. Two studies reported that the cost per attendance of text reminders was 55% and 65% lower than that of phone call reminders142.

Those figures come with the authors' own hedge: the evidence remains insufficient to conclusively inform policy decisions142. A 2018 systematic review of 105 studies of healthcare appointment scheduling found an average no-show rate of the order of 23%, highest in Africa (43.0%) and lowest in Oceania (13.2%)142. We infer from both that missed appointments are a large, measurable loss in health care generally, and that a dental practice should measure its own.

A peer-reviewed or ADA overall no-show rate for US dental practices is NF. The nearest evidence is a 2014 peer-reviewed paper on a university dental clinic, which quotes published failure-to-show rates for US dental appointments that run from 7.3% at a university pediatric clinic to 24% at 57 safety-net clinics in Illinois to 47.0% at a community pediatric clinic143. None is a private general practice143. Henry Schein One† puts the average dental no-show rate at 4% for practices with one to seven locations, against under 1% for its top 10%49. It gives 3% for dental groups and defines neither rate144.

These figures come from different clinics under different definitions, so we do not average them. For a private general dental practice, the no-show rate is ~4% to 7.3% of booked visits (from Henry Schein One†'s average and the lowest published rate; likely low). A software vendor's customers probably schedule more tightly than the average practice, and a published rate describes a clinic that someone chose to study.

The same Henry Schein One† benchmark gives average patient retention of 70%, against 94% for the top 10%, without defining retention, so it is not an annual attrition rate49. In 410,000 appointment records over three years at University of Michigan dental school clinics, 94.9% of patients who were emailed a reminder and 92.2% of those who were texted attended, an association in a teaching clinic and not a trial142.

The dental answering service and dental AI

The ADA Health Policy Institute's quarterly panel of dentists in mid-2026 found that 43.3% of responding dentists used AI for at least one type of task and another 26.4% planned to40. Of all responding dentists, 13.6% used AI for insurance verification, 13.2% to explain clinical findings to patients and 10.1% for reception or front desk check-in40. These are shares of responding dentists in a panel. A randomized or field study of an AI agent answering inbound calls or scheduling in dental settings is NF. The nearest peer-reviewed figure is for outbound calls at one hospital: in the first phase of a prospective study at Mount Sinai with no control group, an AI voice assistant making pre-procedure calls completed 86.4% of its calls47. That is a guide, not a measure of a dental front desk.

The ADA's own papers on AI address other questions. The one ADA technical white paper this paper cites, approved in December 2022, has a stated scope that covers clinical disciplines145. An AI policy adopted by the ADA House of Delegates is NF. The nearest evidence is the Council on Dental Practice's 2026 annual report, which says it will submit a proposed "Statement on Use of Artificial Intelligence in Dentistry" to the 2026 House of Delegates146. It also lists the 2021 policy statement on augmented intelligence among the policies it determined should be rescinded146. Both are council proposals and not adopted policy. In February 2026 the ADA told HHS that inconsistent definitions of non-medical devices across agencies and industry create uncertainty around liability, patient consent, privacy and security for AI tools in dental practices147. That is advocacy to a federal agency, not guidance to dentists147.

A dental answering service staffed by people and an AI receptionist raise the same legal question, because a caller gives a name, a symptom or an insurance plan. If the practice is a covered entity, HHS describes a business associate as a person or organization that performs functions on its behalf involving creating, receiving, maintaining or transmitting protected health information55. A subcontractor that does so for a business associate is itself one55. The Privacy Rule permits disclosure to a business associate on satisfactory assurances, documented in a written contract that meets the rule's content requirements148. An AI receptionist that passes a call's content to a separate language-model vendor raises the question whether that vendor is a subcontractor in the sense HHS describes, and the written agreements answer it.

Our own check of 28 software tools that small practices commonly use found that 9 will sign a business associate agreement without requiring a special tier, 15 sign one only on certain plans and 4 offer none, according to the vendors' own pages54. Chapter 10 builds on that check.

Chapter 7 covers the disclosure statutes for AI in health care. California's section 1339.75 names health facilities, clinics, physician's offices and group practice offices and not dental offices149. It excludes appointment scheduling and billing from its disclaimer duty149. Texas's section 552.051(f) requires the provider of a health care service or treatment in which an AI system is used to give a disclosure no later than the date the service or treatment is first provided150. It took effect on January 1, 2026150. Whether a given dental office, or a front-desk tool, is inside that text is for counsel. Our rule for our own service does not depend on the answer: callers and site visitors are always told that they are talking to an AI assistant.

Within the rules: HIPAA compliant marketing and tracking tools on a dental site

People search for "hipaa compliant marketing" about 40 times a month and "hipaa marketing rules" about 7027. We use the phrase in a heading because people type it, and we do not apply it to any vendor or tool. HHS states that no standard or implementation specification in the Security Rule requires a covered entity to certify its compliance151. It does not endorse or recognize private organizations' certifications regarding the Security Rule, which do not absolve covered entities of their legal obligations151. That FAQ speaks of the Security Rule, not the whole of HIPAA, and we keep that scope. It is one reason this paper says "HIPAA-ready", never "HIPAA compliant", of any tool, including our own.

The marketing rule in plain terms

The Privacy Rule defines marketing, in general, as making a communication about a product or service that encourages recipients to purchase or use it, and the definition carries exceptions in its second paragraph10. A covered entity must obtain an authorization for any use or disclosure of protected health information for marketing, except a face-to-face communication or a promotional gift of nominal value152. HHS's plain-words gloss is that a communication that is marketing can generally occur only if the covered entity first obtains the individual's authorization11. It adds that covered entities may not sell lists of patients without authorization from each person on the list11. We rest the rule on the regulation; HHS's page is a gloss.

On its text, the authorization duty attaches to a use or disclosure of protected health information. We think an advertisement that uses none is outside this provision, though the advertising rules above still apply. A practice that wants to feature a patient's name, photograph or story is asking whether that is a use of protected health information for marketing, and the text requires an authorization for any such use. Which of a practice's recall and reminder messages fall inside the exceptions is open. Chapter 7 covers authorization, financial remuneration and the business associate chain.

Tracking tools on the site

A dental website can send information to third parties through the tools that measure it. HHS's bulletin on online tracking technologies says that tracking technologies on many unauthenticated webpages do not have access to individuals' protected health information, in which case the HIPAA Rules do not regulate their use, though in some cases they may have access to it12. It says banners that ask users to accept or reject cookies do not constitute a valid HIPAA authorization12. It also says it is insufficient for a tracking vendor to agree to remove protected health information before saving it12. One element of the bulletin was vacated by a federal court in 202412. Chapter 7 gives the full treatment of the bulletin, the court decision and the FTC's cases, which concern health sites generally.

The design question for a dental site is narrower: which pages and forms can carry protected health information. A new-patient form, an online booking flow, a call-tracking number that records, and a chat box are places where it can. In our check of 28 tools, Google Analytics, Meta Pixel, Calendly and Zapier are the four that offer no business associate agreement, according to the vendors' own pages54. We state that as a fact about those vendors' pages, not as a finding about any practice's use of them.

Measured: what a dental practice can count each month

A dental practice can count most of what matters from systems it already has, as Table 5 lists. None needs a patient's health information to be sent to a third party: each can be kept as a number or a category.

What to countWhere it comes fromWhat it can showWhat it cannot show
Calls received and answered, by hour and locationphone system or answering-service logwhether the front desk is reachable when patients callwhy a caller rang, or whether a missed call was lost
Booking requests by channel (phone, form, text, online)scheduling systemwhere first contact happenswhether a request became a visit
New-patient appointments scheduled and kept, by source asked at bookingpractice management systemwhich channels end in a visitwhat caused the visit (Chapter 11)
Cancellations and no-showsscheduling systemaccess lost, and the effect of reminderswhich reminder caused which attendance
Reviews: new, average rating, replies sent under policyreview profilesreputation signal and policy compliancewhether reviews changed any choice
Profile audit: name, address, phone, hours, plans accepted, new-patient statusmanual check of the main listingswhether the facts assistants read agreewhich listing an assistant used
Assistant check: the same questions put to three assistants on a fixed date each monththe method of Chapter 4whether the practice is named, and what is said about itwhy; a single draw proves little1
Recall: patients due for a visit, and the share scheduledpractice management systemretention work in progresswhether a patient went to another dentist

Table 5. A monthly count for a dental practice, with its limits.

On recall, Gupta, Lehmann and Stuart report that improving customer retention by 1% is likely to improve customer and firm value by 3% to 7%109. That is the authors' general result, not a dental one, and Chapter 11 takes it up. For a practice, recall is a form of marketing with a measurable output.

A practice's monthly counts can be small enough that a swing of a few appointments is chance, not a change in the marketing. The same difficulty appears among large advertisers: in field experiments on advertising, the median standard error on return on investment for the retail experiments was 26.1%, implying a confidence interval over 100 percentage points wide103. We therefore ask of a monthly count only that it be recorded the same way each month, and we read trends over quarters. Chapter 10 covers the privacy limits on measurement and Chapter 11 the small-numbers problem.

What a dental patient is worth, and what a dental marketing system has to win back

A practice buys a system: a plan that produces demand, an agent that answers it and, for the tools that touch patient information, a HIPAA-ready program as the compliance layer. Each fee is set against what it buys: the plan against new patients, the agent against answered calls and front-desk hours, and the program against everything that runs on it.

Plans start at $999 a month per location56. Marketing alone carries no patient information, so a dental practice can buy a plan with no HIPAA program; the plan alone is the least a practice can buy57. The voice agent is $299 a month with 250 call minutes included, about 62 four-minute calls (250 / 4), and $0.39 a minute after that56. The worked cases here stay inside the included minutes, so the agent is $299 a month. The Essentials program is $699 a month after $1,500 setup57, which is $824 a month in the first year ($699 + $125, the setup over 12 months). It is the compliance layer that every tool handling patient information runs inside, and it does not include the agent57. The agent on patient calls, the site assistant, appointment forms, texts to patients and call tracking each need it57, so its fee is shared by everything that runs on it and no single gain is asked to carry it. Table 6 gives the stacks (calc. from those prices).

What is boughtMonthly costSet against
A plan alone, the marketing$999new patients
The agent, added to a plan the practice has$299answered calls and front-desk hours
HIPAA-ready Essentials, the compliance layer, first year$824everything that runs on it
Essentials after the first year$699everything that runs on it
The agent and Essentials, first year: the narrow case, a practice whose only patient-facing tool is the agent$1,123the agent's gains only
Plan, agent and Essentials, first year$2,122all of it
The same after the first year$1,997all of it

Table 6. Monthly cost of each purchase for one dental location (calc. from the published prices56,57), with the $1,500 setup spread over 12 months: the agent is $299 with 250 minutes included, Essentials is $699 + $125 in the first year, and the narrow case is the two together ($299 + $824).

Four front-desk add-ons run inside the program: reminders and rebooking at $149 a month, recall and reactivation at $199, waitlist backfill at $99, and intake with insurance capture at $149 plus $250 setup, or $449 a month for all four, per location plus usage57. They run on the program, so they share its fee, and their own prices are not in Table 6.

What a patient and a visit are worth

Four cited figures price a dental patient and a visit. AHRQ's survey puts average dental spending at $887 a year for each person with any dental visit in 2021118. The HPI puts a general dentist's average gross billings in private practice at $965,660 in 2025 and average net income at $215,32021. Statista's series on yearly patient visits per dentist gives roughly 3,429.5 for 2025, and its page does not name the survey behind it, which may count hygiene visits153. Billings come to $80,472 a month for one dentist ($965,660 / 12), and net income is 22.3% of billings (calc., above). Billings per visit are ~$282 ($965,660 / 3,429.5; joins two sources, likely low), out of about 286 visits a month (3,429.5 / 12).

The plan against new patients

The whole system costs 2.6% of one dentist's monthly billings in its first year ($2,122 / $80,472) and 1.3% of two dentists' (calc.). The plan alone, at $999, is covered by 1.1 new patients a month on first-year spending ($999 / $887). Against Henry Schein One†'s average of 30 new patients a month49, that is 3.8% more patients ($999 / $887 / 30). This paper counts a patient's first year only, which errs low; the one retention figure on file, Henry Schein One†'s 70%, cannot be turned into years49. The whole system needs more new patients, and how many depends on how much of a patient's spending the practice keeps. Table 7 shows four shares. The last is the average dentist's net income as a share of billings, which we think is a cautious share, because rent and salaries are already paid when one more patient is seen.

Share of a patient's spending the practice keepsNew patients a month needed, first year only
100%2.4
50%4.8
25%9.6
22.3%, the average general dentist's net income share10.7

Table 7. New patients a month that cover the whole system at $2,122 (calc. from Table 6 and the $887 of first-year spending).

In visits, the whole system is 7.5 more kept visits a month ($2,122 / $282) at ~$282 of billings a visit. At the ~4% to 7.3% no-show rate, a dentist with about 286 visits a month misses 11 to 21 of them (4% and 7.3% of 286), which is why recovering missed visits alone would have to win back between a third and two-thirds of them, or a smaller share if the estimate errs low. So the dental case rests on new patients and on calls rather than on no-shows alone.

A full schedule changes the form of the gain

The ADA Health Policy Institute's poll of 704 private-practice dentists found nearly two in five not busy enough in the third quarter of 2026 and able to have treated more patients102. In the fourth quarter of 2025, one-third reported being not busy enough, 12% said they were too busy to treat all patients requesting care, and 18% treated all patients but were overworked101. Owner dentists accepting new patients reported an average wait of 12 days for a new patient's first appointment in the third quarter of 2026102. For the dentists with open chairs, a new patient fills one. For the busy minority, demand changes form. We think there are four forms: canceled slots filled from a waiting list, better-paying work chosen from a longer list of patients, a higher minimum fee for the work the practice does take, and another dentist or hygienist added. None of the four is measured in this paper.

The calls against the agent

Patient Prism†'s 11,552,668 calls at 8,280 locations in 2025 are 116 patient calls a month per location (11,552,668 / 8,280 / 12)24. In its data, 31 of every 100 calls hung up before reaching an agent, which is 36 a month (116 x 31 / 100)24. Of the 69 that connected, its own classification gives 34 as real booking opportunities and 35 as not bookable, and 21 of the 34 ended in a booked appointment24. If the 36 hang-ups resemble the callers who get through and every one were answered, the result is 11.0 bookings a month (36 x 21 / 69). The callers are new and existing patients together, so a booking is valued at two ends: one visit at ~$282 of billings, and a new patient's first year at $887118, which comes to $3,089 to $9,730 a month on the hang-ups (calc.).

The chain assumes that hang-ups resemble callers who get through, which the source does not show. It counts none of the callers who ring back, which lowers every figure, most of all at the low end where the callers are existing patients.

One link is open: the share of those hang-ups that an agent would turn into conversations. A field study of that share is NF; the nearest figure is the 86.4% of outbound pre-procedure calls completed at one hospital47. So the chain is solved for the share instead. The agent, at $299, must turn 3% of the hang-ups into conversations if each booking is a new patient's first year (0.3 bookings, $299 / $887) and 10% if each is one visit (1.1 bookings, $299 / $282). The chain is set against the agent's own fee, because answered calls are the agent's gain. The plan's product is new demand, which Table 7 counts, and charging the plan against answered calls as well would charge a fee with nothing credited for it.

The program is not charged to answered calls either: it is a layer and not a tool, and the site assistant, texts to patients and the front-desk add-ons run inside the same one57, so its fee is shared by everything on it. The narrow case is a practice that buys the program for the agent and runs nothing else on it: against that $1,123 the agent must turn 12% of the hang-ups into conversations if each booking is a new patient's first year (1.3 bookings, $1,123 / $887) and 36% if each is one visit (4.0 bookings, $1,123 / $282). Table 8 shows the open link at four values. It is a range for the one link nobody has measured, not a forecast, and no row in it is a finding.

Share of hang-ups turned into conversationsCalls a monthBookingsGain, one visit eachGain, a new patient eachLess the agent ($299)Narrow case: less the agent and a program bought for it alone ($1,123)
10%3.61.1$309$973$10 to $674-$814 to -$150
25%9.02.7$772$2,433$473 to $2,134-$351 to $1,310
50%18.05.5$1,544$4,865$1,245 to $4,566$421 to $3,742
75%27.08.2$2,317$7,298$2,018 to $6,999$1,194 to $6,175

Table 8. Value of recovered hang-ups, and that value net of the agent's fee, at four values of the open link (calc. from Patient Prism†'s rates). Each of the last two columns is a range from valuing a booking at one visit to valuing it at a new patient's first year; the first is net of the agent's $299 (at 10%, $309 - $299 = $10 and $973 - $299 = $674), and the last is the narrow case.

The system loses money in one corner: a practice that is not growing and whose intake and scheduling are already sufficiently optimized, with no new patient wanted, few callers lost and few visits missed. The cited base rates put a typical practice far from it. In Patient Prism†'s data 31 of every 100 calls hung up before reaching an agent, nearly two in five dentists say they could treat more patients, and Henry Schein One†'s average practice with one to seven locations sees 30 new patients a month.

The agent alone can still lose at a low share: in Table 8, at 10% of hang-ups it is ahead of its $299 at both ends, by $10 if each booking is one visit and by $674 if each is a new patient's first year, and a share below 10% leaves it short at the one-visit end; at 25% it covers its fee at both ends. In the narrow case, the last column of Table 8, it does not cover its fee at 10% at either end, and at 25% only if each booking is a new patient. The cases do not price errors. A wrong answer or a lost call that the agent causes would have to consume the gap in Table 8: at 50% of hang-ups, $1,245 to $4,566 a month net of the agent's fee, or $421 to $3,742 in the narrow case. Chapter 12 treats quality and error at the front desk.

A worked design for a two-location practice

This section is an illustration. Suppose a general dental practice with two offices, an owner-dentist who works at both, a front desk at each office and an office manager who also handles billing. Nothing in the example is a finding, and a figure computed in it shows its working.

Framing

The first step is a system map with one named owner per stage. The office manager owns the accuracy of the profiles and the review-reply policy. The owner-dentist owns what the practice claims, including any statement of specialty, in line with the ADA provisions above. Each front desk owns call handling at its office. One person owns the monthly report. The map also records which tools hold patient information, because that list decides which vendors need an agreement.

Plumbing

Each office gets one profile record with the same name, address, phone, hours, accepted plans and new-patient status wherever it appears. Phones route to the front desk during hours and to an answering service or an AI receptionist outside them, each under a signed business associate agreement and each with a written rule for when a staff member takes over. An AI receptionist says at the start that it is an AI assistant. Scheduling sends reminders by text or by a staff member's call; an AI-voiced reminder call raises a consent question that Chapter 7 describes.

The booking and new-patient forms collect only what the visit needs. The site measures visits, calls and form submissions as counts, and any tool that would receive form content or a booking page's contents is checked against the business associate and tracking rules before it is installed. A vendor register lists every party that touches patient information, with its agreement and its subcontractors.

Finish

Each office has its own page with its services, the plans it accepts, its team and their credentials, written to answer the questions patients ask in the form Chapter 4 used: a kind of practice in a city, with a need attached. Claims of specialty use the ADA's wording, and in this design claims of superiority are held until someone can substantiate them. The review program asks every patient the same way, offers nothing that depends on the sentiment of the review, and sends every reply through the written policy.

Operating

Once a month the report shows the counts in Table 5 and the assistant check, with the date. Changes are made one at a time and logged, so a later change in a count can be read against what was done. The report names what could not be known, and a practice that cannot yet measure a row says so.

Our own practice in health care is to work only under a signed business associate agreement. The architecture is by Keegan Justis, and the build is by Mike Golliet and Kyle Grenier.

The design needs no program for the marketing itself, because a plan carries no patient information; its least cost is a plan at each office56,57. The program covers the tools that touch patient information, here the booking and new-patient forms, call tracking and recordings, and the reminders.

Essentials is $1,500 setup plus $699 a month and Integrated is $3,500 setup plus $1,199 a month, per location, after a compliance consultation57. At those per-location rates, Essentials at two offices would be $3,000 in setup (2 x $1,500) and $1,398 a month (2 x $699). The program is a compliance layer added to a plan and does not include the AI voice agent, a separate add-on at $299 a month with 250 call minutes included57. HIPAA-ready is not a claim of compliance, which depends on the agreement and on how a tool is set up.

What this chapter has not shown

The evidence for dentistry is thinner than the dentist row suggests. That row is 16 questions asked once on one day1, and it says nothing about whether any named practice is good.

A study of dental patients' choices would ask them what they read and weigh. A call audit of dental offices, repeated over weeks, would measure access. A panel of assistant answers, repeated, would separate real disagreement from run-to-run variation. Chapter 7 turns to physician practices and clinics, where the privacy law shapes the system most and where our study has no row.

Chapter 7

Medical practice marketing under HIPAA

Medical practice marketing: ownership, size and the cost of running a practice

In the American Medical Association's 2024 Physician Practice Benchmark Survey, 42.2% of physicians were in private practice, meaning a practice wholly owned by physicians, which is 18 percentage points lower than in 201214. This chapter applies the paper's framework to physician practices and clinics, where we think privacy law does more than any other body of rules to shape the system.

Table 1 gives the AMA's series. The survey vendor changed in 2024, from M3 Global Research to WebMD/Medscape, and the sample grew from 3,500 to 5,000 physicians, so comparisons across that year carry a caveat14.

Survey yearPhysicians in private practiceCitation
201260.1%14
201456.8%14
201655.8%14
201854.0%14
202049.1%14
202246.7%14
202442.2%14

Table 1. Share of physicians in private practice, AMA Physician Practice Benchmark Survey, 2012 to 2024.

Other measures in the same survey point the same way, though each differs from the one in Table 1. Only 35.4% of physicians had an ownership stake in their practice in 2024, against 53.2% in 201214. In 2024, 6.5% of physicians characterized their practice as private equity-owned, higher than the around 4.5% of 2020 and 202214. And 47.4% worked in practices with 10 or fewer physicians, compared with 61.4% in 201214. These are shares of physicians in a survey, not of practices.

The federal counts are different instruments again, and we set none against another. The Bureau of Labor Statistics counts 862,800 physician and surgeon jobs in 2025 and reports a median annual wage of $275,930 for physicians and surgeons in May 2025154. A median wage does not measure practice owners' income. The Bureau's Quarterly Census of Employment and Wages shows 239,840 private-sector establishments in offices of physicians in 2024, with 2,954,911 employees on annual average155.

What it costs to run a practice

The Medical Group Management Association (MGMA) reports in a June 23, 2026 poll, with 251 applicable responses, that 84% of medical groups had year-to-date operating costs higher than at the same point in 2025, with respondents reporting an average increase of about 11%156. We read the poll as describing its respondents, not all practices. MGMA's DataDive analysis reports that from 2011 to 2024 median total operating cost per full-time-equivalent physician rose 71.6% in physician-owned multispecialty groups and 83.3% in hospital- or IDS-owned peers, while the consumer price index rose 39.5%157. The analysis concerns multispecialty groups, so it does not describe a single-specialty or solo practice.

Two arguments follow. For an independent practice, marketing is a purchase made from a base whose operating costs have risen, so it has to show what it returned. For a physician employed by a larger organization, marketing may be another person's decision, and this paper does not measure how many employed physicians work where the practice itself controls its profiles, its site or its phone.

Found: what our study can and cannot say about physicians

Our study has no row for physicians: it asked about 23 kinds of business, none of them a primary care or specialty physician practice1. The nearest rows are other health care kinds, which we report only to mark where the study stops. Of the names the assistants gave for chiropractors, 10 of 228 (4.4%) were named by all three; for med spas the figure was 13 of 228 (5.7%); and for physical therapists, asked about in three cities only and so a small sample, 1 of 51 (2%)1. Those are not physician figures, and a medical practice should not read its own prospects from them.

The study's limits apply to anything we say about other rows: each question was asked once, names were matched by spelling, up to eight names were read per answer, and API answers are a baseline and not what every signed-in user sees1. A study of physicians would need several wordings, because patients ask for a specialty, a plan, an urgent slot or a language, and it would have to check named practices against an independent register. We have not done that study.

The ten most-cited sites across all 23 kinds include zocdoc.com, in 44 of the 597 answers1. The pool mixes trades and professions, so we cannot say what share of those citations came from health questions. A related result comes from outside our study: in a 2023 study of the verifiability of generative search engines, on average only 51.5% of generated sentences were fully supported by citations, and only 74.5% of citations supported their associated sentence158. The study concerns the systems of that date; we think it is a reason for a practice to expect that an assistant may state its plans, hours or wait times wrongly, and to keep those facts accurate at the sources the assistant reads.

The search phrases of medical practice marketing

The phrases in Chapter 5 are the ones an owner, a manager or a vendor would type. Within that limit, "healthcare marketing agency" averages 5,400 US searches a month, "medical answering service" 5,400 at $35.92 a click, "medical marketing agency" 4,400 and "hipaa compliant ai" 6,600, while "healthcare marketing" averages 1,600, "medical marketing" 590, "medical seo" 390, "medical practice marketing" 320 at $87.76 a click and "ai medical receptionist" 14027. Volume measures interest, not purchase, and the data cannot separate owners from vendors, students and others. We think the pattern is consistent with practice owners looking first for someone to do the work or to answer the phone.

Chosen and reached: access, and the medical answering service

We think that for a physician practice the product a patient chooses is often the next available appointment. AMN Healthcare's 2025 survey, in which research specialists called 1,391 physician offices in 15 large metro areas to schedule a new-patient appointment between January 15 and February 24, 2025, found an average wait of 31 days across six specialties, up 19% since 2022 and 48% since 200422. Table 2 gives the other figures from the survey. The comparisons with 2022 and 2004 are the survey's own, and we do not say whether the specialties compared are the same in every year22.

MeasureResultCitation
Average new-patient wait, six specialties, 15 large metros31 days22
Family medicine, average wait23.5 days, ranging from 1 day to 207 days22
Dermatology, average wait36.5 days22
Cardiology, average wait32.7 days22
Offices accepting Medicaid53%22
Offices accepting Medicare82%22

Table 2. Findings of AMN Healthcare's 2025 physician appointment wait-time survey, 15 large metro areas.

Wait times say nothing by themselves about how often a physician's office answers the phone. For calls, the figures that exist come from VA call centers and from vendors, and none measures a private physician office; the no-show figures below have the same limit. The call figures count different things, so we give each with its owner and do not average them or set them on one scale. In October 2014 the mean telephone abandonment rate across Veterans Health Administration medical centers, calls ended by the caller before a staff member answered, was 12.0%, with a mean average speed of answer of 87.1 seconds159; by September 2016 the rate had fallen to 8.3% and the speed of answer to 69 seconds159. Invoca† (call tracking) reports that a person answered 54% of inbound calls to health care organizations in its data, across behavioral health and rehab, dental, eye care, hearing, hospitals and physical therapy26. CallMyDoc† (answering service) reports that 33.1% of 4,687,808 inbound calls to 297 practices on its own platform were abandoned before the caller selected a call type160.

Insurance acceptance is another filter on access. In a 2011 audit of access to specialty care for children with public insurance, 66% of Medicaid-CHIP callers (179 of 273) were denied an appointment, against 11% of privately insured callers (29 of 273)161. Among 89 clinics that accepted both insurance types, the average wait for Medicaid-CHIP enrollees was 22 days longer than for privately insured children161.

Long waits do not make demand worthless to a practice. In the AMN survey a new patient waits 31 days on average, and in the studies this chapter cites, 18.8% to 23% of booked visits are missed. The 18.8% is the mean no-show rate across 10 main clinics of one VA medical center over fiscal years 1997 to 2008162, and a 2018 review of 105 studies of healthcare appointment scheduling found an average of the order of 23%142. The AMN offices are not those clinics, so we do not claim that one practice has both a long wait and a high no-show rate. A practice short of appointments can still lose patients and visits it could have kept, because a slot left empty by a patient who does not come, and a call that no one answers, are losses in a full practice as in an empty one.

What extra demand is worth to a busy practice depends on its form; a full calendar changes the form without erasing the worth. We think there are four forms, none measured in this paper. Slots left by missed or canceled visits can be filled from a waiting list. A practice with more requests than slots can choose better-paying work from them, and payment per visit differs by payer: in 2016 the median for most specialties was highest for privately insured people and lowest for people on Medicaid61, and 53% of offices in the AMN survey accepted Medicaid and 82% Medicare22. For work that no payer prices, a practice can set a higher minimum fee. Demand that outruns every slot is the case for adding a clinician.

A new practice, a practice adding a clinician or a specialty with short waits faces the plainer problem of filling its calendar. In either case the useful work includes making capacity visible and reachable: the open slot, the plans accepted and the phone number that gets an answer, stated the same way wherever a patient or an assistant looks.

A medical answering service raises the questions of Chapter 6 with higher stakes, because a caller may describe symptoms. If the clinic is a covered entity, whoever takes the call, whether a person at a service or the vendor behind an AI agent, falls within HHS's description of a business associate when it performs functions on the clinic's behalf that involve creating, receiving, maintaining or transmitting protected health information55. We think taking a message is an administrative act, while judging how urgent a symptom is, or advising what to do, is clinical work that belongs to the clinic's clinicians, not to a front-desk tool. Chapter 12 sets out what such a service may and may not do.

Reminders and no-shows

Chapter 6 reported the Cochrane review of text message reminders and its authors' hedge that the evidence is insufficient to conclusively inform policy decisions142. The 2018 review of 105 studies, with its average of the order of 23%, found the rate highest in Africa (43.0%) and lowest in Oceania (13.2%)142. In a randomized quality-improvement initiative at one safety-net primary care clinic, adding model-driven live reminder calls for appointments with a predicted no-show risk of 15% or higher to standard automated reminders lowered the no-show rate to 33% against 36% in the standard-care group163. The study was not of an AI voice agent.

A 2016 systematic review found 31 randomized trials of reminder systems and reported that only one of them did not show a significant reduction in non-attendance164. A 2026 meta-analysis of 10 randomized trials of reminders for hospital outpatient appointments, in nine countries, pooled a risk ratio for attendance of 1.11165. In the VA study that reported the 18.8% no-show rate, a centralized phone reminder lowered the rate only from 16.3% to 15.8%, in an uncontrolled before-and-after comparison162. In an August 2025 MGMA poll of 265 medical practices, 27% said no-show rates had increased relative to the prior year, 60% said they stayed the same and 13% said they decreased, which gives a direction, not a level166. The cases below take 0.5 and 3 points from the VA before-and-after and the randomized initiative.

Healthcare marketing and the HIPAA rule: authorization and business associates

Chapter 6 gave the HIPAA marketing rule in plain terms for dentistry, and the rule is the same for a clinic. The Privacy Rule defines marketing, in general, as making a communication about a product or service that encourages recipients to purchase or use it, subject to exceptions in the definition's second paragraph10. A covered entity must obtain an authorization for any use or disclosure of protected health information for marketing, except a face-to-face communication or a promotional gift of nominal value152. The definition also covers financial remuneration, meaning direct or indirect payment from or on behalf of a third party whose product or service is being described, and it excludes payment for treatment of an individual10.

HHS adds that covered entities may not sell lists of patients or enrollees to third parties without authorization from each person on the list11. We rest the rule itself on the regulation and take HHS's page only as a plain-words gloss. Whether a recall or newsletter list built from the chart is a use of protected health information for marketing, and whether a given message falls inside an exception, is for the clinic's counsel. A list handed to an agency, a lead-generation company or an advertising platform is a disclosure of that information to a third party.

Business associates, agreements and agencies

A business associate, in HHS's description, is a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information; a subcontractor that does so on behalf of a business associate is itself one55. A covered entity may disclose protected health information to a business associate if it obtains satisfactory assurances, documented in a written contract that meets the rule's content requirements55,148. The contract must establish the permitted and required uses and disclosures. It may not authorize uses that would violate the Privacy Rule if done by the covered entity, with narrow exceptions for the associate's own management and for data aggregation167. HHS publishes sample provisions but says that using them is not required for compliance and may not alone be enough to form a binding contract under state law168.

The question for a clinic is who sees the data. A marketing agency falls within HHS's description only if it creates, receives, maintains or transmits protected health information on the clinic's behalf. An agency that works from public information alone does not, and one that receives call recordings, form submissions or patient lists does. Of 28 software tools that small practices commonly use, 9 will sign a business associate agreement without a special tier, 15 only on certain plans and 4 offer none, according to the vendors' own pages, and 6 of the 15 plan-gated tools do not publish the price of the qualifying plan54. Chapter 10 sets that check inside its architecture.

HHS also says that no Security Rule standard requires a covered entity to certify its compliance and that it does not endorse private organizations' certifications of the Security Rule151. The FAQ speaks of the Security Rule, not all of HIPAA. A clinic that reads "HIPAA compliant" on a vendor's page has therefore read the vendor's description and, as to the Security Rule, not a certification that HHS recognizes. We do not use the phrase of any tool.

The proposed Security Rule changes on AI

In a proposed rule published January 6, 2025, HHS stated that electronic protected health information, including that in AI training data, prediction models and algorithm data maintained by a regulated entity for covered functions, is protected by the HIPAA Rules169. It also stated that a regulated entity's risk analysis must consider the type and amount of electronic protected health information accessed by an AI tool, to whom the data is disclosed and to whom the output is provided169. The Unified Agenda entry we read lists the notice of proposed rulemaking as January 6, 2025, the stage as long-term actions and the final action as July 2027, so on that entry the proposal was not a final rule170.

Online tracking and health data outside HIPAA

A clinic's website can pass information about its visitors to every third party whose script it loads. The rules below are described, not applied.

What HHS's bulletin says

The original content of HHS's bulletin on online tracking technologies by HIPAA covered entities and business associates was last reviewed on December 1, 2022171. The court order discussed below refers to the bulletin of March 18, 2024, and the page we read now carries HHS's notice of that order12,172. The bulletin is guidance, not a rule12. It says that regulated entities are not permitted to use tracking technologies in a manner that results in impermissible disclosures of protected health information to tracking vendors, and gives as an example disclosures for marketing purposes without the individuals' authorizations that HIPAA requires12. It says tracking technologies on many unauthenticated webpages do not have access to protected health information, in which case HIPAA does not regulate their use, though in some cases they may have access to it12.

The bulletin also says that website banners asking users to accept or reject tracking technologies such as cookies do not constitute a valid HIPAA authorization, and that it is insufficient for a tracking vendor to agree to remove protected health information or de-identify it before saving it12. In one example, if a patient makes an appointment through a covered clinic's website and the site uses third-party tracking technologies that may send the appointment and the IP address to a vendor, the vendor is a business associate and a business associate agreement is required12. Our check found that Google Analytics, Meta Pixel, Calendly and Zapier offer no such agreement, according to the vendors' own pages54. That is a fact about those pages, not about any clinic's use.

What the court vacated

On June 20, 2024, in American Hospital Association v. Becerra, the U.S. District Court for the Northern District of Texas declared unlawful and vacated one element of the bulletin: the guidance to the extent it provides that HIPAA obligations are triggered where an online technology connects an individual's IP address with a visit to an unauthenticated public webpage addressing specific health conditions or healthcare providers12,172. The court denied a permanent injunction but granted vacatur172. Its footnote said the vacatur was not intended to limit the legal operability of other guidance in the HHS document172. The bulletin was not struck down; HHS's page says the court vacated the guidance only to that extent12. We found no later HHS statement on that page.

The statements above about banners, vendor promises and the appointment example are other guidance in the same document, which the court's footnote says the vacatur does not limit. Whether a given booking page counts as authenticated, and what the order means for it, is a question for counsel.

The FTC's cases against health apps and sites

The Federal Trade Commission has acted against companies that sent health information to advertising platforms. Table 3 summarizes four announcements. They are the FTC's descriptions, in press releases, of proposed orders; the conduct is as alleged in its complaints, no finding of a court is stated, and we did not check whether a court entered each order.

CompanyAnnouncedWhat the FTC allegedProposed outcomeCitation
GoodRxFebruary 1, 2023shared sensitive health information for years with advertising companies and platforms, contrary to its privacy promises, and failed to report the disclosures as the Health Breach Notification Rule requires$1.5 million civil penalty, in the FTC's first action under the rule173
BetterHelpMarch 2, 2023used and revealed email addresses, IP addresses and health questionnaire information to Facebook, Snapchat, Criteo and Pinterest for advertising$7.8 million to consumers174
Premom (Easy Healthcare)May 17, 2023violation of the Health Breach Notification Rule$100,000 civil penalty175
CerebralApril 15, 2024provided sensitive information of nearly 3.2 million consumers to third parties such as LinkedIn, Snapchat and TikTok through tracking tools on its website or appsnearly $5.1 million in refunds (for cancellation practices) and a $10 million civil penalty suspended after a $2 million payment176

Table 3. FTC announcements of proposed orders on health data and advertising tools, 2023 and 2024.

The FTC's complaint against GoodRx alleged violations of the FTC Act and the Health Breach Notification Rule173. We do not state whether any of these companies was a HIPAA-covered entity. The pattern the FTC alleged in several of them is health information passing to advertising companies and platforms, in Cerebral's case through tracking tools on its website or apps and in two of the cases contrary to what the company had promised its users173,174,176.

The Health Breach Notification Rule, as amended with effect from July 29, 2024, does not apply to HIPAA-covered entities or to any entity to the extent it acts as a business associate of one177. The FTC said the amendments clarify that the rule applies to health apps and similar technologies not covered by HIPAA178. Its definition of a breach of security includes an unauthorized acquisition of unsecured personal health record identifiable health information that occurs as a result of a data breach or an unauthorized disclosure, with notice to affected individuals no later than 60 calendar days after discovery177,178.

Washington's My Health My Data Act

Washington's My Health My Data Act defines consumer health data as personal information linked or reasonably linkable to a consumer that identifies the consumer's past, present or future physical or mental health status179. Beginning March 31, 2024, a regulated entity may not collect consumer health data except with consent for a specified purpose or as necessary to provide a product or service the consumer requested179. It may not share such data except with consent that is separate and distinct from the consent to collect, or as necessary for the requested product or service179. A small business, which the Act defines by thresholds that include the number of consumers whose data it handles, must comply with those provisions beginning June 30, 2024. That status delays the dates and does not remove the duties179.

The Act does not apply to information that meets the definition of protected health information under HIPAA and related regulations179. We think the exemption covers information that meets that definition, not covered entities as such, so data a clinic's website holds outside protected health information can remain in scope; counsel should confirm that reading.

Enforcement runs through the Consumer Protection Act: a violation is an unfair or deceptive act in trade or commerce for purposes of chapter 19.86 RCW179. Under RCW 19.86.090 a person injured in business or property may sue in superior court to enjoin further violations and to recover actual damages, costs and a reasonable attorney's fee179. The court may increase the damages up to three times, with the increase for a violation of RCW 19.86.020 capped at $25,000179. We did not research how the injury requirement and that cap apply to a claim under this Act. The Act is Washington law, and this paper does not survey other states' rules.

What a clinic's site may measure

The rules above leave a design question in place of a prohibition. Three questions sort the tools. Can the page or form carry protected health information, or health data in the Washington sense? Who receives what the tool collects? And is there an agreement, or a consent of the kind the law names, covering that recipient? A booking flow, an intake form, a patient portal login and a call-tracking line are places where the answer to the first question can be yes. We think a measurement design that records visits, calls and submissions as counts, keeps third-party scripts off the pages where the answer can be yes unless the vendor is under an agreement, and never sends form content or a page's health subject to an advertising platform, reduces the exposure the cases describe. It does not eliminate it.

Claims, reviews and testimonials in medical practice marketing

The AMA's Code of Medical Ethics, Opinion 9.6.1, says there are no restrictions on advertising by physicians except those that can be specifically justified to protect the public from deceptive practices9. It also advises that patient testimonials about a physician's skill or quality of service tend to be deceptive when they do not reflect the results that patients with comparable conditions generally receive9. The opinion is ethics guidance, not law.

The FTC's Health Products Compliance Guidance of December 2022 says that claims about the health benefits or safety of foods, supplements, drugs and other health-related products require substantiation in the form of competent and reliable scientific evidence, and that advertisers should not make claims through testimonials or expert endorsements that would be deceptive or could not be substantiated if made directly180. The guidance is written for health-related products. We infer that the guidance applies to a practice's clinical services, and counsel should confirm that.

We read two state rules, California's and Texas's. California's Business and Professions Code section 651 applies to every person licensed under Division 2, including physicians130. It requires "before" and "after" views in advertising to be comparable in presentation and to carry a statement that the same results may not occur for all patients130. It treats any statement, endorsement or testimonial likely to mislead because of a failure to disclose material facts as false or misleading130, and a claim of professional superiority as false or misleading unless it is relevant to the service and can be substantiated with objective scientific evidence130. A physician may state that the practice is limited to specific fields but may not state board certification or eligibility unless the board is an American Board of Medical Specialties member board, a board with equivalent requirements approved by the Medical Board before January 1, 2019, or a board or association with an ACGME-approved postgraduate training program130. Section 2271 makes advertising in violation of section 17500 unprofessional conduct for physicians181.

In Texas, Medical Board rule 22 TAC 164.3(b) deems it false, deceptive or misleading for a physician to provide a testimonial without a disclaimer or warning as to the credentials of the person making it and a disclosure of any compensation provided in exchange for it182. It deems the terms "board eligible" and "board qualified," or similar language calculated to convey the same meaning as "board certified," false, deceptive or misleading in physician advertising182. Counsel confirms the current text of both states' rules and how each applies to a given page.

Ratings, reviews and replies

The survey evidence on physicians and reviews dates from 2012. In a national survey conducted in September 2012, 65% of respondents were aware of online physician ratings, lower than their awareness for cars (87%) and for non-health-care service providers (71%)4. Among those who had sought ratings in the past year, 35% reported selecting a physician based on good ratings and 37% had avoided a physician with bad ratings4. Among those who had not sought ratings, 43% reported a lack of trust in the information on the sites4. A 2012 analysis of physician ratings found them generally positive: the mean was 3.93 on a scale of 1 to 5, and 45.80% of physicians received a 5 out of 579.

One later study measures bookings. In a retrospective analysis of 1.12 million new patient appointments and 12,882 physician reviews at one large single-specialty practice in the New York and New Jersey area, mean new patient volume decreased 2.3% to 2.6% after a 1-star review, with effects lasting at least 16 weeks81. Only 6.7% of the reviews were 1-star81. It is one practice in one metro area and one specialty, and the result is an association, not a trial. It shows that a review can move new-patient volume at a physician practice, not a rate for physicians in general.

A clinic's reply to a review carries the HIPAA risk shown in Chapter 6. Manasa Health Center, a New Jersey psychiatric provider, paid $30,000 and agreed to a corrective action plan over its replies to negative Google reviews, in which, according to the resolution agreement, it impermissibly disclosed the information of four patients124,126. HHS's page describes the payment as resolving potential violations, and the civil money penalty tiers are in Table 3 of Chapter 6. The FTC's reviews rule took effect on October 21, 2024 and applies to a clinic as to any business138. It prohibits fake reviews, compensation for reviews of a particular sentiment and certain threats or accusations aimed at suppressing a review83,84,85. The FTC cited civil penalties of up to $53,088 per violation in warning letters of December 22, 2025139.

AI in patient communication: disclosure laws, quality and HIPAA compliant AI

People search for "hipaa compliant ai" about 6,600 times a month27. We use the phrase because it is what they type, and we apply it to no tool. None of this section is legal advice.

Disclosure laws

California Health and Safety Code section 1339.75, known as AB 3030, applies to a health facility, clinic, physician's office or group practice office that uses generative AI to generate written or verbal patient communications pertaining to patient clinical information. Such a communication must include a disclaimer that it was generated by generative AI and clear instructions for reaching a human149. For audio communications the disclaimer is to be given verbally at the start and the end of the interaction149. The duty does not apply if a licensed or certified human health care provider reads and reviews the communication149. Patient clinical information does not include administrative matters such as appointment scheduling, billing or other clerical or business matters149.

On the statute's text, an AI receptionist that only books appointments is outside the duty, and an AI that drafts a message about a patient's symptoms or results is inside it unless a provider reviews the message. We give no operative date; counsel should confirm it.

Utah's general duty applies to a supplier using generative AI in a consumer transaction. The supplier must disclose that the individual is dealing with AI and not a human, but only if the individual asks or otherwise prompts it, and the request must be clear and unambiguous183. A separate duty has no ask-first condition. It requires an individual providing services in a regulated occupation to prominently disclose AI use in a high-risk interaction183. The statute defines that to include the collection of sensitive personal information such as health data and personalized medical advice or services184. The disclosure is given verbally at the start of a verbal interaction and in writing before a written one183. The page we read defines a regulated occupation as one regulated by the Department of Commerce that requires a license or state certification, and does not say which health professions that includes184.

Texas's H.B. 149 took effect January 1, 2026150. Its general consumer disclosure is addressed to governmental agencies150. A separate subsection requires the provider of a health care service or treatment in which an AI system is used to give the disclosure no later than the date the service or treatment is first provided150. Civil penalties are not less than $10,000 and not more than $12,000 for a curable violation and not less than $80,000 and not more than $200,000 for an uncurable one150. The attorney general may not sue before the 60th day after written notice, or if the violation is cured and a written statement given150.

The AMA's own policy, adopted November 12, 2024, says AI use should be clearly disclosed at the beginning of an interaction where patients engage directly with AI, and AI-generated patient-facing content should be disclosed within the content185. That is a policy position, not law.

Voice calls raise a separate rule. The Federal Communications Commission confirmed in 2024 that the Telephone Consumer Protection Act's restrictions on an artificial or prerecorded voice encompass current AI technologies that generate human voices, and that such calls require the prior express consent of the called party to initiate, absent an emergency purpose or exemption186. The ruling concerns calls initiated using an AI-generated voice, and a reminder call placed in such a voice is one. For inbound calls, the FCC said in a September 2024 notice of proposed rulemaking that the TCPA's requirements do not extend to technologies used to answer them187. The notice proposed to define an AI-generated call as an outbound one187. That is a proposal, not a ruling, and we do not follow what became of it. Chapter 12 treats the law of the call.

What the evidence says about quality

In a 2023 cross-sectional study of 195 public patient questions posted to an online forum, licensed health professionals preferred an AI chatbot's response to the verified physician's response in 78.6% of 585 evaluations142. The questions were public posts rather than messages from a practice's own patients, and the evaluators were licensed health professionals, not patients, so the result does not show that an AI agent is safe to answer a clinic's patients unsupervised. We think the design it points toward is draft and review.

A randomized or field study of an AI agent answering inbound calls or scheduling in a medical office is NF; the nearest evidence is outbound. In a prospective study at Mount Sinai with no control group, an AI voice assistant making pre-procedure calls completed 86.4% of its calls in the first phase47. In that phase 295 calls (36.6%) were fully automated and needed no human intervention, and about a third needed a quick protocol-driven nurse callback47. That is a guide to what an outbound agent can do, not a measure of a front desk answering inbound calls. On the practice side, an MGMA poll of April 2025 (375 applicable responses) found that about one in five (19%) medical group practices used some version of a chatbot or virtual assistant for patient communication188. In February 2026 practice leaders (177 applicable responses) named scheduling (31%) and calls (27%) as the areas where they were focusing AI and automation to improve access189. Those are self-selected members' answers about use and focus, not results.

Two sources describe the failure mode and who answers for it. NIST's generative AI profile defines confabulation as the production of confidently stated but erroneous or false content by which users may be misled or deceived190. In Moffatt v. Air Canada, 2024 BCCRT 149, the Civil Resolution Tribunal found the airline liable for negligent misrepresentation because its website chatbot gave wrong information, and wrote that the company is responsible for all the information on its website, whether it comes from a static page or a chatbot191. We make no claim about the weight of such a decision in an American court; we cite it for the principle it states, the one this paper relies on: a human organization stays responsible for what its machine says.

Measured: counting without exposing patient data

A clinic can count most of what Chapter 6's Table 5 lists for a dental practice: calls received and answered, booking requests by channel, new-patient appointments scheduled and kept by source, cancellations and no-shows, and a profile audit. Four counts are specific to a physician practice. The first is days to the next available new-patient appointment, the clinic's own version of the measure AMN used22. The second is calls and requests that ended because the caller's plan is not accepted, which shows whether the profile states accepted plans correctly. The third is referral source, because a physician practice can receive patients on other clinicians' referrals and a count by source should keep them apart from patients who found the practice themselves. The fourth is the monthly assistant check, with the same questions put to three assistants on a fixed date and with the limits of Chapter 4 in force1.

Each count can be kept as a number or a category. The source of a new-patient appointment is a label such as search, referral or assistant, recorded at booking, so no health information leaves the practice. The assistant check needs none either, because its questions are public ones a patient would ask. We think that makes it the one measure in this chapter with no privacy exposure. The statistics of small counts, and the reasons a clinic should read trends over quarters, are in Chapter 11.

A system design for a clinic

This section is an illustration. Suppose a primary care clinic with one site, four physicians, a nurse practitioner, a front desk of three, a billing lead and a part-time practice manager. The design follows the rules above and separates the places where patient information exists from the places where it must not. Table 4 sets out the layers.

LayerWhat it holdsUnder a business associate agreement?The rule it answers
Public site and profilesservices, hours, accepted plans, clinician credentials, general educational pagesNo patient data, so the content needs noneclaims that are true and substantiated (AMA Opinion 9.6.1; FTC)
Site measurementcounts of visits, calls and form submissionsThird-party scripts kept off booking and form pages unless under an agreementHHS tracking bulletin; Washington's My Health My Data Act
Booking and intake formsnames, reasons for visits, insurance detailsYesHIPAA Privacy Rule
Phone, answering service or AI receptionistcallers' names, reasons, call content, any recordingYes, with subcontractors namedHIPAA; AI disclosure laws; recording consent
Reminders and messagesappointment times, contact detailsYesmarketing rule; consent for AI-voiced calls
Recall listspatient listsYesmarketing rule and authorization
Reportingcounts by source category; the assistant checkNo patient data leaves the practicemeasurement

Table 4. Layers of a clinic's system, what each holds and whether it sits under an agreement.

The phone layer needs the most care. A person at an answering service says that the call is being taken for the clinic and that the caller is not speaking to a clinician. An AI agent says at the start that it is an AI assistant, which is our rule for our own service: callers and site visitors are always told. Neither gives clinical advice or interprets a result. The clinic's clinicians write a rule for urgent symptoms, and a marketer does not.

Any call recording is announced. Washington makes it unlawful to record a private telephone communication without the consent of all participants, and California's penal code punishes recording a confidential communication without the consent of all parties149,179. The federal wiretap statute is satisfied by the consent of one party, subject to an exception for a criminal or tortious purpose192. The hand-off to a staff member is written down, and a licensed provider reads any AI-drafted message about clinical information before it is sent, which is the condition of the California exemption above.

A vendor register lists every party that touches patient information, with its agreement and its subcontractors. We work in health care only under a signed business associate agreement; that is our practice, not a recommendation. The architecture is by Keegan Justis and the build is by Mike Golliet and Kyle Grenier.

The clinic's marketing needs no program, because a plan carries no patient information, so the least it can buy is a plan56,57. A program is for the tools that can tell a named person is a patient: the voice agent on patient calls, the site assistant, appointment forms, texts to patients, call tracking and recordings, and any connection to a scheduling system57. Its fee is shared by everything that runs on it, and no single gain is asked to carry it.

Essentials is $1,500 setup plus $699 a month and Integrated is $3,500 setup plus $1,199 a month, per location, after a compliance consultation57. A program is a compliance layer added to a plan, and the AI voice agent is a separate add-on that it does not include57. Four front-desk add-ons run inside it: reminders and rebooking, recall and reactivation, waitlist backfill, and intake with insurance capture57. HIPAA-ready is not a claim of compliance, which depends on the agreement and on how a tool is set up.

What a medical practice's system costs, and what a patient and a kept visit are worth

A practice buys a system: a plan that produces demand, an agent that answers it and, for the tools that touch patient information, a HIPAA-ready program as the compliance layer. Each fee is set against what it buys: the plan against new patients, the agent against answered calls and front-desk hours, and the program against everything that runs on it. Kept visits are set against the program and the reminders add-on together, because reminders are what the two reminder results measured.

Plans start at $999 a month per location56. The agent is $299 a month with 250 call minutes included, about 62 four-minute calls (250 / 4), and $0.39 a minute after that56. The worked cases here stay inside the included minutes, so the agent is $299 a month. The Essentials program is $699 a month after $1,500 setup57, which is $824 a month in the first year ($699 + $125, the setup spread over 12 months). It is the compliance layer that every tool handling patient information runs inside, and it does not include the agent57. The agent on patient calls, the site assistant, appointment forms, texts to patients and call tracking each need it57, and the front-desk add-ons run inside it at their own prices57, so its fee is shared by everything on it and no single gain is asked to carry it. With the reminders add-on at $149 a month, the program and reminders come to $973 a month ($824 + $149). Table 5 gives the stacks.

What is boughtMonthly costSet against
A plan alone, the marketing$999new patients
The agent, added to a plan the practice has$299answered calls and front-desk hours
HIPAA-ready Essentials, the compliance layer, first year$824everything that runs on it
Essentials after the first year$699everything that runs on it
Essentials and the reminders add-on, first year$973kept visits
The agent and Essentials, first year: the narrow case, a practice whose only patient-facing tool is the agent$1,123the agent's gains only
Plan, agent and Essentials, first year$2,122all of it
The same after the first year$1,997all of it

Table 5. Monthly cost of each purchase for one medical location (calc. from the published prices56,57), with the $1,500 setup spread over 12 months: the agent is $299 with 250 minutes included, Essentials is $699 + $125 in the first year, the reminders add-on is $149, and the narrow case is the agent and Essentials together ($299 + $824).

Table 6 lists the cited inputs for the primary care cases.

InputFigureSource
Payment per primary care office visit$186 mean, 201661
Payment per office visit to any medical doctor$347 mean, 2023193
Physician office visits per 100 persons320.7, 2019194
Share of physician office visits made to primary care50.3%, 2019194
New patients as a share of primary care office visits12.5%, 201950
Appointments per primary care physician51.1 a week, Epic data, January to May 2021195
No-show rate18.8% across 10 clinics of one VA medical center; of the order of 23% in a review of 105 studies162;142
Effect of reminder calls on the no-show rate16.3% to 15.8%, a before-and-after comparison at one VA medical center; 36% to 33% in one randomized initiative among high-risk appointments at one clinic162;163
Years a patient stays with a physician8 median, patients 65 and over, one state112
Front-office staff cost with benefits, health care and social assistance$34.59 an hour, June 2026196

Table 6. Cited inputs for the primary care cases, with the source of each in the last column.

The appointments figure comes mostly from health-system organizations, not small independent practices195. The new-patient share is a share of visits, not of patients, and the 8-year stay is for older patients in one state, so it may not fit a working-age practice. A person makes 1.61 primary care visits a year (3.21 visits a person x 50.3%), which at $186 a visit is $300 a year and $2,400 over 8 years (8 x $300). One physician has 221 scheduled appointments a month (51.1 x 52 / 12), which include the visits that are missed. The $300 a year, which is an average over everyone including people who made no visit, likely understates what a patient who seeks a practice out brings.

Table 7 sets each purchase against what it must win back, using Tables 5 and 6.

One physicianFour physicians
Scheduled appointments a month221886
Missed visits a month at 18.8% to 23%42 to 51167 to 204
Kept visits a month171 to 180682 to 719
Payments a month at $186 a kept visit$31,714 to $33,444$126,855 to $133,774
New-patient visits a month now, at 12.5% of kept visits21 to 2285 to 90
Plan, $999: new patients a month needed, first year only3.3 (15.6% more)3.3 (3.9% more)
The same, counting three years1.11.1
The same, counting 8 years0.40.4
Program and reminders, $973: kept visits needed5.2 (2.4 points off the no-show rate)5.2 (0.6 points off the no-show rate)
Reminders alone, $149, where a program is already in place: kept visits needed0.80.8
What one point off the no-show rate is worth a month$412$1,647
Kept visits if no-shows fall by 0.5 to 3 points1.1 to 6.6 ($206 to $1,236)4.4 to 26.6 ($824 to $4,942)
The agent's fee, $299, in front-desk hours at $34.598.6 a month (about 25 minutes a working day)8.6 a month (about 25 minutes a working day)
Whole system, $2,122: share of payments for kept visits6.3% to 6.7%1.6% to 1.7%
Kept visits needed, as a share of missed visits22.4% to 27.4%5.6% to 6.9%
New patients a month needed, first year only7.1 (33.2% more)7.1 (8.3% more)
The same, counting three years2.42.4
The same, counting 8 years0.90.9

Table 7. What each purchase must win back for one physician and for four physicians (calc., Tables 5 and 6). Kept visits are the fee over $186 ($973 / $186 = 5.2), points are the fee over what one point is worth ($973 / $412 = 2.4 and $973 / $1,647 = 0.6), and the agent's hours are its fee over $34.59 ($299 / $34.59 = 8.6; 8.6 hours over 21 working days is 0.41 hours, about 25 minutes a day).

The agent's own fee, $299 a month, is what 8.6 hours of front-desk time cost at $34.59 an hour ($299 / $34.59), about 25 minutes a working day at 21 working days; its answered calls are counted in Table 9. Those hours are the fee and not the talk time that covers it: on talk time alone the fee is covered at about 18 hours of calls a month, because minutes past the 250 included are billed (1,080 minutes cost $622.70, $299 + 830 x $0.39, and release $622.62, 18 x $34.59). The program is the larger fee, $824 a month in the first year (Table 5), shared by everything that runs on it. Set against kept visits, the program and the reminders add-on together are $973 a month ($824 + $149), because reminders are what the two reminder results measured. A four-physician practice covers the two with 0.6 points off its no-show rate, which is inside the 0.5 to 3 points of the two reminder results. For a practice that already has a program, reminders alone are $149 a month, or 0.8 kept visits ($149 / $186).

For a solo primary care physician the agent's own fee is the cost of 8.6 front-desk hours a month, or of the answered calls of Table 9. The program and reminders need 2.4 points off the no-show rate, toward the top of the 0.5 to 3 points of the two reminder results, and the top figure, 3 points, was reached only among high-risk appointments in one randomized initiative. So a solo office covers the program by running more than one tool on it: reminders and rebooking, recall and reactivation, waitlist backfill and intake with insurance capture run on the same layer57, beside the agent's answered calls.

The narrow case is a practice whose only patient-facing tool is the agent. It pays for both, $1,123 a month in the first year, and has only the agent's gains to set against it: the answered calls of Table 9, and front-desk hours, 32.5 a month ($1,123 / $34.59). A practice that runs reminders, the site assistant or the other add-ons on the same program spreads the program's fee across them.

Payment per visit changes every count, and Table 8 repeats three of them at the 2016 primary care median61, the 2016 mean used above61 and the 2023 mean for a visit to any medical doctor193. A few costly visits pull the mean above the typical visit, which the median shows.

Payment per visitA patient's payments in a yearNew patients a month to cover the planKept visits a month to cover the program and remindersKept visits a month to cover everything
Median primary care visit, 2016: $107$1735.89.119.8
Mean primary care visit, 2016, used above: $186$3003.35.211.4
Mean visit to any medical doctor, 2023: $347$5601.82.86.1

Table 8. The cases at three payment levels (calc., Tables 5 and 6 and the three payments per visit). The program-and-reminders column is $973 over each payment ($973 / $107 = 9.1).

Payments are what insurers and patients paid, before the practice's cost of the visit. We think a kept visit in a slot that is already staffed adds nearly all of its payment, a new patient who needs added capacity adds less, and a practice should use its own margin. The missed-visit counts join a no-show rate from one hospital system and a worldwide review with appointment volumes from a third source, so they are a scale, not a forecast. The 2023 mean for all doctor visits is nearly twice the 2016 primary care mean ($347 / $186).

At the loaded rate of $34.59 an hour196, staff hours released are worth $692 at 20 hours a month (20 x $34.59), $1,384 at 40 and $2,767 at 80. That rate is for office and administrative support occupations across the whole health care and social assistance sector, not for offices of physicians alone. A study of how many hours an agent releases in a medical office is NF; the nearest evidence is an MGMA poll of March 2026, in which practice leaders named eligibility and prior authorization in 45% of responses and scheduling in 31% as the time-intensive tasks staff handle on the phones197. That says which tasks, not how many hours, so the hours appear only as a break-even.

The calls against the agent

Call Box† reports that practices in its study of more than 1,700 US health care practices received an average of 299 calls per office location in February 2020 and 505 in June 2020, in a pandemic year198. Two measures say how many calls reach no one, from different settings: 8.3% abandoned in VA call centers in September 2016159, and 46% not answered by a person among Invoca†'s health care customers (calc. from its 54% answered)26. The second spans six fields of health care.

For conversion, Invoca† reports that 43% of the calls answered by a person were leads and 45% of those leads converted on the call26. Because the sources count new and existing patients together, a booking is valued at two ends: one primary care visit at $18661, and a new patient's first year of payments at $300 (calc., above). Table 9 runs the chain once for each miss rate; its rows are separate scenarios, not a range, and we do not average them.

Share of calls reaching no oneCalls a monthCalls missedBookings lost if they convert like answered callsValue, one visit each to a new patient eachShare the agent must turn into conversations to cover its own fee ($299)Narrow case: the same, against the agent and a program bought for it alone ($1,123)
8.3% abandoned (VA call centers, 2016)299254.8$893 to $1,44121% to 33%78% to more than all of them
8.3% abandoned (VA call centers, 2016)505428.1$1,509 to $2,43312% to 20%46% to 74%
46% not answered by a person (Invoca†, health care, 2026)29913826.6$4,950 to $7,9854% to 6%14% to 23%
46% not answered by a person (Invoca†, health care, 2026)50523245.0$8,361 to $13,4872% to 4%8% to 13%

Table 9. The call chain for one office location (calc. from the cited rates). A booking is valued at one primary care visit at the low end and a new patient's first year at the high end. The table treats missed callers as if they converted like answered callers, which the source does not show. The last two columns are the share of missed calls the agent must turn into conversations to cover a fee, as a range across the two ends ($299 / $1,441 = 21% and $299 / $893 = 33% in the first row), and the last is the narrow case.

The chain is set against the agent's own fee, $299 a month (Table 5), because answered calls are the agent's gain. At the VA rate the missed calls are worth more than the fee in every row, so whether they cover it depends on the share the agent holds: for an office with 299 calls a month the agent must turn 21% to 33% of the 25 missed calls into conversations, and at 505 calls a month 12% to 20% of the 42 (calc., Table 9).

Every share in the $299 column of Table 9 is below the 86.4% of its calls that an AI voice assistant completed in the first phase of a hospital study of pre-procedure calls47, and the highest, 33%, is below the 36.6% of those calls that were fully automated and needed no human intervention47. Neither study is a front desk answering calls and the VA is not a private office, so these are points of comparison and not a forecast of what an office would reach.

The program is not charged to answered calls: it is a layer that the site assistant, texts to patients and the front-desk add-ons also run inside, and Table 7 sets its fee, with the reminders add-on, against kept visits. In the narrow case, a practice that buys the program for the agent and runs nothing else on it, the agent must hold 78% to more than all of the missed calls at 299 calls a month and the VA rate, and 46% to 74% at 505 (calc., Table 9). The plan is not charged here, because its product is new demand, which Table 7 counts, and setting answered calls against the plan as well would charge a fee with nothing credited for it. The lead and booking shares are Invoca†'s figures for answered calls, and the kept visits and front-desk hours of Table 7 are separate gains on top of this one.

The system loses money in one corner: a practice that is not growing and whose intake and scheduling are already sufficiently optimized, with no new patient wanted, few calls unanswered and few visits missed. The cited base rates put a typical practice far from it: missed visits at 18.8% to 23% and calls that reach no one at 8.3% in VA call centers in September 2016159 and 46% among Invoca†'s health care customers (Table 9), none of them from a physician office. For any one practice the distance is something its own phone log and schedule can show, as Chapter 11 describes. The cases do not price errors: a wrong answer or a lost call that the agent causes has a cost this paper does not measure, and the margin it would have to consume is the gap between each gain and its fee in Tables 7 and 9. Chapter 12 treats quality and error at the front desk.

Limits of this chapter

The evidence for physician practices is thinner than the rules are detailed. The nearest evidence on AI agents is outbound, so the cases solve for the share an agent must hold instead of assuming one. This paper does not survey state medical boards' positions on AI, and the state laws we describe differ from one another.

What would settle the open questions is a panel of assistant answers for physician queries, repeated and checked against a register of practitioners, an audit of physician offices' telephone access measured over weeks, and a field study of an AI agent answering a clinic's inbound calls. Chapter 8 turns to law firms, where the binding constraint is the rules of professional conduct and the conversion point is intake.

Chapter 8

Law firm marketing, intake and the rules of professional conduct

Law firm marketing after Bates v. State Bar of Arizona

A lawyer's advertisement is protected speech and regulated speech at once, and one Supreme Court decision shows both halves. In Bates v. State Bar of Arizona, decided June 27, 1977, the Supreme Court held that a State could not prevent publication of a lawyer's truthful advertisement of the availability and terms of routine legal services, because applying the disciplinary rule against it violated the First Amendment5. The appellants had advertised a fee of $175 plus a $20 court filing fee5. The syllabus summarizes commercial speech as serving individual and societal interests in assuring informed and reliable decisionmaking, and as entitled to some First Amendment protection5.

The same opinion limits itself: "Advertising that is false, deceptive, or misleading of course is subject to restraint."5. Those two statements give this chapter its shape. A firm may tell the public what it offers and on what terms, and a state's rules decide what counts as misleading. We cite the ABA Model Rules of Professional Conduct and ABA Formal Opinions as the common reference. They are model or advisory texts, and a state's version of a rule may differ, as the advertising rules of four states read for Chapter 10 show. When we write "Rule 7.1 says", we mean what the Model Rule says.

The profession being marketed is large. The ABA's 2025 Profile of the Legal Profession reports 1.37 million lawyers in the U.S. in 2025, up from 1.35 million the year before199. The 2024 edition of the Profile counted 1,322,649 active lawyers as of January 1, 2024200, so the two editions do not give the same count for 2024, and we name the edition whenever we give a figure. That edition also reports growth of roughly 1% a year from 1,022,462 lawyers in 2000, an increase of 29% by 2024200. The Bureau of Labor Statistics counts 863,700 jobs for lawyers in 2025 and projects 5% growth from 2025 to 2035201, with a median annual wage of $159,670 in May 2025201. The ABA counts active lawyers and the Bureau counts jobs for lawyers, so we do not set one against the other.

We use these counts to place the profession. The Bureau's Quarterly Census of Employment and Wages counts offices. For 2024 it shows 188,520 private-sector establishments in offices of lawyers, as an annual average, with 1,098,918 people employed in them202. In the first quarter of 2024, 141,518 of the 187,335 offices of lawyers had fewer than 5 employees and 23,989 had 5 to 915,16,17. That is 75.5% of offices with fewer than 5 employees (141,518 / 187,335) and 88.3% with fewer than 10 (calc.). The unit is the establishment: a firm with two offices counts twice, and a lawyer with no employees is not counted at all, so these are not counts of firms or of solo practitioners.

The Census Bureau's advance estimate puts the 2025 revenue of employer firms in legal services at $446,060 million, up 9.9% on 2024203. That class is wider than offices of lawyers. The ABA's profile of the legal profession would give the share of lawyers in solo practice, but the page would not open for us, so we estimate from the Bureau's counts. No more than about three offices in four are likely to have a single lawyer (from the 75.5% of offices with fewer than 5 employees above; likely high, because that group includes offices of two or three lawyers and their staff).

Law firm SEO and attorney marketing: what three assistants named for lawyers

In our study the three assistants rarely named the same family law attorney: asked in all 16 cities, they named 240, and only 5.4% were named by all three1. Table 1 gives the three legal rows of the study.

Kind of businessCitiesBusinesses namedPer questionNamed by all threeNamed by one onlySource
Family law attorneys162401513 (5.4%)75.8%1
Personal injury lawyers3, small sample4615.33 (6.5%)76.1%1
Estate planning attorneys3, small sample4715.71 (2.1%)83%1

Table 1. Agreement among ChatGPT, Gemini and Perplexity on the lawyers they named, from our study. Rows marked "small sample" rest on three cities each.

Four features of the method qualify every row: each question was asked once, names were matched across assistants by spelling, up to eight names were read per answer, and answers obtained through an API are a baseline, not what every signed-in user sees1. Chapter 4 works through what each does to the figures. Personal injury lawyers and estate planning attorneys were asked about in three cities only, so those two rows are small samples. In a row of 46 names, one more business named by all three assistants would move the share by about 2.2 percentage points (1 / 46). We report these rows without ranking them against family law.

City-to-city spread is the better yardstick. For the ten kinds of business asked about in every city, the share of named businesses that all three assistants named ranged from 1.9% in Charlotte to 7.9% in Seattle1. The family law figure of 5.4%1 sits inside a range that one design produced across cities, so we draw nothing from the gap between it and other professions. The row supports only a narrower statement: in this run, a person who asked all three assistants for a good family law attorney in one of these cities would usually have seen a different set of names from each.

Behind the answers, the most-cited sites are not law firms' own. Across all 23 kinds of business in the study, the 25 most-cited websites drew 29% of citations and were directories, review platforms, list sites, publishers and a forum1. Those counts pool every kind of business, so they do not tell us which sites the assistants used for lawyers, and we did not classify the sites cited for lawyers by type or by whether a lawyer paid for a listing. For a law firm the consequence is practical: the pages an assistant reads include pages the firm does not write, so the rules on paid recommendations, below, bear on marketing as well as on ethics.

Table 2 lists the law-firm phrases in our Google Ads data from November 2026, with US monthly volume and cost per click. Chapter 5 sets out the data and its limits: the phrases are ones a firm owner or a vendor would type and not ones a client would, volume shows interest and not purchase, and cost per click is what advertisers pay and not what a client is worth.

Search phraseSearches a monthCost per clickSource
legal ai18,100$31.6227
law firm seo3,600$84.5327
attorney seo2,900$108.7827
ai for lawyers2,400$58.2127
ai for law firms1,900$33.8727
law firm website design1,600$96.0627
law firm marketing1,300$75.6827
attorney marketing720$134.6627
legal answering service720$214.5127
law firm crm390$108.0427
legal intake software140$106.4427
law firm intake70$92.4127

Table 2. US monthly search volume and cost per click for law-firm phrases, Google Ads data, November 2026. The phrases "lawyer seo" and "lawyer marketing" carry the same figures as "attorney seo" and "attorney marketing" in the data27, which suggests Google grouped them as variants, so each pair is shown once.

"Legal ai" is searched 18,100 times a month and "law firm marketing" 1,300 times27, so the AI phrase draws about 14 times the volume (18,100 / 1,300). Dentistry does not follow this pattern: "dental ai" draws 880 searches a month against 1,300 for "dental marketing"27. For lawyers, we think owners are searching for the tool before the system around it, which is the gap this chapter tries to fill. The highest cost per click in the table is for "legal answering service"27, 2.8 times that of "law firm marketing"27 ($214.51 / $75.68).

Truthful and still misleading: Rules 7.1 and 7.2 in attorney marketing

ABA Model Rule 7.1 says that "A lawyer shall not make a false or misleading communication about the lawyer or the lawyer's services." It adds that a communication is false or misleading if it contains a material misrepresentation of fact or law, or omits a fact necessary to make the statement considered as a whole not materially misleading6. The comments carry the rule past falsehood. Comment [2] says "Misleading truthful statements are prohibited by this Rule."204. The comment treats a truthful statement as misleading if there is a substantial likelihood that it will lead a reasonable person to a specific conclusion about the lawyer or the lawyer's services for which there is no reasonable factual foundation204.

Comment [3] applies the idea to claims a website makes every day. It says that an unsubstantiated claim about a lawyer's or law firm's services or fees, or an unsubstantiated comparison with other lawyers, may be misleading if presented with such specificity that a reasonable person would conclude it can be substantiated204. It adds that a truthful report of past results may mislead if it creates an unjustified expectation of the same results for other clients204. Read as drafting questions, the comments ask of each sentence on a page what a reasonable reader would conclude from it and what supports the conclusion. A page that ranks the firm, compares it with others, reports a result or quotes a price makes the kind of statement the comments describe.

Our study adds one new kind of statement: a firm named by an assistant may want to say so on its site. Being named is not a measure of quality, and our study did not check whether any named lawyer exists, holds a license or is good (Chapter 4). A page that reports an assistant's recommendation is a communication about the lawyer's services, and we think Rule 7.1 as written applies to it as it applies to any other communication. How to word such a statement, and whether a state's version of the rule allows it, are for the firm's counsel.

Rule 7.2(c) bars a lawyer from stating or implying certification as a specialist in a particular field of law unless an organization approved by an appropriate authority has certified the lawyer68. The communication must clearly name that organization68. Because the rule says "state or imply", the question is not confined to the word "specialist". Rule 7.2(d) requires any communication made under the rule to include the name and contact information of at least one lawyer or law firm responsible for its content68.

Paid recommendations and lead generation are the part of Rule 7.2 that we think assistants make newly relevant. Rule 7.2(b) bars a lawyer from compensating or giving or promising anything of value to a person for recommending the lawyer's services68. Its listed exceptions include the reasonable costs of permitted advertisements, the usual charges of a lawyer referral service and nominal gifts of appreciation68. Comment [5] says a lawyer may pay others for generating client leads, such as Internet-based leads, as long as the lead generator does not recommend the lawyer, the payment is consistent with Rules 1.5(e) and 5.4, and the generator's communications are consistent with Rule 7.1205. It adds that a lawyer must not pay a lead generator that states, implies, or creates a reasonable impression that it is recommending the lawyer, is making the referral without payment from the lawyer, or has analyzed a person's legal problems when determining which lawyer should receive the referral205.

These texts matter more when assistants read list sites and directories. A firm that pays for a listing, a lead or a "best lawyers" feature is paying a third party whose pages an assistant may cite, and the line the comment draws between generating a lead and recommending the lawyer is a line about what that third party says and implies. Our data do not show how many of the cited sites are paid placements. In the design below, no paid listing goes live until the page that displays it has been read before launch (Table 5). That is our design choice, not a statement of what the rule requires.

Solicitation is a separate rule. Rule 7.3(a) defines it as a communication initiated by or on behalf of a lawyer or law firm, directed to a specific person the lawyer knows or reasonably should know needs legal services in a particular matter, that offers to provide legal services for that matter206. Rule 7.3(b) bars solicitation by live person-to-person contact when a significant motive is the lawyer's or firm's pecuniary gain206. The bar does not apply if the contact is with a lawyer, a person with a family, close personal, or prior business or professional relationship with the lawyer or firm, or a person who routinely uses such legal services for business purposes206. The system in this chapter is built around contact that the prospective client starts. Rule 7.3 applies to outbound contact to a named person, and Chapters 10 and 12 treat the law that governs calls.

Referrals, directories and reviews: how people choose a lawyer

The paper's theory chapter (Chapter 2) treats professional services as credence goods: the buyer cannot judge quality before purchase and often cannot after. Day and Barksdale argued that purchasing professional services differs from buying physical products because services are typically intangible and customized, which makes evaluation before purchase difficult3. On that account a prospective client leans on signals. Two such signals matter for a law firm, the referral and the review.

Referral is the first of two channels that one survey reports. In a 2019 consumer expectations survey run by Clio† (law practice management software) and summarized by the Illinois Supreme Court Commission on Professionalism, 59% of clients sought a referral to an attorney and 57% searched for one on their own91. The two shares add to 116% (59% + 57%), so if both describe the same group, some respondents did both. The same source reports that 60% of Boomers were likely to value referrals against 46% of Millennials91.

These figures reach us through the Commission's summary, which gives no sample size, so they are that survey's results and not a measure of the size of either channel. In its 2024 survey, FindLaw† (a Thomson Reuters legal site) covered 2,000 US adults who had a legal need in the past 12 months87. Among the respondents who learned about the attorney they contacted, 48% had used referrals from family members, friends or trusted coworkers, the most popular offline resource87. The page does not describe the sampling, and using a referral as a source of information is not the same as finding the lawyer by referral.

Rule 7.2(b), described above, shapes what a firm may do about the first channel, because it bars paying a person for recommending the lawyer's services within its listed exceptions68. A referral program, or a reward to a source of referrals, is therefore a question under that rule before it is a marketing question, and what the state allows is for counsel. The assistants add a reading of the second channel. We infer that the signals a firm has built in review sites and directories are read twice, once by the person and once by the machine, if an assistant reads those places on the buyer's behalf.

The second signal is the review. A survey by BrightLocal† (local search) reported that 81% of consumers said reviews were "important" or "very important" in the financial and legal sector, up from 66% in 202286. Its 2024 edition reported that 75% of consumers always or regularly read online reviews when researching local businesses and that 3% never do207. The first groups legal with financial services, and the sources we read do not print either sample. FindLaw†'s 2024 survey reports that 82% of the respondents who contacted an attorney after learning about the attorney online used online reviews as part of their decision87. Nearly 40% said reviews were their primary source of information87. The base is that subgroup and not everyone with a legal need, and the page does not give its size. In our study the most-cited sites included review platforms1, so a review may be read by an assistant as well as a person. Our data do not trace any name to any review.

On replies to reviews, ABA Formal Opinion 496, dated January 13, 2021, says that a negative online review, alone, does not meet the requirements of permissible disclosure in self-defense under Model Rule 1.6(b)(5)208. It says that as a best practice lawyers should consider not responding to a negative post or review, because doing so may draw more attention to it208. It adds that lawyers may ask the host website or search engine to remove it208. In our design the reply policy is written in advance, with counsel, so that no reply is composed under the pressure of an unhappy client's post.

The FTC's rule on consumer reviews adds limits on businesses. It went into effect on October 21, 2024 and authorizes courts to impose civil penalties for knowing violations209. The FTC said the rule addresses reviews that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews, or who did not have actual experience with the business210. It prohibits compensation or other incentives conditioned on the writing of reviews expressing a particular sentiment, positive or negative210. It also prohibits using unfounded or groundless legal threats, physical threats, intimidation, or certain false public accusations to prevent or remove a negative review210. In warning letters of December 22, 2025, the FTC said violations can bring civil penalties of up to $53,088 per violation139. A legal threat against a reviewer therefore meets both a conduct opinion and a federal rule. The two texts address different questions, and we do not say how they combine in a particular case.

Legal intake is the stage at which a person who has found and chosen a firm either becomes a prospective client or does not. It is the conversion problem of this chapter, and the most direct evidence on it comes from secret-shopper studies run by Clio†. Its 2024 study involved emailing and phoning 500 law firms with a client inquiry23. Table 3 lists the results, with the 2019 figures as Clio† restated them.

Measure20242019, as restated by Clio†Source
Firms that responded to an emailed inquiry33%40%23
Firms that answered the phone call40%56%23
Firms that answered or called back52%not given23
Firms "essentially unreachable by phone" (Clio†'s term)48%not given23
Email responders that replied within eight hours84%not given211
Email responders that gave clear next steps or cost information18%not given211
Email responders that referenced similar legal cases as requested2%not given211
Shoppers unlikely to recommend the firms they contacted73%not given211

Table 3. Results of Clio†'s 2024 secret-shopper study of 500 law firms. The 2019 figures are Clio†'s own restatement inside its 2024 report.

Two cautions apply. The page we read does not say how the 500 firms were drawn or give their practice areas or sizes, so the results indicate what one set of shoppers met and are not a census of US law firms. The shoppers also posed as clients, so the figures describe the response to an inquiry, not the firms' own call records. With those cautions, the direction is plain: in this study, fewer than half of the firms answered the phone, and the 2024 figures were lower than the 2019 figures Clio† restated.

Clio†'s 2019 test emailed a random sample of 1,000 US law firms and phoned 500 of them. In it 56% of the firms answered the call, 39% of calls went to voicemail (57% of which were not returned within 72 hours) and 5% went unanswered212. It also reports that 73% of firms either picked up or phoned back212. The 2024 source we read restates only the 56%, so we do not set the 73% beside the 2024 figure of 52% as a trend. A report of a 2016 test by an American Bar Association committee (secondary) says one in three callers did not speak to a person and that callers were on hold 45% of the time213. That counts callers who reached a person, which is a different measure from Clio†'s.

Hennessey Digital† (a law-firm marketing agency) contacted 1,333 US law firms through their online lead forms in 2025. It reports a median response time of 13 minutes, a response from 74% of firms within seven days (59% in 2021) and no response from 26%214. Its firms came from marketing lists, and its test is of web forms and not the telephone.

The buyers' side is stated expectation, not behavior, and it is seven years old. In Clio†'s 2019 consumer survey of 2,000 US adults, 10% of people who reach out expected a response within an hour, 24% within a few hours and 45% within 24 hours212. A reply later than 24 hours therefore misses the expectation of 79%212. In the same Clio† survey, 57% of those who had ever shopped for a lawyer said they contacted more than one firm212. And 42% said that if they like the first lawyer they speak with they will not need to speak with any others212, which means they would stop looking, not that 42% hire whoever answers first.

Timing also matters at the start of an injury matter: in an Insurance Research Council poll fielded in June and July 2016, nearly two-thirds of claimants who contacted an attorney about an auto injury claim did so within one week of the accident215. The Council is funded by property-casualty insurers, and the poll is ten years old.

If the email figures refer to the same 500 firms, as Clio†'s wording suggests, then the 84% of responders who replied within eight hours is about 28% of all firms (84% x 33%), and the 18% who gave clear next steps or cost information is about 6% (18% x 33%).

We think this is an intake problem inside the marketing system, since a firm that is found and chosen has already paid for that moment. A person who cannot judge a lawyer judges what is visible: whether anyone answered, how fast and with what. Clio†'s data show what shoppers received and that 73% of them were unlikely to recommend the firms they contacted211. They do not show that any caller went elsewhere.

A study that follows law firm callers to engagements is NF; the nearest evidence follows a picked-up call to an appointment. In a secret-shopper study led by Ohio State University researchers, which called mental health clinicians in five states, a call that was picked up led to an appointment 24% of the time216. We estimate that about one answered call in four at a law firm could end in a first meeting (from that 24%; likely high for retained matters, because a first meeting is not an engagement). The estimate may err in either direction for the call itself, because those calls went to numbers listed in insurance directories, which the release says were often wrong.

A measurement of how response time changes whether a law firm caller books is NF; the nearest evidence measures speed in another setting. A 2011 news report quotes a Harvard Business Review article by Oldroyd and colleagues on a study of 1.25 million sales leads at 29 consumer and 13 business-to-business US companies217. In it, firms that tried to contact a lead within an hour of its query were nearly seven times as likely to qualify it as firms that tried even an hour later217. They were more than 60 times as likely as firms that waited 24 hours or longer218. We estimate that speed matters for a legal inquiry in the same direction (from the sales-lead ratios), and that the size is unknown. These ratios likely err high for law, because they compare firms that chose to respond fast with firms that did not, and sales leads are not people with a legal problem.

Intake is also where duties may begin. ABA Model Rule 1.18 treats a person who consults a lawyer about the possibility of forming a client-lawyer relationship as a prospective client219. It adds that even when no relationship ensues, a lawyer who has learned information from a prospective client shall not use or reveal it, subject to the exceptions Rule 1.9 gives for former clients219. What a person says to whoever answers a firm's phone, or types into its website, may therefore be information a lawyer has learned from a prospective client. Whether a given exchange with a nonlawyer or a machine amounts to consulting a lawyer is for the firm's counsel and the state bar, and we do not answer it.

A legal answering service, a receptionist or an AI intake agent collects facts: who is calling, what kind of matter it is, how to reach the person and when. In our design, none of them gives legal advice.

ABA Model Rule 5.3 concerns "a nonlawyer employed or retained by or associated with a lawyer". It requires a partner, and a lawyer with comparable managerial authority, to make reasonable efforts to ensure the firm has measures giving reasonable assurance that the nonlawyer's conduct is compatible with the lawyer's professional obligations220. It makes the lawyer responsible for a nonlawyer's conduct that would violate the Rules if the lawyer orders it or ratifies it220. Formal Opinion 512 adds that lawyers have additional supervisory obligations insofar as they rely on others outside the law firm to employ generative AI tools in connection with the legal representation, citing Rule 5.3(b)221. An outside answering service or an AI intake vendor is, in the ordinary sense, outside the firm. Whether Rule 5.3 reaches a particular vendor, and what supervision is reasonable, are questions for counsel.

Rule 1.18, above, protects information from a prospective client. Rule 1.6(c) requires a lawyer to make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client222. Comment [8] to Rule 1.1 says that to maintain competence a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology223.

Florida's Ethics Opinion 24-1, which the Florida Bar's Board of Governors voted unanimously to approve on January 19, 2024, speaks to intake directly224. It says that website chatbots used for client intake present additional risks, including that a prospective client relationship or even a lawyer-client relationship has been created without the lawyer's knowledge225. It also says that generative AI chatbots that communicate with clients or third parties must comply with restrictions on lawyer advertising and must include a disclaimer indicating that the chatbot is an AI program and not a lawyer or employee of the law firm225. It recommends that a lawyer obtain the affected client's informed consent before using a third-party generative AI program if doing so would involve disclosing any confidential information225. The Bar styles it an Ethics Advisory Opinion224, and the copy we read was hosted by a law school because the Bar's own page showed no text, so a reader should check the Bar's own version225.

ABA Formal Opinion 512 says that, because many of today's self-learning generative AI tools are designed so that their output could lead directly or indirectly to the disclosure of information relating to the representation of a client, a client's informed consent is required before inputting such information into the tool221. It also says that adding general, boilerplate provisions to engagement letters purporting to authorize the use of generative AI is not sufficient to obtain informed consent221. The State Bar of California's 2026 guidance says an attorney must not input any confidential information of the client into a generative AI solution that may present material risks to confidentiality or security, absent informed client consent as to the underlying risks226. That version replaced the November 2023 guidance226. Both texts speak of a client's information. Whether a prospective client's first call falls within their wording is left open here, as it is under Rule 1.18.

We think callers and visitors should also be told when a machine is answering, and the Florida disclaimer is one source of that expectation. Utah's definition of a high-risk artificial intelligence interaction includes personalized recommendations, advice or information that could reasonably be relied upon to make significant personal decisions, including legal advice or services184. Utah requires an individual providing services in a regulated occupation to disclose prominently when a person is interacting with generative AI in a high-risk interaction, verbally at the start of a verbal interaction183. The Utah page we read does not say which professions fall within the statute's regulated occupations184. Our own practice is to tell callers and site visitors they are talking to an AI assistant, whatever the state. Chapter 10 compares the disclosure rules across the four professions.

A firm that records intake calls meets a second body of law. Federal law lets a private person intercept a communication where the person is a party to it or one of the parties has given prior consent, unless the interception is for a criminal or tortious purpose192. Washington's statute makes it unlawful to record a private communication without the consent of all the participants179. California's Penal Code section 632 punishes recording a confidential communication without the consent of all parties149. The two state statutes ask for more than the federal text does, so which statute governs a given call is for counsel. Chapter 12 treats the law of the call.

ABA Formal Opinion 512, dated July 29, 2024, says that lawyers using generative AI tools must fully consider their applicable ethical obligations, including their duties to provide competent legal representation and to protect client information221. It says lawyers need not become experts in a tool but must have a reasonable understanding of its capabilities and limitations221. Because the tools are subject to mistakes, it says, uncritical reliance on their content can result in inaccurate legal advice to clients or misleading representations to courts and third parties221. Reliance on or submission of a tool's output without an appropriate degree of independent verification or review could violate the duty of competent representation under Model Rule 1.1221.

On client communication, the opinion says disclosure of AI use may be unnecessary depending on the circumstances221. Lawyers must still disclose their practices if a client asks how the work was done or whether AI was used, or if the client requires disclosure221. On fees, a lawyer who bills an hourly rate must bill for actual time even when a tool makes the work faster221. The lawyer may not charge a client to learn a tool the lawyer will regularly use221, and should treat a tool's cost as overhead to the extent it functions like equipping and maintaining a practice, absent a contrary disclosure made in advance221. These statements reach marketing, because a claim that AI makes a firm faster or cheaper is a claim about services and fees. Comment [3] to Rule 7.1 treats such a claim as potentially misleading if it is presented with specificity and not substantiated204.

The State Bar of California's 2026 guidance describes two related duties of competence, baseline technological competence about a system's capabilities, data sources, limitations and risks, and independent professional judgment in reviewing, verifying and correcting its outputs226. It says the duty of candor to the tribunal cannot be delegated to AI, and that an attorney should check for rules or orders in the jurisdiction requiring disclosure of AI use226. It also says an attorney must consider disclosing to the client an intention to use AI in the representation226.

One federal district has gone further for filings. The Northern District of Texas adopted Local Civil Rule 7.2(f) by orders of June 10, 2024, effective September 3, 2024227. It requires a brief prepared using generative AI to disclose that fact on its first page under the heading "Use of Generative Artificial Intelligence"228. It is one district's rule and not a nationwide requirement. Judge Brantley Starr's page on the court's site says he no longer requires his own certification on the use of AI in light of that rule229.

The error evidence shows what these duties guard against. Dahl, Magesh, Suzgun and Ho report, in the preprint, that legal hallucinations occurred between 58% of the time with ChatGPT 4 and 88% with Llama 2 when the models were asked specific, verifiable questions about random federal court cases230. A second study found that the AI research tools made by LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research and Ask Practical Law AI) each hallucinated between 17% and 33% of the time231. That is less than general-purpose GPT-4 but contrary to the providers' claims of being hallucination-free231. The tools may have changed since that study, whose authors call its 202 queries small232.

A later and different test, the Vals AI legal research benchmark of October 2025 (secondary), used 200 questions. All the AI products it tested scored 74% to 78% on an average weighted score, within 9 points of its lawyer baseline of 69%233. The second study and the Vals test used different questions and different products, so we do not average them or read either as a current error rate.

Courts have sanctioned lawyers for filing citations that an AI tool invented. In Mata v. Avianca (S.D.N.Y., June 22, 2023), Judge Castel imposed a $5,000 penalty, jointly and severally, on the two lawyers and their firm who filed non-existent cases generated by ChatGPT234. The order says there is nothing inherently improper about using a reliable AI tool for assistance, but that existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings234. In Wadsworth v. Walmart (D. Wyo.), attorneys filed motions in limine that cited nine cases, of which eight did not exist235. By order of February 24, 2025 the court revoked one attorney's pro hac vice admission and fined him $3,000, and fined two other attorneys $1,000 each235. In Johnson v. Dunn (N.D. Ala., July 23, 2025), the court publicly reprimanded three attorneys for ChatGPT-fabricated citations and disqualified them from the case236. It ordered publication of the order and referred the matter to the Alabama State Bar236.

Damien Charlotin's database (secondary) of legal decisions in which a court found or implied that a party relied on AI-hallucinated content listed 2,289 cases when we read it on 2026-11-10237. Its page calls it a work in progress that will expand as new examples emerge.

None of these cases concerned marketing or intake. They show what the verification duties are for. The opinions are written about client representation, and how far their language reaches a firm's marketing and intake tools is for counsel. One enforcement example does concern a legal-AI marketing claim. The FTC alleged that DoNotPay, which claimed to offer the world's first robot lawyer, did not conduct testing to determine whether its AI chatbot's output was equal to the level of a human lawyer and did not hire or retain any attorneys238. The company settled. The final order, approved 5-0 on January 16, 2025, requires $193,000 in monetary relief and bars advertising that the service performs like a real lawyer without sufficient evidence239.

How many firms use AI depends on who is asked and what counts as use. Table 4 sets four instruments side by side. The rows count different things, so they are not a trend and we do not combine them.

InstrumentWho was countedWhat was reportedSource
ABA Legal Technology Survey, 2024512 respondents to the online research volume30.2% of attorneys said their offices were currently using AI-based technology tools, against 17.7% of solo practitioners and 47.8% of firms of 500 or more41
Clio† survey, 20241,028 US legal professionals, lawyers and support staff, June 5 to 23, 202479% using AI "in some capacity"; 25% adopted widely or universally23
Clio†, 2023 baselineLaw firms19% were using AI in their practice23
Thomson Reuters Institute† survey, 2025, as reported by LawSites1,702 respondents across legal, tax, accounting, audit, risk, fraud and government sectors, legal professionals 41%26% of legal organizations actively using generative AI, up from 14% in 2024240

Table 4. Four instruments that count AI use in law, with what each counted.

The ABA figures are for solo practitioners and for firms of 500 or more41, and this paper reports none for a firm of the three-lawyer size described below. Clio†'s wording is "in some capacity" and its survey included support staff, so we use the survey's wording and do not read 79% as the share of lawyers using AI daily23. The Clio† 2023 figure counts law firms and the 2024 figure counts surveyed legal professionals, so the two may not be directly comparable23. The ABA's survey also found accuracy the most pressing concern about AI tools, cited by 74.7%, ahead of reliability at 56.3% and data privacy and security at 47.2%41.

Measuring a law firm's system without exposing its intake

A firm can count what happens at each stage without statistics: calls offered and answered, intake forms started and finished, consultations booked and held, and matters opened, each by source. Calls that rang and went unanswered deserve their own line, because in Clio†'s secret-shopper sample 48% of the firms were, in its term, essentially unreachable by phone23. Chapter 11 treats attribution and the small-numbers problem.

First, measurement tools see intake. An advertising tag or a call-tracking script that records what a prospective client types or says sends information that a lawyer may have learned from a prospective client to a third party. Rule 1.6(c), on preventing unauthorized disclosure, and Rule 1.18, on information from a prospective client, both bear on that. In health care, HIPAA permits disclosure of protected health information to a business associate on satisfactory assurances in a contract or other written arrangement55. For law, the nearest texts this chapter cites are Rule 5.3 and Opinion 512's passage on supervising outside vendors, and Chapter 10 adds the Texas, New York and California bar documents on generative AI. In our design, tracking tags stay on marketing pages and off the intake form and the call recording, and the figures that come from intake are totals, not content.

Second, the AI channel needs its own measure. The check we use is a monthly, dated set of questions put to each of the three assistants for the firm's practice areas and city, with the names and cited sites recorded. It inherits the method limits of our study1. One check proves little, and the minimum useful check is dated, repeated and covers all three assistants (Chapter 4). A result records what an assistant answered on a day. It guarantees no ranking, and the firm should not describe it as a statement of the firm's quality.

A three-lawyer firm, designed as a system (illustration)

Suppose a firm of three lawyers and two staff members in a mid-sized city. It practices criminal defense and personal injury, the two kinds of matter for which this chapter can cite a fee or a claim payment. It takes most new matters by referral and by search, and has a website, a handful of directory listings and no written intake procedure. The firm is invented for illustration; the figures in the cost section that follows come from cited sources. The design is ours. Table 5 sets out the design by stage of the paper's through-line.

StageWhat the system doesTexts that bear on itWhat it never does
FoundKeeps the firm's name, address, phone, hours and practice areas identical on its site and on the third-party sites assistants read; gives each practice area a page that answers the questions people ask; names a responsible lawyer on every pageRule 7.1; Rule 7.2(c) and (d)Implies a specialty no body has certified; goes live with a paid listing before its page has been read
ChosenShows how the firm works, what a first consultation involves and how fees are explained; asks clients for reviews with a neutral request; follows a reply policy written in advanceRule 7.1 comments [2] and [3]; Opinion 496; FTC reviews ruleOffers anything for a favorable review; answers a review with client information; promises results
ReachedAnswers every call and form: an AI intake agent says at the start that it is an AI assistant, collects name, contact details, kind of matter and a good time for a call back, and hands the facts to a named staff memberRules 1.18, 5.3 and 1.6(c); Opinion 512; the state's opinionsGives legal advice or says whether the person has a case; handles an urgent matter in place of a person
Within the rulesCounsel reads each page, script and vendor term before launch; the managing lawyer is the named responsible person; the script is reviewed against the state's version of the rulesRules 5.3 and 7.1; Opinion 512; Florida Opinion 24-1 and California guidance as examplesRecords a call without the notice and consent the state's recording law requires
MeasuredCounts calls answered and missed, forms, consultations and matters by source; runs a monthly, dated check of three assistantsRules 1.6(c) and 1.18 for what the tools seePuts tracking tags on the intake form; treats being named as proof of quality

Table 5. A system design for a three-lawyer firm, by stage. An illustration for an invented firm, not a recommendation for any state.

In the design, the website is a rep for the front office: it answers the questions people ask, collects the facts a staff member needs and hands them off to a person. The intake agent does not replace the lawyer. Under Rule 5.3 the managing lawyer remains responsible for the nonlawyer's conduct in the terms the rule gives. The design assigns that role to a named person who reads the script, the vendor terms and the monthly figures. Chapter 12 sets our prices against wage figures with the arithmetic shown.

What the system costs against collected hours, DUI matters and injury cases

We compare two purchases: the AI voice agent added to a plan the firm already has, and the lowest plan with the agent, which is the least a firm without one of our plans would buy. Our plans start at $999 a month, and the voice agent is $299 a month with 250 call minutes included and $0.39 a minute after that56. The included minutes are about 62 four-minute calls (250 / 4). The cases price the agent at $299, which holds while a month's calls stay inside them; a firm whose agent takes more pays $0.39 for each minute past them, or $1.56 for a four-minute call (4 x $0.39). The agent on an existing plan costs $299 a month, or $3,588 a year (12 x $299), and the lowest plan with the agent costs $1,298 a month ($999 + $299), or $15,576 a year (12 x $1,298).

A purchase is set against what it produces. The plan produces inquiries, so it is set against new matters. The agent answers inquiries, so it is set against matters won from calls that would have gone unanswered and against lawyer time. Table 6 lists the inputs with the source and limit of each.

InputFigureSource and limit
Hourly billing rate$300 median, 2023State Bar of Michigan survey of 2,191 attorneys; Michigan only, self-reported59
Average hourly rate of lawyers: criminal, personal injury, family, wills and estates$217, $335, $343 and $370, 2025Clio†52
Firms billing flat fees alone or beside hourly rates, 202459%Clio†52
Share of the working day billed38%; 25% for solo lawyers in 2022Clio†52,241
Billed work invoiced, and invoices collected88% and 93%Clio†52
Attorney's fee for a DUI$3,500 average, Washington, Spring 2020Washington State Office of Public Defense survey of private criminal defense attorneys; excludes investigators and experts242
Average auto bodily-injury claim paymentmore than $20,000 in 2022Insurance Research Council study of more than 7.4 million claims; a payment on a claim, not a lawyer's fee243
Usual personal injury contingency feeone-thirdState Bar of California advisory for fee arbitrators; California only244
Firms that answered a test call, 202440%; 48% essentially unreachable by phoneClio† secret-shopper study of 500 firms23
Shoppers who contacted more than one firm57%, 2019Clio† consumer survey212
Would stop looking if they like the first lawyer they speak with42%, 2019Clio† consumer survey212

Table 6. Inputs for the law case, with the source and limit of each. None is a price list for any firm.

An hour billed at $300 brings in $245.52 once 88% of billed work is invoiced and 93% of invoices are collected ($300 x 0.88 x 0.93). A lawyer who bills 38% of an eight-hour day, as the average does in Clio†'s data52, bills 3.0 hours and has 5.0 unbilled. One-third of a $20,000 injury payment is $6,667. Clio†'s own translation puts collected work at 2.4 hours per eight-hour day52. The Clio† family-law and wills-and-estates rates in Table 6 are above the $300 median, though they are national averages and the median is Michigan's, so fewer collected hours would probably cover each purchase in those fields. Because 59% of the firms in Clio†'s data bill flat fees52, an hourly rate times hours is not a safe estimate of what a client pays. We use it only to express the cost in hours.

What is boughtMonthlyA yearCollected hours a month to cover itDUI matters a yearInjury cases a year
Agent added to a plan the firm already has$299$3,5881.21.00.5
Lowest plan plus agent$1,298$15,5765.34.52.3

Table 7. What each purchase costs and the collected hours, DUI matters or injury cases that cover it (calc., from Table 6). Each matter counts as one year's fee: $3,500 for a DUI matter, which is the Washington average, and $6,667 for an injury case, which is one-third of $20,000.

Against Clio†'s 2.4 collected hours a day, the 5.3 collected hours a month that cover the lowest plan with the agent are about 2.2 days of one lawyer's collected time (5.3 / 2.4). Courts set the fee in a third kind of matter. In chapter 13 bankruptcy, the Eastern District of Wisconsin's presumptively reasonable debtor's-attorney fee is $5,500 for cases filed on or after May 1, 2024245. The Central District of California allows up to $7,000, or $8,500 where the debtor is engaged in a business, without a detailed fee application246. These amounts vary by district and are not an average of what clients pay.

How many matters the agent saves is the open link. A published count of calls a month at a law firm is NF, and we invent none. The nearest evidence is cited income and fees, from which matters can be bounded. The State Bar of Michigan survey reports a median 2022 gross income of $165,000 for 384 sole practitioners with an office outside the home59. At the $3,500 DUI fee that is ~47 matters a year ($165,000 / $3,500; likely high), and at $6,667 an injury case ~25 ($165,000 / $6,667). Against those, the 4.5 DUI matters or 2.3 injury cases that cover the lowest plan and the agent are ~9% of one such lawyer's year (4.5 / 47, and 2.3 / 25) and ~3% of the year of three (9% / 3). The matter counts are likely high because real caseloads include matters larger than these two. They also rest on one state's survey whose report does not say whether gross income is firm receipts or the lawyer's own income.

The system loses money only in one corner: a firm that is not growing and whose intake is already sufficiently optimized, with nearly every call and web form answered in time and no wish for more or better matters. The tests above put the firms they reached far from the intake half of that corner, with the sampling cautions given above. In Clio†'s 2024 test six firms in ten did not answer the call (100% - 40%), 33% answered the email23, and in Hennessey†'s 2025 test of firms drawn from marketing lists 26% did not respond to a web form214. This chapter cites no growth rate for law firms, so it cannot say how many firms sit in the no-growth half.

Billed hours are unlikely to be the limit for the typical firm in Clio†'s data, though not for every firm. Unbilled time is not idle time. Full-time private practitioners in the Michigan survey reported a median of 10 hours a week on non-billable activities59. A firm whose lawyers are full still gains from demand, in another form: it can choose the better-paying matter, raise its minimum fee or justify another lawyer.

The calls themselves give a reason to think the first conversation counts. The 2019 Clio† rows in Table 6, on shoppers who contacted more than one firm and on those who would stop looking if they liked the first lawyer they spoke with212, suggest that an unanswered call has competitors. They do not say how many of a firm's callers are in that position.

A per-matter fee for a divorce, a will or an estate-planning engagement is NF in any survey or other source we could open, so this chapter quotes none and makes no estimate of one. The nearest evidence is Clio†'s average hourly rates of $343 for family law and $370 for wills and estates in Table 6, which are rates and not matter fees.

What the evidence does not show

We estimate the share of callers who become clients, and the effect of speed on it, from a mental health study and a study of sales leads. The figures on unanswered calls come from one call per firm in two Clio† tests, a secondary report of one ABA committee's 2016 test and Hennessey†'s test of web forms. This paper does not measure the effect of any intake design. The conduct rules we cite are ABA model texts and formal opinions, two state bar documents and one district court's local rule, and none of them describes the rules of every state.

To settle the open questions, a panel of firms could log calls and inquiries by source for a year, with matters opened as the outcome. Repeated assistant checks would separate real disagreement from run-to-run variation, and a state-by-state table of the rules cited here would show where the Model Rules and the states diverge. Chapter 14 returns to this research agenda.

Chapter 9

Accounting firm marketing and the limits of capacity

Accounting firm marketing starts from a market of people and hours

The Bureau of Labor Statistics counts 1,595,200 jobs for accountants and auditors in 2025 and projects 5% growth from 2025 to 2035247. Median pay was $83,680 a year, or $40.23 an hour247. The Bureau counts jobs for an occupation, so the figure is not a count of the people who work in accounting firms247. The Bureau's Quarterly Census of Employment and Wages counts offices instead. For 2024 it shows 63,177 private-sector establishments in offices of certified public accountants, employing an annual average of 539,890 people248. It shows 37,615 establishments in tax preparation services, employing 99,216249, and 71,408 in other accounting services, employing 287,389250. We think tax preparation work is concentrated in the filing season, so an annual average counts fewer people than work in that class in the spring.

These count offices and not firms. In the first quarter of 2024, 44,800 of the 62,828 CPA offices had fewer than 5 employees and 9,333 had 5 to 916,17,18. That is 71.3% of offices with fewer than 5 employees (44,800 / 62,828) and 86.2% with fewer than 10 ((44,800 + 9,333) / 62,828). NASBA reports 650,667 actively licensed CPAs on August 27, 2026, from 53 of the 55 licensing jurisdictions251. The IRS lists 886,338 individuals with current preparer tax identification numbers for 2026, among them 209,076 CPAs, 71,966 enrolled agents and 26,256 attorneys, though some hold several credentials252.

NASBA counts licenses and not firms, and a count of CPA firms is NF at NASBA or the AICPA; the nearest evidence is the 63,177 offices248. We estimate that employer CPA firms number no more than that, since each has at least one office. It errs high, because a firm with several offices is counted once for each, and it leaves out owners with no employees. The Census Bureau's advance estimate puts 2025 revenue of employer firms in accounting, tax preparation, bookkeeping and payroll services at $220,030 million, up 2.3%203. These figures place the profession and do not size the market for a firm's services.

Graduates who earned a bachelor's or master's degree in accounting fell to 55,152 in the 2023-24 academic year, down 6.6% from the year before, according to the AICPA's 2025 Trends report253. Provisional federal data in the same release show 40,817 accounting bachelor's degrees, down 3.3%, and accounting or taxation master's degrees down about 15% to 14,335253. The two add to 55,152 (40,817 + 14,335), which matches the total.

Entry to licensure moved differently. In 2023, 42,626 new candidates entered the CPA Exam pipeline, the highest tally since 2016, and in 2024 the figure was 28,082253. That is a fall of 14,544 (42,626 - 28,082), or 34.1% (14,544 / 42,626). The comparison may overstate the change, because the AICPA says a surge of that kind typically comes the year before a major exam change253. Public accounting firms that responded to the AICPA survey reported hiring 11,985 new graduates in 2024, but low response meant the national total could not be projected with confidence253.

Earlier years show the same direction. Accounting bachelor's degrees numbered 47,067 in 2021-22, down 7.8% from the year before, and master's degrees fell 6.4% to 18,238254, which is 65,305 together (47,067 + 18,238). In 2019-20 the AICPA's 2021 Trends report shows decreases of 2.8% at the bachelor's level and 8.4% at the master's level255. Unique candidates sitting the CPA Exam, a different count from new candidates, fell from 72,271 in 2021 to 67,335 in 2022254.

These are counts of degrees, exam candidates and survey respondents, and NASBA's count of active licenses is for one date. They show a graduate inflow that fell in the 2023-24 school year and an exam-candidate inflow that is volatile, and they are one input to the next section's argument and not a measure of any firm's capacity.

Capacity as the constraint: why an accounting firm markets for fit, not volume

An accounting firm sells professional hours. In the short run those hours are set by the people the firm employs, and the pipeline figures suggest that adding people is not quick. The inferences from that premise to what a firm's marketing should do are ours.

First, hours are the scarce resource, so an inquiry the firm has no hours to serve does not become revenue; it becomes a delay, a disappointed prospect or a rushed engagement. Second, the quantity to maximize is therefore the share of the firm's hours committed to engagements it does well, at a fee that covers them, for clients who stay. Third, that share depends on fit, which we define as the match between what a prospect needs and what the firm does well, at the firm's fee and on its timetable. Fourth, a marketing system for such a firm should select as well as attract: its pages say who the firm serves and who it does not, its intake asks the questions that test fit, and its booking shows real availability. This is the working-rep idea in a specific form: the website qualifies inquiries before they reach a partner.

The evidence supports the premise in part. The pipeline figures show fewer graduates in 2023-24 and a volatile inflow of exam candidates, and busy-season studies show hours stretched when work peaks. In a longitudinal study of public accountants at one national firm, reported weekly workload rose from an average of 49 hours before the busy season to about 63 hours during it256. A survey of over 700 auditors found that they work, on average, five hours a week above the threshold at which they believe audit quality begins to deteriorate257. At the peak of busy season they often work 20 hours above it257.

A staffing firm's 2025 survey, as a trade publication reports it, found that nearly 80% of respondents worked more than 51 hours a week in busy season and 31.4% more than 61258. The first study is one firm in 2002, the second measures auditors' perceptions, and the third is a vendor survey of self-selected respondents. None is a time-record study of small practices, and this paper does not measure any firm's utilization, its rate of turning work away, or the share of inquiries that become engagements. The conclusion therefore rests on logic and on the premise that hours are fixed in the short run, with the busy-season studies supporting the peak.

A full calendar changes the form of the gain from demand. It does not make demand worthless. A firm with more inquiries than places can do three things that a firm with no spare demand cannot: choose better-fitting and better-paying clients, raise the minimum fee at which it takes a client, and fill the months outside busy season.

The first is choice. Fees per client differ by large factors in the sources we cite. The median revenue per advisory client of a CPA firm was $17,867 a year in the 2024 CPA.com and AICPA benchmark of 206 self-selected respondents, whose figures cover the client advisory line only113. The average base Form 1040 fee was $236 in the National Association of Tax Professionals' 2025 fee study60. These are different kinds of work with different hours, and this paper cites no hours per engagement, so the fee gap is not a gap in pay per hour. It does show why whom to take matters to a firm whose hours are fixed: when a place is full, an added inquiry replaces a client, and the gain is the difference between the two.

The second is price. Fees vary: among preparers who charge a minimum fee plus a charge for complexity, half charge between $150 and $328 for a base Form 104060, and the AICPA's MAP survey reported a median 6.7% rise in net client fees in fiscal 2024259. That survey reports growth in fees and not whether it came from price or from volume.

The third is the calendar. The IRS had received 99,802,000 individual returns through April 3, 2026, out of 144,992,000 through May 8, which is 68.8% (calc.)260,261. May 8 is the latest weekly count we read and not necessarily the season's final total. Tax professionals e-filed 75,316,000 of the 141,046,000 e-filed returns through that date, or 53.4% (calc.)261. We think a firm whose work is mostly returns is full in the spring and has room in the rest of the year, and that recurring work fills those months. The median typical monthly fee for a client advisory client was $3,000113.

One objection is that firms answer scarcity by raising prices, and that price is itself a filter for fit. We agree: a firm that prices by its scarce hours selects among prospects. The argument then applies to the signals that come before price. A prospect who cannot tell from the site what the firm does and roughly how it charges will not reach the price conversation, and the firm will not learn that it lost the prospect. A second objection is that a new or growing firm has idle hours and needs volume. That is correct for such a firm, and the argument concerns a firm whose hours are mostly committed. A firm with idle hours still gains from fit, because a client who does not fit costs more hours than the fee covers.

A third objection is that AI will expand capacity and loosen the constraint. The evidence is mixed and none of it concerns accounting. In an experiment on mid-level professional writing tasks, access to ChatGPT reduced completion time by 37%, or 10 minutes262. In a field experiment with consultants, those using AI completed tasks 25.1% more quickly on tasks inside the technological frontier263. They were 19 percentage points less likely to produce correct solutions on a task outside it263. If AI does raise capacity for some accounting tasks, hours become less scarce for those tasks. The firm still has to decide which clients to serve, and the review burden the tax-practice guidance describes below falls on the same professionals.

CPA marketing: found and chosen when no accountant was named by all three assistants

In our study, asked in all 16 cities, ChatGPT, Gemini and Perplexity named 272 accounting businesses, 17 per question, and none was named by all three; 90.4% were named by only one assistant1. Table 1 sets that row beside the study's two other finance rows.

Kind of businessCitiesBusinesses namedPer questionNamed by all threeNamed by one onlySource
Accountants16272170 (0%)90.4%1
Financial advisors1628117.65 (1.8%)87.5%1
Bookkeepers3, small sample5016.73 (6%)84%1

Table 1. Agreement among ChatGPT, Gemini and Perplexity on the accounting and finance businesses they named, from our study. The bookkeeper row rests on three questions.

The accountant row is the starkest in the study, and the method qualifies it. Each question was asked once, names were matched across assistants by spelling, up to eight names were read per answer, and answers obtained through an API are a baseline and not what every signed-in user sees1. Matching by spelling can split one business into two when the assistants write its name differently, which would lower measured agreement. The row is also one of four 16-city rows with 16 or more names per question1, and Chapter 4 shows that those four have the four lowest shares named by all three. That may be partly mechanical: the more distinct names the assistants produce between them, the less likely it is that all three name the same one.

For those reasons, and because the gaps between neighboring 16-city rows are smaller than the 6.0-point spread between cities1, we do not rank accountants against financial advisors or dentists. The row supports a narrower statement: in this run, no accounting business was named by all three assistants in any of the 16 cities, and about nine in ten were named by one assistant only. The bookkeeper row, from three cities, shows 3 of 50 (6%)1, a small-sample figure that says little. Being named is not a measure of quality, and the row does not show that any assistant is wrong or that any named firm is good or poor.

One possible reading of the zero is that "accountant" is not one service. A person who asks for a good accountant may want a tax return, bookkeeping, an audit, help forming a business or advice before a sale, so the assistants may be answering different questions from the same words. Our single generic question cannot test that. It does suggest where a firm has room to act. If an assistant writes from pages it reads, a page that states a specific service for a specific kind of client gives it something to match to a specific question. A page that says only that the firm offers full-service accounting matches almost any question and separates the firm from none. A test would put matched pairs of generic and niche questions to the assistants repeatedly, and ours has not done so.

Third-party pages are the other readable signal. The 25 most-cited websites in the study drew 29% of all citations and were directories, review platforms, list sites, publishers and a forum1. Those counts pool 23 kinds of business1 and say nothing specific to accountants. For an accounting firm the practical point is that the facts on the pages an assistant may read should match the firm's own site. Niche positioning and consistent listings are the two parts of "found" that the firm controls.

How clients choose an accountant is measured, but by instruments that count different things. Table 2 sets four side by side, and we do not combine them.

InstrumentWho was countedWhat was reportedSource
TaxDome†, 2025, a vendor surveyMore than 350 US business decision-makers58% found their current accounting firm through a referral and 17% through online search; 85% said responsiveness is "very" or "extremely" influential when selecting an accountant; 26% of business clients leave their firm because they have outgrown it97
Accounting Today and ADP survey, December 2019Accounting firms answering about their clients; the page gives no sample size56% of firm clients had been with their firm eight years or more, and 34% a decade or more; 63% of firms said 6% or less of their client base is new each year; unsolicited client referrals and referrals from fellow professionals were each named a top source of new clients by 79% of firms, solicited client referrals by 43%, and advertising by under 10%51
CPA.com CAS Benchmark Survey, reported November 2024Accounting practices, about leads for client advisory services only64% of all respondents named external referrals as a top source of leads and 67% existing tax clients; 19% named the firm website and 4% advertising96
IRS Comprehensive Taxpayer Attitude Survey, fielded September 30 to October 30, 20212,099 surveys of the general public by telephone and online random samplingAround half of taxpayers used a paid tax professional; paid tax professionals (88%) were among the most valuable sources of tax advice and information98

Table 2. Four instruments on how clients find and keep an accountant, with who was counted. The IRS survey reports how taxpayers value paid professionals as a source, which is not how they choose one.

In the three surveys that ask about sources, referrals lead, in one of them alongside existing clients, and no other channel comes close: advertising is named by under 10% of firms in one survey and 4% of practices in another, and the firm website by 19% of practices for advisory leads (Table 2). The 58%, the 79% and the 64% are different measures and are not averaged (Table 2). We think a referral often ends at a website or an AI answer where the prospect checks the firm, which makes the facts on those pages part of the referral channel. The 85% who name responsiveness in TaxDome†'s survey of stated attitudes (Table 2) are a reason to treat the first reply as part of the product, as the intake section below does.

Table 3 lists accounting phrases in our Google Ads data from November 2026. Chapter 5 sets out the limits: the phrases are ones a firm owner or a vendor would type and not ones a client would, volume shows interest and not purchase, and cost per click is what advertisers pay and not what a client is worth.

Search phraseSearches a monthCost per clickSource
ai accounting6,600$38.9527
cpa marketing1,300$12.3127
accountant marketing260$24.9727
accounting firm marketing170$42.1027
seo for accountants170$10.0827
ai for cpas90$49.3327
accounting firm website design70$13.2227
bookkeeping marketing70$27.6327
accounting lead generation70none reported27
cpa firm marketing50$52.6827
how to get accounting clients50$7.8527
cpa seo40none reported27
tax preparer marketing30$39.1927
accounting firm seo10none reported27

Table 3. US monthly search volume and cost per click for accounting phrases, Google Ads data, November 2026. "Ai for accountants" and "ai in accounting" carry the same figures as "ai accounting"27, and "accounting marketing" and "marketing for accountants" the same as "accountant marketing"27, which suggests Google grouped them as variants, so each group is shown once.

"Ai accounting" is searched about 5 times as often as "cpa marketing" (6,600 / 1,300) and about 25 times as often as "accountant marketing" (6,600 / 260)27. The remaining marketing phrases in Table 3 are each searched fewer than 200 times a month, volumes sensitive to rounding and grouping, so we read the table for its order of magnitude only. AI phrases outrun marketing phrases here, as they do in law, but the pattern is not universal: "dental ai" draws 880 searches a month against 1,300 for "dental marketing"27. The broad AI phrase is searched far more than any marketing phrase, while the phrase that names CPAs, "ai for cpas", is searched less than "cpa marketing" (Table 3), so we think the firms that would benefit from the design below may not be looking for it by name.

Advertising rules for accounting firms: the AICPA code and Circular 230 section 10.30

Accounting firm marketing meets two kinds of advertising rule, one from the profession and one from the Treasury. The AICPA Code of Professional Conduct's Advertising and Other Forms of Solicitation Rule (1.600.001) says that "A member in public practice shall not seek to obtain clients by advertising or other forms of solicitation in a manner that is false, misleading, or deceptive." It adds that solicitation by coercion, over-reaching or harassing conduct is prohibited8. State boards of accountancy can have advertising rules of their own, and one is described below. The Code text we read came from a hosted copy because the AICPA's own viewer is script-rendered8.

The Code's interpretation 1.600.010 says that promotional efforts are false, misleading or deceptive if they create false or unjustified expectations of favorable results8. That is the accounting counterpart of the lawyer's concern about past results. It also says that a member who performs services for a client or customer of a third party, such as a platform, should determine that the third party's promotional efforts comply with the advertising rule8. Members, it says, must not do through others what they are prohibited from doing themselves8. Whether this reaches a firm's own marketing vendor is for counsel.

Circular 230 section 10.30(a)(1) is the Treasury's text. It says that a practitioner may not, with respect to any Internal Revenue Service matter, "in any way use or participate in the use of any form of public communication or private solicitation containing a false, fraudulent, or coercive statement or claim; or a misleading or deceptive statement or claim."264. A page that promises a refund, a savings figure or a tax outcome raises the question these rules ask. Whether the firm's people are practitioners under the section is for counsel.

The AICPA texts cited here say less than the lawyer's Rule 7.1. The advertising rule we read is two sentences and does not mention rankings, comparisons or testimonials, but we do not infer that the profession permits what its rules do not mention. A state board can say more, and Texas's does. Rule 501.82(b)(1) of the Texas State Board of Public Accountancy treats certain advertising as false, fraudulent, misleading or deceptive265. Among other things, it covers self-laudatory statements not based on verifiable facts, untrue comparisons with other accountants, and testimonials or endorsements not based upon verifiable facts265. We read it on November 10, 2026 on Cornell's Legal Information Institute, which updates state regulations quarterly, and it is one state's rule. The FTC's rule on consumer reviews, described in Chapter 8, is a separate text. It went into effect on October 21, 2024 and authorizes courts to impose civil penalties for knowing violations209, and it prohibits compensation or other incentives conditioned on the writing of reviews expressing a particular sentiment210.

A firm that also gives investment advice meets a different rulebook, which this paper does not treat because financial advisors are not one of its four professions. The SEC's marketing rule for investment advisers (Rule 206(4)-1) became effective on May 4, 2021, with a compliance date of November 4, 2022266. It permits testimonials and endorsements in an advertisement if the adviser satisfies disclosure, oversight and disqualification provisions266. These include clear and prominent disclosure of whether the promoter is a client and whether the promoter is compensated266. FINRA's Rule 2210 requires member communications to be fair and balanced267, and whether it reaches a given firm depends on whether the firm is a FINRA member. FINRA's Regulatory Notice 24-09 says its rules are technology neutral and continue to apply when firms use generative AI268.

Reached: intake that qualifies before it books

First contact is where a found and chosen firm either gets a conversation or does not. A measurement of missed calls, call volume or lead conversion at accounting firms is NF in trade and vendor sources; the nearest evidence is two neighbors, which count different things, so we do not average them. The first is law, where Chapter 8 sets out secret-shopper tests of law firms: Clio†'s 2024 test calls were answered by 40% of 500 law firms, so 60% went unanswered (calc.)23. The second is Invoca† (call tracking), which reports that 54% of inbound calls to the health care organizations in its data were answered by a person, so 46% were not (calc.)26.

We estimate that a first call to a small accounting firm reaches no one at the first try somewhere between four and six times in ten (from the two neighbors above; likely high for a firm with a full-time receptionist). It may err low in busy season, when the partners are on returns. For the share of answered calls that lead to a first meeting we use the one neighbor that followed picked-up calls to appointments. In a study led by Ohio State University researchers of calls to mental health clinicians in five states, 24% of picked-up calls ended in an appointment216. We estimate about one in four for an accounting firm (from that study; likely high for engagements). A first meeting is not an engagement.

We infer, from the capacity constraint alone, that if hours are the scarce resource, intake has two jobs. It must respond, because a prospect who is ignored may go elsewhere. It must also qualify, so that the partner's time goes to conversations with prospects who fit. A form that asks what kind of business the prospect runs, which services it wants, how large it is and when it needs help does that work before a person reads the message. An AI receptionist that answers the phone can ask the same questions, say plainly that it is an AI assistant, and pass the answers to a named person. That is what our own AI voice agent does for a practice: it answers, qualifies, books and hands off to the office team. Chapter 12 treats the front desk across all four professions; the addition here is a capacity-aware booking page that shows when the firm can take new work and offers a place on a waiting list when it cannot.

The intake system collects the information needed to decide fit and nothing furnished for a return. It does not take Social Security numbers, income figures or documents. Those enter the firm through whatever secure channel it uses for clients, after an engagement exists. The rule is our design choice, drawn from the rule texts below and not a statement of what they require.

Beyond what an accounting firm may say, the binding constraint on its marketing systems is what it may do with what clients tell it. Four sources bear on that, each protecting a different thing in a different way. We describe each as the evidence states it.

Under 26 U.S.C. 7216, a tax return preparer who knowingly or recklessly discloses or uses information furnished for a return, other than to prepare the return, "shall be guilty of a misdemeanor". The statute sets a fine of not more than $1,000 ($100,000 for a disclosure or use to which section 6713(b) applies), imprisonment of up to 1 year, or both269. Section 6713 adds a civil penalty of $250 for each disclosure or use, capped at $10,000 for any calendar year270.

The regulations define who is covered and when consent is needed. Treasury Regulation 301.7216-1(b)(2) defines a tax return preparer to include any person engaged in the business of providing auxiliary services in connection with the preparation of tax returns271. That includes a person who develops software used to prepare or file a tax return and any Authorized IRS e-file Provider271. The definition therefore reaches beyond the person who signs the return. Whether a particular marketing vendor, answering service or AI tool is a person providing auxiliary services depends on what it does with return information. The design consequence is ours: the fewer vendors that ever see return information, the fewer such questions arise.

Regulation 301.7216-2 allows a preparer to share tax return information inside the firm within the United States without consent272. It requires the taxpayer's consent under 301.7216-3 before any disclosure to an officer, employee or member who is located outside the United States272. Where the people who can see the information sit is therefore a design question from the start. Regulation 301.7216-3(a)(1) says that unless section 7216 or 301.7216-2 specifically authorizes the disclosure or use, a preparer may not disclose or use a taxpayer's tax return information "prior to obtaining a written consent from the taxpayer"273.

For marketing, the practical question is whether a given use of information drawn from a client's return or file is one the statute and regulations reach, or one that section 301.7216-2 permits. We state only what the regulation says: where neither the statute nor section 301.7216-2 authorizes a use, written consent comes first.

The AICPA confidentiality rule and third-party providers

The AICPA's Confidential Client Information Rule (1.700.001) says: "A member in public practice shall not disclose any confidential client information without the specific consent of the client."8. Interpretation 1.700.040 addresses the vendor directly. It says a member should take one of two steps before disclosing confidential client information to a third-party service provider. One is a contractual agreement with the provider to maintain the confidentiality of the information, together with reasonable assurance that it has procedures to prevent unauthorized release. The other is the client's specific consent8. We found no reference to AI in the Code text we read, including that interpretation, and no Code interpretation on generative AI. An AICPA ethics-staff article in the Journal of Accountancy of August 5, 2026 says the Code still applies despite AI. It says members should ask, before pasting client data into any AI system, whether it is confidential client information and whether consent is required and has been given. They should also ask whether the use could result in disclosure to third parties274. The article is staff-written and non-authoritative.

In structure this resembles the business associate chain in health care (Chapter 10): a contract, or the client's consent, before the information goes to the vendor. Our check of 28 tools, which found that 9 will sign a business associate agreement without a special tier, 15 only on certain plans and 4 not at all54, concerned that health-law contract. It tells us nothing about which of those vendors will sign the confidentiality agreement the AICPA describes, and a firm has to check each vendor's own terms.

The FTC Safeguards Rule

The FTC says that section 314.2(h) of its Safeguards Rule lists examples of covered financial institutions that include tax preparation firms275. IRS Publication 5708 (Rev. 8-2024) says that under the GLBA and the Safeguards Rule, "tax and accounting professionals are considered financial institutions, regardless of size". It adds that the Rule requires a written, accessible information security plan276. Under section 314.4(j), a financial institution must notify the FTC as soon as possible, and no later than 30 days after discovery, of a notification event involving the information of at least 500 consumers277. That paragraph has been effective since May 13, 2024278.

What this means for forms, chat, AI tools and tracking

Taken together, the four sources push a system in one direction: know where client information goes, and keep the number of places small. We state the consequences as design choices and not as readings of the law.

On forms and chat, the texts cited here do not say whether what a prospect types into a marketing form or a website chat counts as information furnished for a return or as customer information under the Rule, and we do not guess. In our design the forms and the chat on marketing pages collect contact details and fit questions and tell the visitor not to enter return details. Documents move through the firm's secure client channel after an engagement begins.

On AI tools, Circular 230 section 10.22(b) presumes that a practitioner exercised due diligence when relying on the work product of another person if the practitioner used reasonable care in engaging, supervising, training and evaluating that person279. The regulation speaks of "another person"279. Section 10.35 requires the knowledge, skill, thoroughness and preparation the matter needs, and says competence can be gained by consulting experts or studying the law280. An AI tool that touches client information also raises the consent and contract questions above. A tool that touches no client information, such as one that drafts a page about the firm's services, raises fewer.

Tracking is the hardest. An advertising tag or analytics script on a page sends visit events to a third party. None of the texts cited here addresses whether such events are a disclosure under section 7216, and we do not claim they are. In our design, tracking stays on marketing pages and off any page or form where a client enters return information. Our check of 28 tools found that Google Analytics and Meta Pixel were among those that offer no business associate agreement54. That finding concerns a health-law instrument and does not transfer to tax.

AI for accountants and AI in accounting: demand, accuracy, guidance and what the surveys count

On demand, "ai accounting" is searched 6,600 times a month in the US27, and "ai for cpas" 90 times27. The accuracy evidence we cite dates from 2023 to 2025 and comes from academic studies. Wood and co-authors tested ChatGPT as of January 2023 on 28,085 accounting assessment and test-bank questions from 14 countries and 186 institutions. Students averaged 76.7 percent on assessments, against 47.5 percent for ChatGPT without partial credit and 56.5 percent with partial credit281.

Eulerich, Sanatizadeh, Vakilzadeh and Wood tested the CPA, CMA, CIA and EA exams. They report that the ChatGPT 3.5 model could not pass any exam, with an average score of 53.1% across all assessments282. After moving to ChatGPT 4, adding few-shot training and allowing reasoning and acting such as using a calculator, ChatGPT passed all the exams with an average score of 85.1%282. We think passing exams does not by itself show that AI can do the work of accountants. A 2025 paper in Artificial Intelligence and Law tested models on 390 questions, including perturbed versions, drawn from IRS-endorsed Volunteer Income Tax Assistance certification283. Top models such as GPT-4 and Claude 3.5 Haiku reached 83.33% accuracy on basic multiple-choice questions, against 54.17% for GPT-3.5283. These are certification questions and not client returns, and this paper cites no test of current models on client work.

Surveys report what the profession says it is doing, and they count different things. Table 4 lists two vendor surveys, which are not a trend and which we do not combine.

InstrumentWho was countedWhat was reportedSource
Intuit QuickBooks survey, April 2025700 US accounting professionals, in an online survey Intuit commissioned81% reported that AI has positively impacted their productivity (reported by respondents)284
Karbon†, State of AI in Accounting 2025Input from more than 500 accounting professionals across six continents70% expressed concerns about data security285

Table 4. Two vendor surveys on AI in accounting. The first is US-only and the second is global, so their figures describe different groups.

Intuit's figure is a perception of productivity and not a measured gain, and the Karbon† figure is a share of respondents who were concerned, not a rate of any incident. We use them only to show that practitioners report both benefit and worry.

On June 24, 2026 the IRS Office of Professional Responsibility issued an alert, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice"286. It ties its guidance to Circular 230 due diligence, section 10.27 and section 10.36286, and we think it is guidance and not a new rule. It says that technology serves as a powerful tool, not a substitute for professional judgment, and that final decisions must always rest with qualified professionals286. Due diligence, it says, requires verifying the accuracy of facts, citations and calculations produced by AI, and practitioners cannot rely solely on AI286. It adds that billing clients for time not actually spent, or double billing for AI-assisted tasks, may violate section 10.27 depending on the facts286. And it says firm procedures under section 10.36 should cover staff training, secure data handling and AI accuracy monitoring, and that outsourced or third-party AI tools should be vetted286.

On professional liability, a Journal of Accountancy column of April 1, 2025 states that no specific federal law or professional standard applicable to CPAs mandates disclosure when generative AI is used or if client data is used within it287. It is a column and not an AICPA rule, it predates the IRS alert, and a reader should check for later guidance before treating it as current. In our own practice, a caller or visitor who deals with our AI voice agent or chat is told it is an AI assistant, in every profession. Whether a firm tells clients that AI helped prepare their work is a separate question, and one for the firm's counsel and the guidance above.

Measuring a practice that sells hours

The counts a firm can keep without statistics are inquiries by source, inquiries that meet its fit criteria, consultations held, engagements signed, hours and fees per engagement, and clients retained from one year to the next. The first four measure the front of the system. The last two connect marketing to capacity, because an engagement that took twice the hours its fee covered, or a client who left after one year, is a result of how the firm was found and chosen as well as of how it worked. Chapter 11 treats retention, lifetime value and the small-numbers problem, which is acute for a firm that signs only a handful of new engagements a year.

The AI channel needs its own record. The check we use is a monthly, dated set of questions to each of the three assistants for the firm's services and city, with names and cited sites recorded. It carries the limits of our study1. One check proves little. A result records what an assistant answered on a day, it is not a guarantee, and the firm should not describe it as a statement of the firm's quality, for the reasons in the advertising section above.

A five-person accounting firm, designed as a system (illustration)

Suppose a firm of five people: two owners who are CPAs, a senior accountant and two staff accountants. It prepares returns and keeps books for owner-managed businesses and their owners in one metropolitan area. After its busiest weeks the owners can take only a limited number of new engagements, and we do not suppose how many. The fit problem is which inquiries fill the places. The firm is invented for illustration and the design is ours. Table 5 sets the design out by stage.

StageWhat the system doesTexts that bear on itWhat it never does
FoundOne page for each service and kind of client, each saying whom the firm serves and whom it does not; the firm's name, address, phone and hours identical on its site and on the third-party sites assistants readAICPA 1.600.001 and 1.600.010; Circular 230 section 10.30Describes the firm as serving everyone; claims savings or outcomes it cannot substantiate
ChosenStates scope, how fees are set and what working together involves; shows current availability and a waiting list; asks for reviews with a neutral requestAICPA 1.600.010; FTC reviews rulePromises a refund, a savings figure or a tax position; offers anything for a favorable review
ReachedA form and an AI voice agent ask only fit questions (kind of business, services wanted, size, timing), say they are AI, and route to a named partnerSection 7216; AICPA 1.700.001; FTC Safeguards RuleTakes Social Security numbers, income figures or documents in a chat, form or call
Within the rulesA written information security plan; a list of every vendor that ever sees client information, with the contract or consent that covers it; written consent for any use of return information beyond preparing the return; an AI use policy with review of outputsSection 7216 and regulations; AICPA 1.700.040; Safeguards Rule; Circular 230 sections 10.22, 10.35, 10.36Sends client information to a vendor with no contract or consent; relies on AI output without review
MeasuredInquiries by source and fit, engagements signed, hours and fees per engagement, retention; a monthly, dated check of three assistantsSafeguards Rule, for the tools that measurePuts tracking tags on a portal or on any page where return information is entered

Table 5. A system design for a five-person accounting firm, by stage. An illustration for an invented firm, not a recommendation for any state or firm.

The capacity logic runs through every row. Niche pages draw inquiries of one kind, and pages that say what the firm does and how it charges let a prospect who does not fit leave before costing anyone time. When the firm's places are filled, the page says so and offers a waiting list, and the system keeps marketing. A waiting list is a supply of inquiries the firm can choose among, so the next place goes to the best-fitting, best-paying client. It supports a higher minimum fee, because a firm with a queue can decline work below it. And it shows the owners which months the queue empties, so the pages can steer recurring bookkeeping and advisory work to the months outside busy season.

What the system costs against fees, clients and hours

Two purchases are compared, as in Chapter 8: the AI voice agent added to a plan the firm already has, and the lowest plan with the agent. Our plans start at $999 a month, and the voice agent is $299 a month with 250 call minutes included and $0.39 a minute after that56. The included minutes are about 62 four-minute calls (250 / 4). The figures below price the agent at $299, which holds while a month's calls stay inside them; a firm whose agent takes more pays $0.39 for each minute past them, or $1.56 for a four-minute call (4 x $0.39).

The agent on an existing plan costs $299 a month, or $3,588 a year (12 x $299), and the lowest plan with the agent costs $1,298 a month ($999 + $299), or $15,576 a year (12 x $1,298). Against the AICPA's MAP survey median of $189,695 in net client fees per full-time professional in fiscal 2022288, those are 1.9% ($3,588 / $189,695) and 8.2% ($15,576 / $189,695) of one professional's fees. At the survey's median billing rate of $159 an hour across all positions288, $1,298 a month is what 8.2 billed hours bring in ($1,298 / $159). Table 6 lists the fee inputs with their limits.

InputFigureSource and limit
Revenue per advisory client of a CPA firm$17,867 a year, median, 2024CPA.com and AICPA benchmark of 206 self-selected practices; client advisory services only113
Typical monthly fee for an advisory client$3,000, medianSame benchmark113
Advisory clients lost a year10%, medianSame benchmark113
Advisory clients served, and new clients67 and 14, mediansSame benchmark; its table gives no period for the new clients113
Hourly billing rate, all positions$159 median, fiscal 2022AICPA MAP survey of 1,117 firms288
Net client fees per full-time professional$189,695 median, fiscal 2022AICPA MAP survey288
Monthly bookkeeping and accounting fee, most common band$250 to $499, 2025Ignition† benchmark of 219 customers; a modal band, with solo bookkeepers in the sample289,290
Business tax return fee, most common band$1,000 to $1,499Ignition†289
Individual tax return fee, most common band$400 to $599Ignition†289
Base Form 1040 fee$236 average, 2025NATP fee study of preparers who charge a minimum fee plus a complexity charge60
Returning clients as a share of all clients87%NATP, read from a chart that does not say whether it counts clients or revenue114
Growth in net client fees6.7% median, fiscal 2024AICPA MAP survey259

Table 6. Fee and client inputs for the accounting case, with the source and limit of each. The Ignition† bands are price lists that firms reported, not what clients paid, and the AICPA figures are medians of firms that chose to take part.

An advisory client who brings $17,867 a year and leaves at a rate of 10% a year stays about 10 years (1 / 0.10) and brings about $178,670 (10 x $17,867) over that time. We infer this from the benchmark's two medians, and a quotient of two medians is not a median tenure. A monthly bookkeeping client in the most common band pays $3,000 to $5,988 a year ($250 to $499 a month). Firms report long tenure too: in the Accounting Today and ADP survey, 56% of firm clients had been with their firm eight years or more51. Table 7 turns those fees into the number of new clients a year that covers each purchase.

Kind of clientFee a yearNew clients a year to cover the agent on a plan ($299 a month)New clients a year to cover the lowest plan plus agent ($1,298 a month)
Advisory client of a CPA firm (AICPA benchmark)$17,8670.20.9
Monthly bookkeeping client, most common band (Ignition† sample, solo bookkeepers included)$3,000 to $5,9880.6 to 1.22.6 to 5.2
Business tax return$1,000 to $1,4992.4 to 3.610.4 to 15.6
Individual return, Ignition† benchmark$400 to $5996.0 to 9.026.0 to 38.9
Individual Form 1040, NATP average$23615.266.0

Table 7. New clients a year needed to cover each purchase, by kind of client, on one year's fee (calc., Table 6 and the costs above). A fee is revenue and not margin, and the table does not subtract the hours each client takes.

In Table 7, one new advisory client a year more than covers the lowest plan and the agent at a CPA firm, and stays about ten years, while a firm that sells only base 1040s needs about 66. The count depends on which clients the firm markets for, which is the case for marketing for fit. For a firm short of hours the table reads the other way: a firm that fills its scarce places with the higher-paying kind needs far fewer clients to cover the same purchase. Which kind of work earns more per hour is not in the table.

The corner where the system loses is a firm that is not growing, has no wish to choose among clients, raise its minimum fee or fill off-season months, and whose intake is already sufficiently optimized, with nearly every contact answered in time. The cited base rates put a typical firm only part of the way toward it. On growth, 48% of responding tax preparers in NATP's 2025 fee study said their practice was growing, so about half did not say so114, while the median firm's fees grew 6.7% in fiscal 2024259. The median CAS practice in the benchmark reported 14 new clients against 67 served and 10% turnover, and 10% of 67 is about 7 clients to replace (calc.).

Not growing is therefore common, and the loss needs every other condition at once. On intake, we estimated above that four to six in ten first calls reach no one, and a firm with a full-time receptionist sits below that. A firm that wants no better clients and no higher fees is the remaining condition, and this paper cites no measure of how many firms are like that.

What the evidence does not show

The rules we cite are federal statutes and regulations, a professional body's code read from a hosted copy, an IRS guidance alert and one state board's advertising rule. They do not include the other state boards' rules, which may differ.

A survey of firms could record capacity, inquiries and signed engagements across a year, which would test the capacity premise. A repeated panel of assistant answers, with matched generic and niche questions, would test the niche argument. A review of state board advertising and confidentiality rules would show where they depart from the AICPA's text. Chapter 14 returns to the research agenda.

Chapter 10

Compliance architecture: one system, four rulebooks

Four rulebooks side by side

A dental office, a clinic, a law firm and an accounting firm run the same kind of front office. Each has a website, a phone line, a booking or intake form, online reviews and, in some offices, an AI assistant that answers some of the contacts. Each is bound by a different set of texts about what it may say, what it may do with the information it is given, whom it may pass that information to, and what it must tell people when a machine is involved. This chapter serves the "within the rules" part of the paper's through-line.

Chapters 6 to 9 explain each profession's rules in depth, with the enforcement cases, and this chapter does not repeat them. Table 1 sets the four subjects against the four professions.

ProfessionAdvertising claimsPatient or client informationThird-party tools and vendorsAI use
DentistryNo advertising that is false or misleading in any material respect; announce specialization as "specialist in" (ADA ethics code, not law)7Use or disclosure of PHI for marketing needs an authorization, with listed exceptions (HIPAA regulation)152PHI may go to a business associate if the practice obtains satisfactory assurance, documented in a written contract (HIPAA regulation)148No adopted ADA policy on AI was found; the ADA Council on Dental Practice's 2026 report proposes a statement on AI in dentistry to the House of Delegates, and the House's action on it is NF; the nearest evidence is the proposal itself146. Proposed Security Rule changes would require the risk analysis to consider the electronic PHI an AI tool accesses169. California's AI disclaimer law names health facilities, clinics, physician's offices and group practice offices, and our source does not list dental offices149
MedicineNo restriction beyond what protects the public from deception; testimonials that do not reflect the results comparable patients generally receive tend to be deceptive (AMA ethics opinion, not law)9The same HIPAA marketing rule152. Washington's act exempts information that meets the PHI definition179The same contract rule. In its example of an appointment made through a clinic's website, HHS's tracking bulletin treats the vendor that receives the appointment and the IP address as a business associate12California and Texas disclosure provisions reach physician offices, clinics and health care providers (Table 6)149,150. AMA policy favors disclosure at the start (policy, not law)185
LawNo false or misleading communication; no paying for a recommendation, with listed exceptions; no stating or implying specialist certification unless an approved organization has certified the lawyer (ABA Model Rules 7.1 and 7.2, model texts)6,68Reasonable efforts against unauthorized disclosure; duties to a prospective client who consults (Model Rules 1.6(c) and 1.18)219,222Supervision of nonlawyers the lawyer retains (Model Rule 5.3); added supervisory duties when others outside the firm run AI tools (Formal Opinion 512)220,221Informed consent before client information goes into many of today's self-learning tools; an appropriate degree of independent verification of output (Formal Opinion 512, which we treat as advisory)221. Florida: a client-facing chatbot says it is an AI program, not a lawyer225
AccountingNo false, misleading or deceptive advertising or solicitation (AICPA Code; Circular 230 section 10.30)8,264A preparer may not disclose or use return information without the taxpayer's written consent unless section 7216 or 26 CFR 301.7216-2 authorizes it (26 CFR 301.7216-3)273. AICPA: no disclosure of confidential client information without specific consent8Before disclosing to a third-party service provider: a confidentiality contract and reasonable assurance of safeguards, or the client's specific consent (AICPA 1.700.040)8. "Tax return preparer" includes a person who provides auxiliary services, including developing software used to prepare or file returns271IRS Office of Professional Responsibility alert of June 24, 2026: verify AI output; vet outsourced and third-party AI tools (guidance which, we think, applies existing duties)286. We found no reference to AI in the AICPA service-provider interpretation we read8. An AICPA ethics-staff article of August 5, 2026 says the Code still applies and asks, before client data goes into an AI system, whether it is confidential client information (staff-written, non-authoritative)274

Table 1. What the four rulebooks say about advertising claims, personal information, third-party tools and AI use. Each cell names the kind of text it summarizes. PHI is protected health information, the HIPAA term for the patient information its rules protect.

The first three columns have an entry for every profession, and the entries ask for similar things in different words: say only what you can stand behind, do not let information move without authority, and put a contract or a consent behind any outside party that handles it. The fourth column is thinner and newer, and its entries differ most in force. In the professional texts the party answerable is the dentist, the physician, the lawyer or the preparer.

What kind of text sits in each cell

The weight of a cell depends on the kind of text behind it, and Table 1 mixes six kinds. Regulations and statutes bind those within their scope: the HIPAA provisions, the Treasury regulations under section 7216 and Circular 230 are regulations, and the California and Texas provisions in the AI column are statutes148,149,150,264,273. Professional codes bind members or serve as models. The ADA and AMA texts are ethics statements, not law7,9. The AICPA Code binds AICPA members in public practice8. The ABA Model Rules are model texts, so a statement about them describes the model, not the law in any one state, which counsel must check6.

Opinions and guidance are a third and fourth kind. ABA Formal Opinion 512, dated July 29, 2024, is the ABA's own opinion; this chapter cites no state's adoption of it, so we treat it as advisory221. State bar opinions on AI are advisory in the same way. Florida's, titled an Ethics Advisory Opinion and approved by the Florida Bar's Board of Governors on January 19, 2024224, says on its face that advisory ethics opinions are not binding291. The IRS Office of Professional Responsibility's alert says that Circular 230 due diligence requires verifying facts, citations and calculations produced by AI286. We think it applies existing duties to AI and is not a new regulation. HHS's tracking bulletin is guidance, and a federal court vacated one element of it12.

The fifth kind is the proposed rule. HHS proposed Security Rule changes on January 6, 2025 that mention AI169. The Unified Agenda entry we read lists final action as 07/00/2027, so those provisions are not final170. The sixth is policy: the AMA's House of Delegates adopted its position on disclosure on November 12, 2024, and it states policy, not a legal duty185.

Where two cells look alike they may differ in force, so which text binds a practice in its state has to be checked separately. Table 1 is also incomplete: its cells summarize national or model texts, and the state texts we read for this chapter sit in Tables 2 and 3 below. Those are advertising rules of state bars, dental and medical boards and one accountancy board. This chapter cites no state dental, medical or accountancy board statement on AI in the front office.

In the AI column, the ADA's standards committee on dental informatics approved a white paper on artificial and augmented intelligence in dentistry in December 2022, with a stated scope of clinical disciplines145. Our review found no ADA policy on front-office or patient-messaging AI145. For accountants, a Journal of Accountancy column of April 1, 2025 states that no specific federal law or professional standard applicable to CPAs mandates disclosure when generative AI is used or if client data is used within it287. That column is one secondary source and predates the June 2026 IRS alert286,287. For AI and confidentiality the AICPA's ethics staff wrote on August 5, 2026 that members' core responsibilities of competence, due care, confidentiality and independence do not disappear because new technology is involved274. We found no Code interpretation on generative AI.

The rulebooks also differ in the route by which an outside party may receive information. For a vendor that performs a function for a practice, the HIPAA texts run on a contract148, and the AICPA interpretation offers a contract or the client's specific consent8. For lawyers, Formal Opinion 512 says that, given the risks of many of today's self-learning tools as it describes them at publication, a client's informed consent is required before client information goes into one221. It adds that boilerplate in an engagement letter is not enough to obtain it221. Florida's opinion, which we read in a copy hosted by a law school, recommends client consent where a third-party generative AI program would receive confidential information225. A system that serves all four therefore needs a record of each.

From four rulebooks to five design requirements

The five requirements below are our reduction of Table 1, not a list found in any single source, and a compliance officer may draw the lines differently. The claim is that one system can meet all five once, whichever profession it serves.

  1. Say only what is true and can be substantiated.
  2. Know where every piece of personal information goes.
  3. Have a contract, or a consent, behind every party that touches it.
  4. Tell people when they are dealing with a machine.
  5. Keep a named human responsible.

Say only what is true and can be substantiated

Every cell in the advertising column reduces to one standard with different edges: truth, completeness and a factual foundation for any comparison or expectation. The ADA bars advertising that is false or misleading in any material respect7. ABA Model Rule 7.1 adds omission: a communication is misleading if it omits a fact needed to keep the statement, taken as a whole, from misleading6. Its comment says a truthful statement can mislead204. It adds that an unsubstantiated comparison may mislead when presented with enough specificity that a reasonable person would think it can be substantiated204. The ADA's advisory opinion names statements likely to create an unjustified expectation about results, and objective claims of superiority that are not subject to reasonable substantiation7. The AICPA interpretation treats promotion that creates false or unjustified expectations of favorable results as deceptive8.

The first extension is reviews. The FTC's rule makes it an unfair or deceptive act to write, create or sell a review that materially misrepresents that the reviewer exists, used the service or had the experience described83. It is also an unfair or deceptive act under the rule to give compensation or incentives conditioned on reviews expressing a particular sentiment84. A review program is therefore part of the claims system.

State bars have said when asking clients for reviews is permitted. Texas Ethics Opinion 685 (January 2020) concludes that the Texas rules do not prohibit a lawyer from encouraging current and former clients to leave positive reviews or ratings online292. That holds provided the lawyer does not encourage statements that are false, misleading or unfounded292. North Carolina's 2018 Formal Ethics Opinion 1, adopted July 27, 2018, says a lawyer may ask present or former clients to post reviews on the lawyer's directory profile as long as there is no quid pro quo293. The opinion adds that a lawyer may not solicit, encourage or assist in the posting of fake, false or misleading reviews293. North Carolina is outside the five states in Table 2, and both opinions are advisory and speak to lawyers. For patients, a statement on asking a patient for a review is NF in HHS and OCR text; the nearest evidence is the OCR actions in Chapter 6, which concern replies to reviews. How the HIPAA texts bear on a request for a review is for counsel.

The second extension is statements made by a machine or by someone else. A British Columbia tribunal wrote that an airline is responsible for all the information on its website191. The tribunal added that it makes no difference whether the information comes from a static page or a chatbot191. It is a Canadian tribunal's decision, cited for how the argument was put and answered, not as United States law. Utah's code says so by statute within its scope: it is no defense to a violation of a statute the state's Division of Consumer Protection administers that generative AI made the statement294. Where a member performs services for a client or customer of a third party, the AICPA interpretation tells the member to determine that the third party's promotional efforts comply8. The interpretation's reason is that members must not do through others what they are prohibited from doing themselves8.

One system meets the requirement with a claims register. It is a single list of every statement the practice makes about itself: credentials, specialties, results, comparisons, prices, hours and insurance accepted. Each entry carries the evidence behind it, the licensed professional who approved it and a review date. Site pages, directory listings, review-reply templates, agency copy and the knowledge an AI assistant may draw on are generated from the register or checked against it. The aim is an assistant limited to registered statements, which has less room to invent a specialty. The register also holds what the practice says about the assistant itself. The FTC alleged in a complaint that DoNotPay did not test whether its chatbot's output equaled a human lawyer's238, and the final order requires $193,000 in monetary relief239. Our own copy describes the assistant as one that answers, books, qualifies and hands off to the office team, and promises no ranking, AI answer or revenue.

Where the state texts move the edges

The lawyer cells of Table 1 rest on model texts, and the states differ from the model and from each other. Tables 2 and 3 set out the state texts we read for this chapter. We read each on November 10, 2026, from the court's, bar's or legislature's own copy where it opened and from a commercial or university reproduction where it did not. The tables describe what each text says, not how any of them applies to a practice.

StateWhat the text saysVersion read and caution
CaliforniaRule 7.1 comment [4]: a truthful report of a lawyer's achievements, or a testimonial, may mislead if it leads a reasonable person to form an unjustified expectation of the same results for other clients in similar matters295. Rule 7.2(c): an advertising communication must include the name and address of a responsible lawyer or firm295. Rule 7.3(c): a written, recorded or electronic communication soliciting a person known to need legal services in a particular matter must carry the word "Advertisement" or words of similar import, with listed exceptions295Rules PDF headed "Current Rules 2026". The label attaches to a communication that solicits a particular person
TexasRule 7.04(a): unless exempt under Rule 7.05, file a copy of each advertisement of legal services, an application and a fee with the State Bar's Advertising Review Committee no later than ten days after it is disseminated296. Rule 7.05(b): information and links posted on a law firm website are exempt from that filing, except the contents of the homepage296. Rule 7.02(a): publish the name of a responsible lawyer and identify the lawyer's primary practice location296State Bar PDF headed "Including Amendments Effective January 31, 2022". Rule 7 was renumbered in 2021 and older hosts show Rule 7.07. An October 2024 amendment order did not list Rule 7, and we checked nothing later
FloridaRule 4-7.19(a): subject to the exemptions in Rule 4-7.20, file a copy of each advertisement with The Florida Bar at least 20 days before first dissemination297. Rule 4-7.20(g): information on the lawyer's website is exempt from that filing297. Rule 4-7.13(b)(9): a testimonial is deceptive or inherently misleading if, among other things, the lawyer wrote it, it was given in exchange for something of value, it does not represent what clients generally experience, or it lacks a disclaimer that the prospective client may not obtain the same or similar results297. Rule 4-7.12(a): include the name of a responsible lawyer or the firm and the city, town or county of an office297Chapter 4 compiled October 22, 2026. The website exemption is from filing only; the content rules still apply to a website
WashingtonRPC 7.2 is marked [reserved], so there is no separate advertising rule, and we think the general standard of RPC 7.1 governs298. RPC 7.3(a) permits a lawyer to solicit professional employment unless the solicitation is false or misleading, the person cannot exercise reasonable judgment, the person has said they do not want to be solicited, or it involves coercion, duress or harassment299. Comment [3] to RPC 7.1: a truthful report of achievements may mislead if it leads a reasonable person to expect the same results for other clients in similar matters300Court rules PDFs. That RPC 7.1 governs advertising is an inference from its text and not the words of the RPC 7.2 page
New YorkThe former Rule 7.1(f) (22 NYCRR 1200.7.1) required most advertisements to be labeled "Attorney Advertising" on the first page or home page301. The former Rule 7.1(e) required the disclaimer "Prior results do not guarantee a similar outcome" with statements about expected results, comparisons, client testimonials or quality characterizations301. The New York State Bar Association says the Appellate Division adopted changes to Article 7 effective June 1, 2026302The label and disclaimer are the former rule, shown as an example of what changes. We could not read the new text because the court's site returned an error, and this chapter describes only that it changed

Table 2. Lawyer advertising texts in five states, read November 10, 2026. Washington has no separate advertising rule, and New York's row shows a former rule.

Table 3 does the same for dentists, physicians and one accountancy board.

Profession and stateWhat the text saysCaution
Dentists and physicians, California Business and Professions Code 651Applies to every Division 2 licensee, which includes dentists and physicians. "Before" and "after" views must be comparable in presentation and carry a statement that the same results may not occur for all patients130. A claim of professional superiority is false or misleading unless it is relevant to the service and can be substantiated with objective scientific evidence130. A statement, endorsement or testimonial likely to mislead because of a failure to disclose material facts is a misleading claim130Read on FindLaw†, marked current as of January 1, 2026, because the legislature's site returned an error; FindLaw† warns that it may not reflect the latest version
Dentists, California, section 1680Advertising professional superiority, or a guarantee of any dental service or of painless dentistry, is unprofessional conduct131Same FindLaw† reproduction
Physicians, CaliforniaA physician may state that the practice is limited to specific fields but not that the physician is certified or eligible for certification by a board unless it meets listed conditions130. Advertising in violation of section 17500 is unprofessional conduct181Same FindLaw† reproduction
Dentists, Texas22 TAC 108.54(a): "specialist" only in recognized specialty areas that a board certifies and the ADA's accreditation commission accredits132. 108.57(c)(8): a testimonial from a person who is not a patient of record, or that is not readily subject to verification, or that omits disclosures about the person's identity and credentials, is listed as an example of misleading advertising133. 108.57(d): photographs of actual patients only with written consent, and the advertising must say "Actual results may vary"133Cornell's Legal Information Institute, which updates state regulations quarterly. 108.54 was amended effective April 1, 2021 and 108.57 dates from May 1, 2013
Physicians, Texas22 TAC 164.3(b)(1): "board eligible" and similar terms are false, deceptive or misleading182. 164.3(b)(3): a testimonial without a disclaimer as to the credentials of the person making it and a disclosure of any compensation is misleading182Cornell's Legal Information Institute; amended effective January 9, 2025
Dentists, FloridaStatute 466.019(3): an advertisement may not contain laudatory statements about the dentist or group of dentists134. Subsection (1) defines an advertisement to include websites and social media134Statute text from the state's rules gateway, history through chapter 2024-214
Dentists, New York8 NYCRR 29.5(b): a claim of professional superiority or special abilities is unprofessional conduct, except that a dentist with approved specialty training may advertise the specialty135. 29.1(b)(12)(iv): testimonials are permissible only on conditions, including the patient's express written authorization, and copies of each advertisement are kept for one year after its last appearance136Cornell's Legal Information Institute, which updates state regulations quarterly
Dentists, WashingtonWAC 246-817-420(3): a general dentist may render specialty services but shall not advertise or represent as a specialist in the listed areas, and using a title such as specialist, expert, board certified or diplomate without completing the listed requirements is misleading137WAC text as read
Accountants, Texas State Board of Public Accountancy22 TAC 501.82(b)(1): self-laudatory statements not based on verifiable facts, untrue comparisons with other accountants, and testimonials or endorsements not based upon verifiable facts are false, fraudulent, misleading or deceptive265Cornell's Legal Information Institute; one state board

Table 3. Dentist, physician and accountant advertising texts in five states, read November 10, 2026. Each is one state's text and does not describe any other state.

For the claims register, first, a superlative meets a different test in each text. Florida's dental statute bars laudatory statements about the dentist134. California treats a superiority claim as false or misleading unless it is relevant and can be substantiated with objective scientific evidence130. The Texas accountancy rule names self-laudatory statements not based on verifiable facts265. The register therefore needs a field for the state whose text applies.

Second, many statements bring a duty with them: a filing (ten days after dissemination in Texas, 20 days before first dissemination in Florida), a label, a disclaimer or a written consent. Tables 2 and 3 cite each. The register stores the attached duty beside the statement.

Third, websites are treated in more than one way: Texas exempts website posts from filing except the homepage, Florida exempts a lawyer's website from filing but not from its content rules, and Florida's dental statute counts websites and social media as advertisements. Fourth, texts change. New York's lawyer advertising article was rewritten effective June 1, 2026, according to the State Bar Association302, and California's bar guidance on generative AI was replaced on May 14, 2026226. The register carries the date each state text was last read, and a lawyer confirms the current text before a statement goes live.

Know where every piece of personal information goes

The information column rests on rules about the movement of information. HIPAA requires an authorization for any use or disclosure of PHI for marketing, with listed exceptions152. A Treasury regulation bars a preparer from disclosing or using return information before obtaining the taxpayer's written consent, unless the regulations authorize it273. Model Rule 1.6(c) asks a lawyer for reasonable efforts to prevent inadvertent or unauthorized disclosure of, or access to, client information222. The AICPA rule bars disclosure of confidential client information without the client's specific consent8. Each is triggered when information goes somewhere, so a practice cannot show that it kept to the rule without knowing the movement.

The system answer is a data map. It lists every point where information enters the practice (the phone, a form, a chat window, a booking page, email, the review inbox, messages on a listing), classifies what could be said at each point (does it identify a person, does it state a condition, a matter or a return detail), and traces every destination. The map has three uses: it is what the practice brings to a lawyer when asking which rule applies, it decides which tools may run on which pages, and it produces the list of parties that the next requirement needs.

The third column asks for the same thing in three forms: the business associate agreement for health information, supervision of the nonlawyers and outside tools a lawyer relies on for law, and a confidentiality contract or the client's specific consent for accounting. One component serves all three: a vendor ledger that lists, for each tool or provider, its role, the information it can see, the kind of agreement that covers it, the plan that carries the agreement, and the date that position was last checked. A consent record sits beside it for the cases where the rule runs on consent.

Tell people when they are dealing with a machine

Disclosure rules differ in who they address and when they apply, and most ask for the same core act: a person learns that the party answering is an AI. One disclosure layer, written once and applied to the phone line, the chat window and text messages, meets that requirement across channels.

Keep a named human responsible

The professional texts in Table 1 hold a person answerable, and the system has to make that visible. A responsibility map names, for each automated action and each published claim, the human who owns it, the point at which a human reviews it, and the way a person takes over. An audit log records who changed what. Table 4 summarizes how the five components answer the five requirements.

RequirementSystem componentWhat a reviewer can ask to see
1. Say only what is true and can be substantiatedClaims register; pages, listings, review replies and assistant knowledge drawn from itThe entry for any public statement: its evidence, approver and review date
2. Know where information goesData map and page classesEach entry point, its destinations and the tools allowed to run there
3. A contract or consent behind every partyVendor ledger and consent recordFor each tool: role, data seen, agreement type, qualifying plan, date checked
4. Tell people it is a machineDisclosure layerThe opening statement for each channel, and a test record showing it plays
5. A named human is responsibleResponsibility map and audit logA name and a review point for each automated action; the log of changes

Table 4. The five design requirements and the component of the system that answers each.

The strongest objection to a single system is that it applies the strictest reading of every rule everywhere and so over-restricts. The objection has force in two places. A practice that announces AI on every call discloses more than Utah's general duty, which applies only when the individual clearly and unambiguously asks, requires183. A practice that keeps advertising tags off its booking pages gives up platform-reported conversions. We think the first cost is worth paying, because analyzing the trigger in each state for each channel is slower and more error-prone than saying the same thing every time. We treat the second as a real trade, and Chapter 11 sets out what a practice can still count. The professions differ in what goes into each component, such as who approves a claim, what kind of agreement a vendor signs and how the disclosure is worded. They do not differ in which components a system needs.

The business associate chain: what "HIPAA compliant AI" depends on

HHS describes a business associate as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting PHI55. A subcontractor that does the same on behalf of a business associate is itself a business associate55. Consider an AI answering service for a clinic. The practice contracts with a platform vendor, and the platform runs on a telephony carrier and calls a speech and language model provider, each of which may handle the caller's words. On HHS's description, we think the chain is as long as the path the data takes, and the arrangement is only as sound as its weakest link. "HIPAA compliant AI" averages about 6,600 US searches a month in Google Ads keyword data for November 202627. HHS says there is no standard in the Security Rule that requires a covered entity to "certify" compliance151, and this paper uses the phrase only as the name of a query, never as a label for a tool.

A covered entity may disclose PHI to a business associate if it obtains satisfactory assurance that the associate will appropriately safeguard it148. The assurance must be documented in a written contract that meets 45 CFR 164.504(e)148. That contract must establish the permitted and required uses and disclosures167. It may not authorize uses that would violate the Privacy Rule if the practice made them, with narrow exceptions for the associate's own management and for data aggregation167. HHS publishes sample provisions and says that using them is not required for compliance and may not alone be sufficient to form a binding contract under state law168. A practice cannot control whether the vendor will sign.

Our own check of 28 software tools small practices commonly use shows how much that varies. According to the vendors' own pages, 9 will sign a business associate agreement without a special tier, 15 sign only on certain plans, and 4 offer none54. Of the 15 plan-gated tools, 6 do not publish the price of the plan that qualifies54. Table 5 gives each share of the 28 (9 / 28, 15 / 28 and 4 / 28).

Position in our checkToolsShare of 28 (calc.)Tools named in our data
Will sign without a special tier932%Not named in this chapter54
Signs only on certain plans1554%ChatGPT, Gmail, Claude, Jotform, HubSpot, Supabase, Salesforce, Twilio, Vercel, Typeform, GoHighLevel, Acuity Scheduling, ElevenLabs, CallRail, Vapi54
Offers none414%Zapier, Calendly, Google Analytics, Meta Pixel54

Table 5. Business associate agreement positions of 28 tools small practices commonly use, from the vendors' own pages54.

The 28 tools are ones we chose as commonly used by small practices, not a sample of all tools or of practices. The check records what a vendor's page says, not what any practice has signed, and plan terms and pages change, so each position needs a "checked on" date, which the vendor ledger records. A tool that will sign says nothing about whether a given account is configured safely, which is why we say HIPAA-ready and not compliant: compliance depends on a signed agreement and on how the tool is set up. HHS adds that it does not endorse or recognize private organizations' "certifications" regarding the Security Rule151. It says such certifications do not absolve covered entities of their legal obligations151. That FAQ speaks of the Security Rule, not the whole of HIPAA.

In our design a tool in the third row of Table 5 is kept off any path where PHI flows, because, by their own pages, the contract the rule calls for is not on offer. A tool in the second row sits on such a path only on the plan that carries the agreement. The agreement is signed before information flows. For the plan-gated tools that do not publish a price, a practice has to ask what the qualifying plan costs.

Law and accounting have the same structure with different instruments: Model Rule 5.3 and Formal Opinion 512 for lawyers220,221, and the AICPA's contract-or-consent rule for accountants. For practitioners before the IRS, Circular 230 presumes due diligence when the practitioner used reasonable care in engaging, supervising, training and evaluating another person279. The passage we read speaks of "another person" and does not mention software or AI. The ledger records the instrument for each tool, whichever it is.

Tracking and analytics: the hard part of HIPAA compliant marketing

HHS's bulletin on online tracking gives a plain example. If a patient makes an appointment through a covered clinic's website, and the site uses third-party tracking technologies that send the appointment and the patient's IP address to a vendor, the vendor is a business associate12. A business associate agreement is then required12. In our 28-tool check, Google Analytics and the Meta Pixel are two of the four tools that offer no such agreement54. We call tracking the hardest piece of the architecture because it is the one place where a marketing tool and a privacy rule collide by default.

First, a tag moves information without a person deciding, each time, to send it, so the data map has to include flows that nobody sees. Second, whether a page visit is regulated depends on context the practice must judge page by page. HHS's bulletin says tracking technologies on many unauthenticated pages do not have access to PHI and are not regulated by the HIPAA Rules, but that in some cases they may12. In June 2024 the Northern District of Texas declared unlawful and vacated one element of the guidance12,172. That element was the "Proscribed Combination" of an IP address and a visit to an unauthenticated public page about specific health conditions or providers172. The court denied a permanent injunction and said the vacatur should not be read as limiting the legal operability of other guidance in the document172. HHS's page says the court vacated the guidance only to the extent it provides that HIPAA obligations are triggered by that combination12.

Third, the usual fixes do not work. HHS's bulletin says a banner asking users to accept or reject cookies is not a valid HIPAA authorization, and that it is insufficient for a tracking vendor to agree to remove PHI or de-identify it before saving12.

Rules outside HIPAA also reach the same flows. The FTC announced proposed orders against GoodRx, BetterHelp and Cerebral173,174,176. Its complaints alleged that each shared consumers' health information with advertising companies or other third parties, in Cerebral's case by using or integrating tracking tools on its website or apps173,174,176. The FTC announced its GoodRx order as its first enforcement action under the Health Breach Notification Rule173. That rule does not apply to HIPAA-covered entities, or to an entity to the extent it acts as their business associate177. In Washington, the My Health My Data Act bars sharing consumer health data without consent that is separate and distinct from the consent to collect179. It makes an exception where the sharing is necessary to provide a product or service the consumer requested179. A violation is an unfair or deceptive act under the state's Consumer Protection Act179.

For law and accounting the texts are less direct. This chapter cites no bar opinion, court decision or tax authority on advertising tags in legal or tax intake forms. The closest texts are Model Rule 1.6(c) on reasonable efforts222 and, for preparers, section 7216 and its regulations, which speak of disclosing or using tax return information271. The regulations define the preparer to include a person engaged in providing auxiliary services in connection with preparing returns, including one who develops software used to prepare or file them271. Whether an analytics vendor is such a person is a question this chapter's sources do not answer. A state bar opinion on tags in intake forms would bear directly on the question for lawyers, and a statement from the IRS or a state board of accountancy would do so for preparers.

The design we describe sorts pages into three classes. Public information pages state facts about the practice and its services. Action pages are where a person states a condition, a matter or a return detail: booking, intake and contact forms, chat, a patient or client portal. Authenticated pages sit behind a login. On action and authenticated pages, only tools covered by an agreement run, and no advertising-platform tag runs at all. Measurement comes from counting events inside systems the practice has a contract with, such as booking confirmations, call logs and form submissions, and from asking each new patient or client how they found the practice.

Call tracking is itself a vendor in the chain: in our check CallRail signs an agreement only on certain plans54. The cost is real, since the practice loses platform-reported conversions. What those numbers were worth is in doubt: in 15 U.S. advertising experiments at Facebook, observational methods often failed to produce the same effects as the randomized experiments104. Which pages fall in which class, and whether HIPAA, Washington's act or neither applies to a given page, are questions for the practice's counsel.

Telling people they are dealing with a machine: disclosure and recording law

No single text governs the disclosure that a machine is answering, and we think the texts we cite differ more in their triggers than in what they ask. Table 6 sets out the ones that bear on the four professions. The rows are not a complete survey of state law.

TextAddressed toWhen it appliesWhat it asks
Utah Code 13-77-103(1), the general dutyA supplier using generative AI with an individual in a consumer transactionOnly if the individual asks, clearly and unambiguously, whether they are dealing with AI183Disclose that it is generative AI and not a human
Utah Code 13-77-103(2), regulated occupationsAn individual providing services in a regulated occupationA high-risk AI interaction, including collection of sensitive personal information such as health or financial data, or personalized financial, legal or medical advice183,184Prominent disclosure, spoken at the start of a verbal interaction and in writing before a written one183
California Health and Safety Code 1339.75Health facilities, clinics, physician's offices and group practice officesGenerative AI patient communications about clinical information, not scheduling or billing, and not when a licensed or certified human provider reads and reviews the message149A disclaimer that AI generated it and instructions for reaching a human; for audio, spoken at the start and the end149
Texas Business and Commerce Code 552.051, from H.B. 149, effective January 1, 2026Governmental agencies in subsection (b); providers of a health care service or treatment in which an AI system is used in subsection (f)150For providers, when the service or treatment involves an AI systemDisclosure to the recipient or the recipient's representative by the date the service or treatment is first provided, or as soon as reasonably possible in an emergency150
California Business and Professions Code 17941A person using a bot to communicate online with someone in CaliforniaWhen the intent is to mislead about its artificial identity so as to deceive about the content of the communication, to incentivize a purchase or sale or to influence a vote149No liability if the bot discloses clearly and conspicuously that it is a bot149
Colorado S.B. 26-189, which repeals and reenacts the part of its statutes on automated decision-making technology, effective January 1, 2027 apart from sections that take effect on passage303Deployers of automated decision-making technologyA consequential decision in a listed domain, including health-care services and financial or lending services; legal services are not listed303A clear and conspicuous notice before use; it is not a general notice that a person is talking to an AI303
Florida Ethics Advisory Opinion 24-1224Lawyers using generative AI chatbots that communicate with clients or third partiesAny such chatbotA disclaimer that it is an AI program and not a lawyer or employee of the firm225

Table 6. Disclosure texts that bear on AI contact with patients and clients, by trigger: a request, a category of interaction, intent or the use of a chatbot.

Unlike Utah's general duty, its duty for regulated occupations has no ask-first condition183. Which professions fall within a regulated occupation depends on a definition tied to the state's Department of Commerce, and the page we read does not say which professions those are184. Texas's general disclosure section is addressed to governmental agencies, with the separate subsection for health care providers shown in Table 6, so, on the subsections we read, it is not a general duty on private businesses150. California's bot law covers a bot communicating "online"149. Whether it reaches a telephone voice agent is a legal question our source does not answer. Colorado's act was repealed and reenacted in 2026 with a 2027 date, and we describe the current enactment, not the 2024 text303. Chapter 12 applies these texts to the call.

On penalties, Utah lets the director of its Division of Consumer Protection impose an administrative fine of up to $2,500 for each violation of the chapter304. A person whose generative AI discloses clearly and conspicuously at the outset and throughout the interaction has a safe harbor from enforcement305. In Texas, the civil penalty for each violation a court determines to be curable runs from $10,000 to $12,000150. The attorney general may not bring an action before the 60th day after written notice of the alleged violation150. The attorney general may not bring one at all if the person cures the violation and provides a written statement within that time150.

At the federal level, the FCC confirmed in 2024 that the TCPA's restrictions on an "artificial or prerecorded voice" encompass current AI technologies that generate human voices186. Calls that use them therefore require the prior express consent of the called party to initiate, absent an emergency purpose or exemption186. The ruling concerns calls initiated with an AI voice.

In a September 2024 notice of proposed rulemaking (FCC 24-84) the Commission said the prohibition extends only to outbound calls that the caller makes or initiates187. It also said that the TCPA's requirements do not extend to technologies used to answer inbound calls187. That is the Commission's description in a proposal, not a holding or a binding rule, and we could not confirm whether its proposed AI-call disclosure rules were adopted. Its proposed definition of an AI-generated call is limited to outbound calls187. Whether and how the TCPA reaches a given inbound agent is for counsel. The statute allows a state-court action for actual monetary loss or $500 per violation, whichever is greater306. A court that finds a willful or knowing violation may increase the award to not more than three times that amount306.

Every call-handling system also raises the question of recording. Federal law lets a private person intercept a communication if they are a party to it or one party has given prior consent, unless the purpose is to commit a criminal or tortious act192. Five state statutes read differently from the federal text. Washington makes it unlawful to record a private communication without the consent of all the participants179. California punishes recording a confidential communication without the consent of all parties, with a fine not exceeding $2,500 per violation, or imprisonment, or both149.

Three more statutes, read on November 10, 2026, point the same way. Florida makes it lawful to intercept a wire, oral or electronic communication when all of the parties have given prior consent134. The usual label for that is "all-party consent", not the statute's own words. Maryland's section 10-402(c)(3) makes it lawful where the person is a party and all of the parties have given prior consent, unless the interception is for a criminal or tortious purpose307. Massachusetts defines interception as secretly hearing or secretly recording the contents of a wire or oral communication through an intercepting device by anyone other than a person given prior authority by all parties308. We think that, because the Massachusetts definition turns on secrecy, an opening statement that the call is recorded bears on it. Whether the statement does, and whether a transcript counts as a recording under a given state's statute, are questions for counsel. We did not read the statutes of Pennsylvania and Illinois, whose hosts would not open.

The single disclosure layer we describe follows the strictest timing in Table 6. At the first moment of every call and chat it says that the party answering is an AI assistant, says whether the interaction is recorded where recording is on, and says how to reach a person. For audio contacts that touch clinical matters it repeats the statement at the end, because California's audio text asks for both149. Our own practice is to tell every caller and site visitor that they are talking to an AI assistant, a design choice that goes beyond what Utah's general duty requires.

Named responsibility: why a small named team is a design choice

Several sources put responsibility on a named person or on a governance function. Model Rule 5.3 requires a partner or lawyer with comparable managerial authority to make reasonable efforts to ensure that the firm has certain measures220. The measures must give reasonable assurance that nonlawyers it retains behave compatibly with the lawyer's obligations220. Formal Opinion 512 warns that uncritical reliance on a generative AI tool's content can result in inaccurate advice or misleading representations221. California's State Bar guidance says the duty of candor to the tribunal cannot be delegated to AI226. In Mata v. Avianca the court said that existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings234. The court imposed a $5,000 penalty on two lawyers and their firm for filing non-existent cases generated by ChatGPT234.

Four state bar documents, read November 10, 2026, say it in their own terms, and all are advisory or policy statements. Florida Ethics Opinion 24-1 says lawyers may use generative AI but must protect the confidentiality of client information, provide accurate and competent services, avoid improper billing practices and comply with restrictions on lawyer advertising291. Texas Opinion 705 (February 2025) says that where a lawyer intends to give client confidential information to a generative AI program, Rule 1.05 will likely be implicated309. It also says a lawyer should always verify the accuracy of a tool's responses309. The New York State Bar Association's task force report, approved April 6, 2024, is a policy report and not an ethics opinion. It says the duty of confidentiality extends to what client information a lawyer may share when using certain generative AI tools310.

California's guidance, approved May 14, 2026, replaces the November 2023 version and addresses agentic AI226. It says attorneys must not deploy agentic systems in a manner that allows the system to make substantive legal determinations, communicate legal advice, prepare and file pleadings, or otherwise act in a representative capacity without meaningful attorney supervision and review226. Our design has the intake agent answer, book, qualify and hand off, and never give legal advice, so a named lawyer's supervision of its script is the condition these texts point to. Whether a given state's text reaches a given agent is for the firm's counsel.

Other texts point the same way. The IRS alert says technology is a tool and not a substitute for professional judgment286. California's health statute lifts the disclaimer duty for a generative AI message that a licensed or certified human provider has read and reviewed149. Model Rule 7.2(d) requires that a communication include the name and contact information of at least one lawyer or firm responsible for its content68. NIST's AI Risk Management Framework, which is voluntary, calls governance a cross-cutting function infused throughout the framework's other functions311.

The design consequence is a responsibility map, because a system that can act on its own needs limits on what it can do. OWASP defines excessive agency as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated outputs from an LLM312, and prompt injection as user input that alters the model's behavior in unintended ways313. An assistant limited to answering from the register, booking into open slots and handing off has a short list of actions, and each can be given an owner and a review point.

We argue from first principles for a small, named team to build and run the system. Every party in the chain is someone to contract with, supervise and reach. On HHS's description of a subcontractor, the chain grows with every layer of subcontracting55, and Circular 230's due diligence presumption speaks of reasonable care in engaging, supervising, training and evaluating the people relied on279. A practice can supervise only parties it can name. A small named team keeps the chain short and makes each link a person the practice can call, and it gives the question of who changed a setting a short answer. Limiting access to named people and logging what they do makes that answer checkable.

Technical Marketing is the marketing practice of the Strata Modus team, based in Thurston County, Washington. Our HIPAA-ready programs begin with a consultation, are architected by Keegan Justis, and are built only by Mike Golliet and Kyle Grenier. Managed marketing is led by Sezen Güner. The argument concerns supervision and traceability: this paper offers no measurement showing that team size changes compliance outcomes, and a small team has less redundancy than a large one. The questions a practice can put to any vendor, ours included, follow from it:

  • Who can access the practice's data, and who built the integration?
  • Is access limited to named people, and is it logged?
  • Which sub-processors handle the data, and under what agreements?
  • Which plan carries the business associate agreement, and is it signed before any information flows?
  • Who answers when something goes wrong?

What the architecture cannot do

The architecture does not make a practice compliant. Compliance is a property of how a practice behaves under the rules that apply to it, and a design can only make good behavior easier to carry out and to show.

Nor does the architecture change the penalty structure. In the inflation-adjusted table HHS published on January 28, 2026, the per-violation penalty runs from $145 to $73,011 where the entity did not know, and could not have known by reasonable diligence, of the violation129. It runs from $73,011 to $2,190,294 for willful neglect not corrected within 30 days129. HHS's 2019 notice said it would apply different annual limits to the four tiers, from $25,000 for no knowledge to $1,500,000 for uncorrected willful neglect, before inflation adjustment128. The current figures should be confirmed. We infer from the structure of the tiers, and not from any case, that the lowest tier turns on reasonable diligence, which is what a ledger and a log are meant to document.

The architecture cannot choose the rulebook. Which state, which board and which version of a model rule apply to a given practice is for counsel. Even within one profession the texts vary by forum: one federal district requires a brief prepared using generative AI to say so on its first page, and that is one court's rule, not a nationwide one228. Nor can the architecture keep pace with the rules without effort. HHS's Security Rule proposal is not final170. California's bar guidance on generative AI was replaced in 2026226. Colorado's act takes effect on January 1, 2027, apart from a few sections that take effect on passage303. The register, the ledger and Tables 2, 3 and 6 are living documents with review dates, and the responsibility map says who keeps them.

The architecture cannot prevent every error, human or machine. Errors by the machine are documented. A preprint reports legal hallucinations occurring between 58% of the time with ChatGPT 4 and 88% with Llama 2 when general-purpose models were asked specific, verifiable questions about random federal court cases230. Magesh and co-authors found that the AI legal research tools of two providers each hallucinated between 17% and 33% of the time231. Both concern legal questions, not front-desk assistants, and the tools in either may have changed since. They are why requirement 5 puts a human behind the machine.

Nor can the architecture settle the questions the texts leave open, which are noted where they arise. A regulator's statement or a court's decision would settle each. Until then, the practice's counsel or compliance lead decides how every rule in this chapter applies, and this chapter describes.

Chapter 11

Measurement and attribution for a small practice

Why marketing attribution is hard, even for the largest advertisers

Marketing attribution is the attempt to say which marketing activity caused a patient or client to arrive. Last-click attribution credits a sale to the last click before it. The evidence on how hard attribution is comes from experiments at large advertisers and platforms. Those studies compare observational methods with randomized tests, which is a different comparison, and none of them gives a figure for how far the last-click rule or a platform's own count misstates results.

Lewis and Rao report on 25 large field experiments with major US retailers and brokerages, most reaching millions of customers, and conclude that measuring the returns to advertising is difficult103. The median confidence interval on return on investment across those experiments is over 100 percentage points wide103. Their explanation is that individual-level sales are very volatile relative to the per capita cost of advertising, a coefficient of variation of 10 being common103. Informative experiments can therefore easily require more than 10 million person-weeks103. They also conclude that selection bias, due to the targeted nature of advertising, is a crippling concern for widely employed observational methods103. In plain terms, advertisers show ads to the people likeliest to buy, so crediting the ad with their purchases counts sales that would have happened anyway.

A practice is neither a large retailer nor a brokerage, and we take only the general point: when a sale is rare and varies a great deal from person to person, the signal from any one campaign is small next to the noise.

Blake, Nosko and Tadelis ran large-scale field experiments on paid search at eBay. They found that returns from paid search were a fraction of non-experimental estimates, and that brand-keyword ads, meaning ads on search terms that name the advertiser itself, had no measurable short-term benefits105. In an experiment that turned off brand-keyword paid search, 99.5% of the forgone click traffic was immediately captured by unpaid search traffic from the platform105. For non-brand keywords, new and infrequent users were positively influenced by the ads, but frequent users, whose purchasing was not influenced, accounted for most of the expense, so average returns were negative105. They also set the methods side by side. Typical OLS methods, a standard regression technique, gave a return on investment for eBay's paid search of over 4,100% without time and geographic controls and over 1,400% with such controls105. The experiment found -63%, with a 95% confidence interval of -124% to -3%105. These results are for eBay alone.

Gordon, Zettelmeyer, Bhargava and Chapsky used 15 US advertising experiments at Facebook, comprising 500 million user-experiment observations and 1.6 billion ad impressions, to compare experimental results with those of observational models104. The observational methods often failed to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables104.

Two further studies put the size of that gap in numbers. In Lewis, Rao and Reiley's Yahoo! experiment, the randomized comparison found a 5.4% increase in users searching brand keywords314. Observational matching on prior behavior gave estimates of search lift from 871% to 1198%314. In their second experiment, looking at the treatment group before and after exposure, without a control, would have overstated the ad's effect by 350%314. Gordon, Moakler and Zettelmeyer analyzed 663 Facebook experiments in a study published in Marketing Science in 2023315. The median randomized lifts were 29%, 18% and 5% for upper, middle and lower funnel outcomes315. A double/debiased machine learning method applied to non-experimental data gave 83%, 58% and 24%, and stratified propensity matching gave 173%, 176% and 64%315. In every pair the observational estimate sits above the randomized one. Together they show the direction and size of the gap: methods that compare people who saw an ad with people who did not, without randomizing, overstated the effect by multiples.

We estimate that the last-click rule overstates the effect it credits to a click-based channel by a multiple of the same kind (from the Facebook ratios; likely low). Those ratios are 2.9, 3.2 and 4.8 times (83% / 29%, 58% / 18% and 24% / 5%). The estimate is likely low because the methods in that study adjusted for over 5,000 user-level features315 and a last-click count adjusts for none. The estimate covers paid, click-based channels only. A last-click count gives no credit at all to channels that leave no click, such as a friend's recommendation or an AI answer.

Berman's work concerns the attribution rule itself. Last-touch attribution gives the whole credit for a sale to the final interaction before it; it is called last-click when that interaction is a click. In a game-theoretic model of advertising across several publishers, Berman shows that the popular last-touch method over-incentivizes ad exposures, often resulting in lower advertiser profits316. The Shapley value, a rule from game theory that shares credit among the touches, achieves higher profits316. He also cites a survey in which 54% of advertisers indicated that they used a last-touch method316. The model's result holds under the model's assumptions, and the passage cited gives neither the date nor the population of the survey. The family of methods that share credit among touches is searched under "multi touch attribution" about 720 times a month in the United States, at $161.53 a click27. "Marketing attribution" is searched about as often27. As Chapter 5 explains, those are phrases an owner or a vendor would type, and volume shows interest, not purchase.

What follows for a small practice is an argument. The strongest objection is that if the largest advertisers cannot measure what a campaign returned, a practice should not try. We think the objection misreads what the studies found hard to measure, which is the causal return to the marginal advertisement: how many extra sales a campaign produced beyond those that would have happened anyway. A practice can count what happened, in calls, bookings and new patients, without answering that question. It can also decline to report a return that its counts cannot support, and treat a vendor's dashboard as one more observational figure to be read with the studies above in mind.

The management literature gives a reason to count, with limits of its own. Brynjolfsson, Hitt and Kim used detailed survey data on 179 large publicly traded firms106. Firms adopting data-driven decision making had output and productivity 5% to 6% higher than would be expected given their other investments and information technology usage106. The result describes firms that adopted the practice without showing that adopting it caused the difference. We cite it for the direction of the argument, not for a size that would transfer to a five-person practice.

What a practice can count without statistics

A count is a record that something happened, kept the same way each time. The counts a practice can keep belong to the "reached" stage of the through-line in Chapter 1: calls received and calls answered, forms submitted, bookings made and bookings kept, and new patients or new matters opened. Each can be split by source. The split is the hard part, and Table 1 sets out the ways to get it and what each can and cannot show.

MethodWhat it recordsWhat it can bearWhat it cannot show
Asking at booking or intakeThe person's own answer to a fixed question, entered in the practice's systemWhich paths people name, including paths no tool sees, such as a friend's recommendationMemory is imperfect, and a person may name one source when several played a part
A tracking number for each sourceThe calls that reach each number the practice has assigned to a source, such as its website or a listingA count of calls by the number dialed, without asking the callerWhich source made the person decide: it credits the number dialed
A source field on forms and booking pagesThe page or link a form was submitted from, kept in the practice's own form systemWeb paths into a bookingCalls and walk-ins; tags that send visit data to third parties raise the privacy limits below
Platform-reported conversionsA count of conversions that an advertising platform credits to itselfConvenience; a view of what the platform claimsWhether the conversion would have happened anyway
A dated pause or changeCounts before, during and after a change the practice made on purposeThe nearest a small practice can come to a testSeason, staffing and other changes move counts too, and no one was randomized

Table 1. Ways to record where a new patient or matter came from, and the limits of each.

Asking is the simplest method and needs no tool. Its weakness is that people report what they remember. In an online survey of adult dental patients in private clinics in Bucharest, Romania, 99.1% said they learned about their dentist or dental office from family or friends, while 20.5% also searched online reviews92. In a 2012 national survey of the US public, 38% rated word of mouth from family and friends "very important" when selecting a primary care doctor, and 19% rated physician ratings on websites "very important"70. Both surveys count what people say, and neither counts what led to a given booking.

We infer from them that the paths to a practice overlap. A recommendation from a friend can be followed by a search to check it, and a review read on a phone can be followed by a call. A source question that allows one answer credits the friend or the search, and the other path leaves no trace. A question that allows a main answer and an "also" answer keeps more of the record, and we expect it adds little time at the desk. The list of choices should include "an AI assistant", because that path leaves no label of its own.

Call tracking does not depend on memory. In the sense used here, a call-tracking service gives the practice a separate telephone number for each source it wants to tell apart and logs the calls each number receives. Its limit is the limit of any count by channel: it credits the number dialed, not the influence that made the person pick up the phone. The phrase "call tracking" is searched about 1,600 times a month in the United States, at an average cost per click of $49.1727. A call-tracking service is also a vendor in the chain that Chapter 10 describes: in our check of 28 software tools, one call-tracking service is among the 15 that sign a business associate agreement only on certain plans54. Where a service records calls, Chapter 12 sets out the recording statutes that may apply.

A dated pause is the only method in Table 1 that comes near an experiment, and it is a weak one. Suppose a practice pays for ads on searches that name the practice, and pauses them for a period with the dates written down. The change in calls and bookings is some evidence about what those ads added, with the eBay result above as the reason to ask. It is not a controlled test, and the counts are small.

The small-numbers problem, with a worked example

Take a dental practice at the average that Henry Schein One† (software) reports for practices of one to seven locations, 30 new patients a month49. The page does not define a new patient. A month of 24 would be a fall of 6, which is one fifth (6 / 30 = 20%), and it reads like news. Whether it is news depends on how much a count of this size moves by itself from one month to the next, and on the simple model below it moves a good deal.

A count of events that arrive independently, at a steady average rate, wobbles from period to period by about the square root of its size. That is a property of counts themselves. The square root of 30 is about 5.5, so a month would depart from 30 by about 5.5 as a matter of course. Call that one step. A month within one step of the average runs from about 24.5 to 35.5 with no change in marketing at all, and a month of 24 sits at the edge of that range. The fall of 6 is about one step, the size of departure that chance alone produces in an ordinary month.

A physician's office meets the same arithmetic. In CDC data, new patients were 12.5% of office visits to primary care physicians in 201950. In Epic data from health systems, primary care physicians had a mean of 51.1 appointments a week in the first months of 2021195. That is 221 scheduled appointments a month (51.1 x 52 / 12). Of these, 171 to 180 are kept once the no-show rates of 18.8% across ten VA clinics and about 23% in a review of 105 studies are applied142,162. At 12.5% of kept visits, that is ~21 to 22 new-patient visits a month for one physician (171 to 180 x 12.5%; the sources differ in year and setting), with a wobble of about 4.7, the square root of 22.

The model's assumptions are strong. A closed week, an insurance change or a holiday makes real bookings lumpier than steady chance, and a cap on capacity can make a count steadier than chance would. We use the square root as a rough guide to the size of the wobble, not as a measurement of it. We also offer no threshold at which a change becomes real, because any single number would claim more than a practice's records could support. That judgment belongs to a statistician working with the practice's own history.

The wobble shrinks in relative terms as the count grows. The relative wobble is the square root of the count divided by the count, which equals 1 divided by the square root of the count. For a month of 30 that is 1 / 5.5, or 18%. A quarter at that average is 90 patients: the square root is about 9.5, so the relative wobble works out to about 10.5%. A year is 360: the square root is about 19, so the relative wobble comes to about 5.3%. Each time the count is four times as large, the relative wobble halves, so a count of 120 has a relative wobble that equals about 9%. This is why the paper reads trends over quarters and years, and asks of a monthly count only that it be kept the same way each time (Chapters 6 and 13).

A ratio of small counts is less steady than either count. The share of a practice's new patients that one channel brings has not been measured from practice records for any profession; the sources are surveys of opinion and stated behavior. The nearest figures are in MouthWatch†'s 2026 survey of 448 dental practices, where 15.6% named Google Business Profile and 7.6% named Google Ads as their top source of new patients93. Those are shares of practices, not of patients, so the range that follows is ours.

For a range, suppose the channel brings one new patient in ten of the 30, which is 3 a month. The square root of 3 is about 1.7, so a month of 2 and a month of 4 both sit within one step. If the channel is our lowest plan at $999 a month56, the cost of each recorded patient is $499.50 in the month of 2 ($999 / 2) and $249.75 in the month of 4 ($999 / 4). Two more patients halve the cost, which chance alone can produce, and a figure that moves by half on a difference of two patients cannot bear a decision about the channel.

None of the habits that follow needs a formula. Pool across time, so that a decision rests on a quarter or a year and not on a month. Pool across channels where the decision allows it, since the question "is marketing bringing patients" rests on larger counts than the question "which of four channels is best". And fix an expectation in advance and in writing, with a date, so that a result is compared with what the practice said it expected and not with what it later wishes it had said. The first two raise the size of the count, and the third protects the reading from hindsight.

Customer lifetime value and retention in a professional practice

Customer lifetime value is the value of a patient or client over the whole relationship, including everything after the first visit. Gupta, Lehmann and Stuart define the value of a customer as the expected sum of discounted future earnings109. The frame suits a professional practice, because a plan that sets what a new patient or client costs to win against what the first visit or matter pays leaves out everything after it. A patient who returns for years, or a client who refers others, is worth more than the first fee. The phrase "customer lifetime value" is searched about 3,600 times a month in the United States, at $12.98 a click27. As Chapter 5 cautioned, a cost per click is what advertisers pay, not what a client is worth.

The research on retention comes from outside the professions. Gupta, Lehmann and Stuart valued five firms from public data and found that a 1% improvement in retention, margin or acquisition cost improves firm value by 5%, 1% and 0.1% respectively109. The 5% depends on the assumptions of a model fitted to those five firms. Elsewhere they put the effect of a 1% improvement in retention at 3% to 7% of customer and firm value, against a margin elasticity of about 1 and an acquisition elasticity of only 0.02 to 0.3109. They also find that a 1% improvement in retention has almost five times the impact on firm value of a 1% change in the discount rate or cost of capital109.

Reichheld and Sasser's 1990 article (secondary) is the older statement. A summary of it republished by Bain & Company says that companies can boost profits by almost 100% by retaining just 5% more of their customers317. A quotation from the article gives its service examples: reducing the defection rate by just 5% yields 85% more profits in one bank's branch system, 50% more in an insurance brokerage and 30% more in an auto-service chain110. None of the three is a professional practice.

What transfers is the arithmetic of retention. If a practice loses the same share of its patients each year, the expected stay is 1 divided by that share. In the 2024 benchmark survey of client advisory services by CPA.com and the AICPA, the median client turnover was 10% a year113, which gives an expected stay of 10 years (1 / 10%). Suppose a firm cut its turnover to 5%: the expected stay would be 20 years. A change of 5 percentage points in retention doubles the expected stay, which is why recall and reminder work counts as marketing and has a measurable output. The example ignores discounting, which Gupta and his co-authors include, and assumes a constant rate, which patients who move or age out of a practice do not follow. It shows the shape of the effect and says nothing about any practice's actual rate.

Table 2 puts what one client is worth in numbers for each profession from cited figures. The last two columns multiply by a tenure only where one is cited. The values are our arithmetic on other people's figures, and none is a published lifetime value for a practice.

ClientFirst-year valueBasisOver the relationshipBasis
Primary care patient$3001.61 visits a year, which is 3.21 office visits per person194 times the 50.3% made to primary care194, at a mean payment of $186 a visit61$2,4008 years, the median stay of patients aged 65 and over of 23 physicians in one state112
Dental patient$887Spending per person with a dental visit, 2021118Not computedNo tenure is cited
DUI defense matter$3,500Average attorney cost, Washington, 2020242Not computedOne matter
Auto injury case$6,667One third of an average payment of more than $20,000243,244Not computedOne matter
Advisory client at a CPA firm that sells advisory services$17,867Median revenue per client, 2024, from a benchmark of 206 self-selected practices113$178,67010 years at 10% turnover113
Bookkeeping client$3,000Low end of $250 to $499 a month, the most common band among the customers of Ignition† (accounting-firm software)289Not computedNo tenure is cited

Table 2. What one client is worth in each profession, from cited figures (calc. where no source prints the value).

The rows are not like for like. The dental figure is spending by all payers per person with any dental visit, reported by households, not what a practice collects. The primary care row multiplies a visit count by a payment from two surveys that define primary care differently, so the $300 (calc.) is a rough value.

The injury row is a third of what insurers paid on average, not what a lawyer collects, and the one-third is a California advisory's description of a common contingency fee. The advisory row describes practices that chose to answer a survey about a service they sell well, so it is the high end for a small firm that does mostly tax returns, not a figure for all CPA firms or tax-only clients. The bookkeeping row includes solo bookkeepers. Where a practice knows its own figure it should use it. That is why the later chapters set the cost of a purchase against the number of clients needed to cover it, and do not assert what a client is worth.

We think retention in dentistry and medicine is mostly the recall visit: the patients due for a visit and the share who are scheduled, from the practice's own records (Chapters 6 and 7 treat reminders and recall messages). In accounting it is the client who returns for the next year's engagement (Chapter 9). In an Accounting Today survey run with ADP, firms reported that 56% of their clients had been with them for eight years or more, a self-report whose sample the page does not describe51.

In law, much of a client's later value may lie in referrals and not in repeat matters. We think so, but the referral or repeat share of a law firm's clients is NF; the nearest evidence is two surveys of stated behavior. In FindLaw†'s 2024 survey of 2,000 adults with a legal need, 48% of the respondents who learned about the attorney they contacted said they used referrals from family, friends or coworkers87. The Harris poll below puts the share who would ask a friend, family member or colleague at 46%90. We estimate the referral side at about half (from the two surveys; likely high as a share of clients). A person can use a referral and another source, and the surveys count what people say. The share of repeat matters is NF, and nothing near it is published.

The Harris poll was a 2010 survey for an American Bar Association committee, as the ABA Journal reported it, and its 46% is a statement of intention and not a count of referrals90. We think a law firm's source question should include a choice such as "referred by a past client", and that count is the nearest thing the firm has to a retention measure.

Measuring the AI channel: a monthly, dated check

The AI channel is a source that no tracking number or form field will show a practice, because a person who reads an assistant's answer and then calls does not arrive with a label. What a practice can observe directly is what each assistant answers. The minimum useful check, which Chapter 4 promised, applies the design of our own study to one practice: dated, covering all three assistants, using many questions, and repeated on a fixed day each month. A monthly check of this kind is part of our service. We have run the study's design once, on 2026-11-101. This paper reports no monthly series, so what follows is a method and its limits.

The check measures the practice's visibility to assistants. How to credit an AI agent that answers the practice's own calls for the bookings it makes is a separate question, which the companion paper takes up.

The questions, the assistants and the date

Each question in our study had one form, "Who is a good dentist in Atlanta, GA? Name specific businesses.", with the city and the kind of business changed1. The wording asks for names and sets no standard of quality. For a practice, the baseline set uses the same form for its kind of business in its own city and in the nearby towns it serves. A second set, kept apart, carries the specific questions people ask: a service, an insurance plan, an emergency or after-hours need, a language. The wording is fixed in the first month and changed only with a dated note, because a changed wording is a changed experiment; the two sets are never pooled. We think about 10 to 20 questions are enough that one odd answer does not decide a month.

The check uses the same three assistants as the study. In our study each was one model, reached through its programming interface with web search on: gpt-6-luna for ChatGPT, gemini-3.8-flash for Gemini and sonar for Perplexity1. Answers collected that way are a baseline and not what every signed-in user sees1. A practice that runs the check by hand in the consumer apps may see something different. It should record the product, the model name where one is shown, whether it was signed in, and the location setting, choose one way and keep it. Results for these models do not extend to later versions.

The study describes one day, and we expect assistants to change their answers over time. Running the check on the same day each month, with the time noted, lets the practice line an answer up against its own dated changes, such as a corrected listing or a new page.

What to record

Table 3 lists the fields of the record, which the practice keeps as a table with one row for each answer.

FieldWhat to enterWhy
Date and timeWhen the question was putAnswers date, and a record is useful only if its dates are fixed
Assistant and settingProduct, model name if shown, signed in or out, location, web search on or offSettings can change answers; they are held constant
Question number and exact wordingThe text sent, from the fixed listA changed wording is a changed experiment
The full answerSaved as text or a screenshotA summary loses what the answer says about the practice
Names listed, in orderEvery business named, in order, and how many were listedOur study read up to eight names per answer and compared the first-named1; a practice can read the whole list
Practice namedYes or no, and its place in the listThe main count
What the answer says about the practiceAddress, telephone, hours, services, plans accepted, new-patient statusCompared with the practice's own claims register; an error is a fact to correct
Sites citedEvery site the answer citesOur study's 25 most-cited sites drew 29% of citations and were directories, review platforms, list sites, publishers and a forum1
Other practices namedTheir namesContext for the practice's place in a list

Table 3. The monthly record of what each assistant answers.

Names are matched across assistants by spelling in our study, which can split one business into two when the assistants write its name differently1. A practice looking for itself can match by name, address and telephone number, as the stronger study in Chapter 4 proposed. We think the record of cited sites is the most useful field for action: the sites cited most in our study were directories, review platforms and similar sites, not any one practice's own pages1, so those are the places to check that the practice's facts are right and consistent.

How to read a change

The first rule is that one run is one draw. Our study did not measure how far an assistant disagrees with itself when asked the same question twice1, so a practice does not yet know the size of its own run-to-run variation.

A repeat-asking study of AI recommendations for local businesses or professional practices is NF; the nearest is SparkToro† and Gumshoe.ai†'s 2026 study, in which 600 volunteers ran 12 prompts through each of three tools a combined 2,961 times, and which reports less than a 1 in 100 chance that ChatGPT or Google's AI, asked a prompt 100 times, would give the same list of brands in any two responses115. It is not peer reviewed, its author discloses a possible conflict of interest, its prompts were not about dentists, doctors, lawyers or accountants, and volunteers pasted the answers by hand. It is a reason to expect variation and not a measure of it for a practice. We suggest a way to learn it: in the first month, run the full set on three separate days. If the practice is named on one day and not on another, for the same question and assistant, that difference is the noise to expect from then on. Chapter 14 proposes the same measurement at scale.

The second rule is to read counts and not single answers, and to read them by assistant. In our study 79.7% of the businesses named were named by only one of the three assistants, and any two assistants overlapped on between 11.9% and 14.3% of the businesses they named1. A pooled figure would hide an assistant that has stopped naming the practice behind two that still do. The count is therefore written for each assistant, in the form "named in a number of the 10 answers from ChatGPT", and a pooled count is kept beside it only as a convenience.

The third rule is the small-numbers arithmetic again. Suppose a practice puts 10 questions to each of three assistants, which is 30 answers, and is named in 8 of them in one month and 5 in the next. The fall of 3 is about one step of the wobble that a count of that size shows by itself (the square root of 6.5 is about 2.5). The answers are not independent events, since the same question goes to all three assistants, so the real wobble may be larger. We treat the monthly figure as a log, not a test.

The fourth rule is that a change is worth acting on only if it persists over several months, shows in more than one assistant or set of questions, and follows a dated change the practice made. Even then the reading is limited. Our study did not show that a cited page caused a name to appear, so it does not support crediting a change in the answers to a change in the practice's pages.

The fifth rule treats errors as a different kind of finding. A wrong telephone number, hour or insurance statement in an answer is a fact the practice can correct at its source, however often it appears. Our study does not show whether any business named exists, is open or accepts new patients, but a practice can check these things about itself, so the monthly check is partly a fact-check. The remedy is the one for any wrong listing: know about the error and correct the source.

Last, a result from the check is not evidence about quality. Being named does not measure the care or advice a practice gives, and the study does not show that any assistant is wrong or that a named business is good. The check also cannot show whether any answer led to a call or a booking. The AI referral share or conversion for a dental, medical, legal or accounting practice's site, and the calls or bookings that follow an assistant's answer, are NF. The nearest evidence is the share of traffic reported by Conductor† (SEO platform) and Microsoft Clarity† (free analytics tool), from which we estimate below.

Conductor† reports that AI referral traffic made up 1.08% of all website traffic across the 10 industries it analyzed, about 1 in 100 visits, with 0.64% for health care and 0.48% for financials42,43,44. Microsoft Clarity† reports that AI referrals were still under 1% of overall traffic on more than 1,200 publisher and news sites45.

A share of traffic leaves out an answer that nobody clicks. OpenAI's help page says that for a local business a mention counts when ChatGPT shows a result for that location, even if the customer does not open the listing or visit the website318. Pew Research Center's analysis of US adults' browsing found that users who saw a Google AI summary clicked a link inside it in 1% of visits to pages with such a summary30, and it compared clicks on traditional search results with and without a summary30. The traffic and click figures count clicks, not calls or bookings.

We estimate that AI referrals are under 1% of the visits to a dental, medical, legal or accounting practice's website (from Conductor†'s and Microsoft Clarity†'s traffic shares above42,43,44,45; likely low as a measure of influence). An unclicked answer and a later phone call leave no referral.

A conversion rate for AI referrals to a practice's site is NF; the nearest figures are a Clarity† finding of sign-up conversion of 1.66% for LLM traffic against 0.15% for search on publisher sites45 and Adobe's report that AI-referred traffic to US retail sites converted 90% better than other traffic in September 202653, and neither is a practice, a call or a booking. The source question, with "an AI assistant" among its choices, is the practice's only way to see even a self-reported trace of a call or a booking. A practice should not describe a check result in its advertising as a statement of its quality. Chapters 6 to 9 give each profession's rules on that.

Measurement touches personal information at three places: the source question, the call log and the tags on a website.

In health care, Chapter 7 sets out HHS's bulletin on online tracking and the 2024 court decision that vacated a portion of it12. HHS gives the example of an appointment made through a clinic's website with third-party tracking technologies that send the appointment and the IP address to a vendor: the vendor is a business associate, and an agreement is required12. Google Analytics and the Meta Pixel are two of the four tools in our check that offer no such agreement, according to the vendors' own pages54. Washington's My Health My Data Act bars sharing consumer health data without consent that is separate and distinct from the consent to collect, unless sharing is necessary to provide what the consumer requested179.

In tax practice, Chapter 9 sets out section 7216. Unless the statute or the regulation on permissible disclosures authorizes it, a tax return preparer may not disclose or use a taxpayer's tax return information before obtaining the taxpayer's written consent273. The sources cited here do not say whether the events a tag sends are such a disclosure, and we do not claim they are; that is for counsel. In law, Chapter 8 sets out ABA Model Rule 1.18, under which a lawyer who has learned information from a prospective client may not use or reveal it even when no relationship ensues, subject to the rule's exceptions219, and Rule 1.6(c), which requires reasonable efforts to prevent unauthorized disclosure of information relating to the representation222.

The design in Chapter 10 follows from these texts, but it is our choice and not a statement of what any rule requires. It sorts pages into public information pages, action pages and authenticated pages, and lets no advertising-platform tag run where a person states a condition, a matter or a return detail. Measurement then comes from counting events inside systems the practice has an agreement with, such as booking confirmations, call logs and form submissions, and from asking each new patient or client how they found the practice. The source question records a category in the practice's own system, and no health, tax or legal detail leaves it. The monthly check of the assistants uses public questions of the kind a patient or client would ask and records no personal information.

A practice that keeps tags off its action pages gives up platform-reported conversions. What those numbers were worth is itself in doubt, since observational methods in the Facebook experiments often failed to reproduce the experimental effects104, and the counts that remain are the ones this chapter has argued a small practice can read.

A minimum measurement plan

The plan below is the least a practice could run and still learn something from it. It uses no statistics beyond the arithmetic above, needs no health, tax or legal detail to leave the practice, and takes the shape of the through-line: reached, within the rules, measured. It is a design, not a promise of results.

  1. Name the one outcome that matters, which is new patients or matters booked and kept, and count it every month from the practice's own scheduling or matter system.
  2. Add one source question to booking or intake, with a fixed list of choices that includes a friend or family member, a search, a review site, another professional, a past client where the practice is a law firm, and an AI assistant. Record a main answer and an "also" answer.
  3. Count the front desk. Calls received, calls answered, calls missed, and the time to the first reply to a form or message are the counts Chapter 12 uses. Cancellations and no-shows belong here too.
  4. Keep every tool that sees personal information inside the vendor ledger of Chapter 10, and keep advertising-platform tags off action pages.
  5. Run the assistant check of Table 3 on the same day each month, with the first month run on three separate days to learn the noise.
  6. Record retention in the practice's own terms: patients due for recall and the share scheduled, clients retained from one year to the next, or referrals from past clients.
  7. Keep a dated change log of every new page, listing correction, tool, plan change and campaign, so that a change in a count can be set beside a change in the practice.
  8. Review quarterly. Show each count with the number of events behind it, read it against the quarter before and the same quarter a year before, and write down what would change the practice's mind.

The plan reports no return on investment for a channel, because the counts cannot support one and the studies above explain why. It also credits no source with a patient who named two, does not measure an AI agent's own contribution to bookings, and asks for no statistical training.

A practice that runs it for a year will hold a small record: how many people booked, how they said they came, how many calls went unanswered, what the assistants said about the practice month by month, and what the practice changed. That record cannot prove that any one thing worked, but it shows the practice a missed call and a wrong fact in time to act on them. Chapter 12 turns to the third item on the list, the front desk.

Chapter 12

The front desk: response, access and the AI receptionist

First contact is the conversion point: speed to lead and speed of access

First contact is the moment when a person who has found and chosen a practice either reaches someone or does not. It is the "reached" stage of the through-line in Chapter 1, where the website's work as a rep for the front office is completed or lost. Three terms are used in a fixed sense. Response time is the time from a person's contact to the first reply, from a human or a machine. Answer rate is the share of contacts that are answered at all. Wait time is the time from a request to the first available appointment. The studies below measure different ones of these, and Table 1 keeps them apart. The telephone figures in the later tables divide further. "Hung up before reaching an agent", "abandoned", "went to voicemail" and "not answered by a person" count different events, and no sentence here treats them as one rate or averages them.

The economics of information gives a reason to care about response time. In Stigler's account, the optimum amount of search is found when the cost of search is equated to its expected marginal return58. Applied here, we think a call that rings out, or an inquiry that waits a day, makes the next practice cheaper to try. Services research names the same quality. The 1986 working paper that introduced SERVQUAL measures five dimensions of service quality: tangibles, reliability, responsiveness, assurance and empathy73. A 2013 review of the Gaps Model (secondary) describes responsiveness as the firm's willingness to help customers and perform the service promptly72. For a buyer who cannot judge the work itself, how quickly someone replies is one of the few qualities visible before the purchase.

How fast a first response arrives

Oldroyd, McElheran and Elkington wrote in Harvard Business Review in 2011 that their research showed most companies were not responding nearly fast enough to potential customers' online queries319. We have not read the figures that follow in the journal; they are taken from a 2011 news piece that quotes the article. The authors audited 2,241 US companies, measuring how long each took to respond to a web-generated test lead217. In that audit, 37% responded within an hour, 16% within one to 24 hours, 24% took more than 24 hours, and 23% never responded at all217. Among companies that responded within 30 days, the average response time was 42 hours217.

The same article describes a separate study of 1.25 million sales leads received by 29 B2C and 13 B2B companies in the US217. Firms that tried to contact potential customers within an hour of receiving a query were nearly seven times as likely to qualify the lead as those that tried even an hour later217. Qualifying meant having a meaningful conversation with a key decision maker217. The same passage reports that they were more than 60 times as likely as companies that waited 24 hours or longer217.

The audited companies were US companies in general, and the passages cited do not say that any is a dental, medical, legal or accounting practice. The comparison is between firms that tried at different speeds and does not show what speed would change for a given firm, because firms that reply fast may differ in other ways. A multiple of 60 describes one set of leads and one definition of qualifying. We use the article only as evidence that response times varied widely across the audited firms and that, in one set of leads, the time to the first attempt went with whether the lead reached a decision maker. We do not carry its multiples to a practice.

What the audits measured

The other evidence on first contact comes from audits in which callers phone offices and note what happens. Table 1 lists them with what each measured. The telephone audits of physician and dental offices measure appointment access, by insurance status or by wait, and not answer rates. The audits that count answered calls cover law firms, dental offices in four countries and mental health practices, each with a different instrument.

StudyWho was contactedWhat was measuredResultWhat it does not show
Oldroyd, McElheran and Elkington, 20112,241 US companies, web test leadResponse timeAs reported, 37% within an hour and 23% never217Any professional practice
AMN Healthcare survey, 20251,391 physician offices in 15 large metro areas, January 15 to February 24, 2025Wait time for a new-patient appointmentAverage 31 days across six specialties; family medicine 23.5 days, from 1 to 20722How often the phone was answered; metros only
Rhodes and co-authors (calls November 2012 to April 2013)Primary care practices in 10 states, simulated patients by telephoneWhether a caller was given a new-patient appointment84.7% of privately insured callers, 57.9% of Medicaid callers320An answer rate; the outcome is an appointment offered
LeHew and co-authors, 2023Dental offices eligible to bill Medicaid, children in IllinoisOdds of obtaining an appointmentPrivately insured patients had 5.9 times the odds of Medicaid patients321Dental offices in general, or adults
Bisgaier and co-authors, 2011273 specialty clinics, paired callersDenial of an appointment66% of Medicaid-CHIP callers denied, 11% of privately insured callers161Anything about adults or primary care
Clio† (legal software) secret-shopper study, 2019A random sample of 1,000 US law firms emailed, 500 of them also phonedWhether the call was answeredClio† reports that 56% answered, 39% went to voicemail and 5% were unanswered, and that 73% either picked up or phoned back212What the caller did next; any other profession
Clio† secret-shopper study, 2024500 law firms, a client inquiry by phone and emailAnswer rate for calls; response to emailClio† reports that 40% answered the call and 33% responded to the email23How the firms were drawn; any other profession
Hennessey Digital† (law-firm marketing), 20251,333 US law firms, contacted through their online lead formsTime to respond to a web leadMedian 13 minutes; 74% responded within seven days and 26% did not respond214Phone answering; the firms were drawn from marketing lists
Geier† (practice training), 2015Test calls to more than 12,600 dental practices in the United States, Canada, Australia and the United Kingdom over 12 monthsWhether a call in regular business hours went straight to voicemail35% of test calls, 4,415 calls, went straight to voicemail25A US rate; how the practices were chosen; a trade column and not a peer-reviewed study
Ohio State researchers (JAMA Health Forum), 20265,822 psychiatrists and psychologists listed in commercial insurance directories in five statesWhether a call reached a personMore than 6 in 10 calls reached an automated system or answering machine and could not be analyzed; a call that was picked up led to an appointment 24% of the time, and 31% if the caller would see another clinician216Physician or dental offices; the numbers came from insurance directories

Table 1. What each audit of first contact measured.

Each of the three insurance audits counts something a front desk does: telling a caller whether the practice will see them. In the Rhodes audit, among callers who obtained an appointment, the median wait was typically less than 1 week (2 weeks in Massachusetts), with no differences by insurance status or urgency of health concern320. Uninsured callers got appointments 78.8% of the time when full cash payment was possible, but only 15.4% of the time if payment required at the visit was restricted to $75 or less320. What a front desk says about insurance, cash payment and the next open appointment decides the outcome these audits count, so an automated front desk must say what the practice would say and must not guess.

The AMN survey's 31 days is the average wait for a new-patient appointment, and we think it measures capacity. A fast reply to a person who then waits a month solves a different problem from the one a missed call creates.

A long wait means a full schedule, and a full schedule changes the form of what more demand is worth without making it worthless. A full schedule still has gaps, because patients miss visits. The mean no-show rate was 18.8% across ten main clinics of one VA medical center162. A review of 105 studies found the average no-show rate in healthcare appointment scheduling to be of the order of 23%142. More inquiries let a practice fill those gaps, choose among patients, take better-paying work, raise its minimum fee or justify another provider. Chapter 7 treats waits and access for a clinic. Dentistry shows the other side. Nearly two in five dentists told the ADA Health Policy Institute in the third quarter of 2026 that they were not busy enough and could have treated more patients102, so for them the schedule is not the limit.

The law firm study and what it adds

In its 2024 study Clio† reported that 33% of the 500 firms responded to an emailed inquiry, against 40% in 201923, and that 40% answered the phone call, against 56% in 201923. In total, 52% either answered or called back, "meaning 48% were essentially unreachable by phone" in Clio†'s words23. Assume the 40% who answered are among the 52%. Then 12% called back without having answered (52% - 40%), which is 12 of every 60 firms that did not answer, or 20%. On this reading, four in five of the firms that did not answer the call did not call back.

Among firms that replied to emails, Clio† reported, 84% did so within eight hours, but just 18% provided clear next steps or cost information211. It also reported that 73% of the shoppers were unlikely to recommend the firms they contacted211. The study's cautions (Chapter 8) are that it made one call to each firm, the source does not state how the 500 firms were drawn, and the shoppers posed as clients. It shows what one set of shoppers met, not that any caller went elsewhere. The 2019 test used a different design, a random sample of 1,000 firms emailed and 500 of them phoned212, so the fall from 56% to 40%23 compares two studies and not one panel of firms.

Hennessey Digital† also tested web forms. It contacted 1,333 US law firms through their online lead forms in 2025 and reports a median response time of 13 minutes, steady year over year after peaks of 25 and 33 minutes214. It also reports that 74% of firms responded within seven days, up from 59% in 2021, and 26% did not respond to online leads214. The firms were drawn from marketing lists. Clio†'s 2019 consumer survey of 2,000 US adults asked how fast people expect a reply to a message or email. Of those, 10% expected one within an hour, 24% within a few hours and 45% within 24 hours, so a reply later than 24 hours misses the expectation of 79%212. That is a stated expectation, not behavior.

Call volume, answer rates and hang-ups: what each owner measured

Most figures for the telephone come from vendors, and most of the vendors sell a fix for what they measure. Table 2 therefore gives each figure under its owner.

Owner and yearWhose callsWhat it countsResult
Patient Prism† (call analytics); calls in 20258,280 dental practice and DSO locations; 11,552,668 patient callsCallers who hung up before reaching an agentOf every 100 patient calls, 31 hung up and 69 connected with the front desk24
Patient Prism†, same callsSame locationsIts own classification of each callOf every 100 calls, 34 were real booking opportunities and 21 ended in a booked appointment24
PatientNews† (dental marketing); undatedClient dental practices; over two million new patient calls scoredCalls answered; calls converted to appointmentsThe average client answers 75% of calls and converts 53%; the top practices answer 94% and convert 85%140
Invoca† (call tracking); 2026 reportCustomers in six health care fields: behavioral health and rehab, dental, eye care, hearing, hospitals and physical therapyInbound calls answered by a person54% across the six fields, and 45% to 68% between them26
Call Box† (call tracking and answering); 2020More than 1,700 US health care practicesPatient callers connected to a qualified individual who could help73% in 2019 and 63% in 2020198
CallMyDoc† (answering); 20254,687,808 inbound calls to 297 practices on its own platformCalls abandoned before the caller selected a call type33.1%160
Veterans Health Administration (Griffith and co-authors, peer reviewed); October 2014 and September 2016Medical-center call centers of a government health systemCalls ended by the caller before a staff member answeredMean 12.0% in 2014 and 8.3% in 2016; mean speed of answer 87.1 seconds in 2014 and 69 seconds in 2016159

Table 2. Telephone measures in health care, by owner.

Calls per office are measured in three ways that do not match. Dividing Patient Prism†'s 11,552,668 calls by 8,280 locations gives about 1,400 patient calls per location per year, or 116 a month (11,552,668 / 8,280 / 12). That assumes every location was on the platform all year and counts only calls the vendor classed as patient calls, so it is a lower bound, not a measured office average24. Call Box† reports that its practices received an average of 299 calls per office location in February 2020 and 505 in June 2020, a pandemic year, with the mix of dental and medical offices not stated198.

A peer-reviewed study of four clinics of one university health system found incoming call rates from 439 to 1,002 per 1,000 patients between the start and end of the study period322. Spruce Health† tabulated seven primary care practices at about 21 incoming patient calls a day for every thousand panel patients323. From an assumed panel of 1,500 it estimated 53 calls a day for a provider323. CallMyDoc† reports that 89.4% of the patient calls on its network arrived in standard business hours, Monday to Friday, 8 am to 5 pm local time160.

What the callers did next is thinner. Patient Prism† reports that, of the patients who walked away without booking, the average share who then received a follow-up was 38%, against more than 80% for the best operators in its cohort24. It counts 890,621 patients who walked away in 2025 and never got a callback24, a count and not a rate. In the Ohio State audit of Table 1, a call that reached someone led to an appointment 24% of the time. The VA study cuts the other way on what a missed call does to patients. A center's average speed of answer was associated with how patients rated their ability to get timely urgent-care appointments159. Its abandonment rate was not associated with any of the satisfaction outcomes the authors tested159. That is an observational association in a government system. It does not show that missed calls do not matter, and it is not evidence about new-patient acquisition.

Three quantities are NF; the nearest evidence, or another bound, follows for each. The first is the share of calls an accounting firm misses, which for a small firm with no one dedicated to the telephone is ~3 to 6 calls in 10 (from the two ends below; a loose range). The low end is Patient Prism†'s 31 callers in 100 who hung up before reaching an agent at dental offices24. The high end is the 60% of law firms that did not answer Clio†'s 2024 test call (100% - 40%)23. The two ends count different things, a share of calls and a share of firms on one test call, and a firm with a full-time receptionist will sit below it.

The second is calls a month at a law firm, for which we suppose no value. New matters can be bounded another way, from cited hours and fees, and Table 12 does so. The third is the after-hours share of calls. CallMyDoc†'s 89.4% in business hours implies 10.6% outside them (100% - 89.4%)160. No source for any of the four professions could be opened, so that figure is the nearest.

Reminders and no-shows: keeping the appointment once it is made

A reminder serves the same stage: once a contact has become an appointment, the work is to see that the person comes. A 2013 Cochrane review found moderate-quality evidence from seven trials (5,841 participants) that mobile text message reminders improved attendance at healthcare appointments compared with no reminders142. The risk ratio was 1.14 (95% confidence interval 1.03 to 1.26)142. Across the pooled trials, attendance was 67.8% with no reminders, 78.6% with mobile phone messaging reminders and 80.3% with phone call reminders142. The gains over no reminder are 10.8 and 12.5 percentage points (78.6% - 67.8% and 80.3% - 67.8%). The review found the cost per attendance of text reminders to be 55% and 65% lower than that of phone call reminders in the two studies that reported it142. Its authors concluded that the evidence remains insufficient to conclusively inform policy decisions142. A 2026 meta-analysis of 10 randomized trials (8,236 participants) of reminders for hospital outpatient appointments found a pooled risk ratio for attendance of 1.11, an 11% increase in attendance compared with no reminder165.

The 2018 review of 105 studies found the highest average no-show rate in Africa (43.0%) and the lowest in Oceania (13.2%)142. A randomized quality-improvement initiative at a safety-net primary care clinic added model-driven live reminder calls to standard automated reminders for appointments with a predicted no-show risk of 15% or higher. The no-show rate fell to 33%, against 36% in the standard-care group163, a gain of 3 percentage points. The other estimate of what adding a phone reminder does is smaller. In the same VA study as the 18.8% figure, a centralized phone reminder was followed by a fall in the no-show rate from 16.3% to 15.8%162, a gain of 0.5 points in a before-and-after comparison.

In a poll of 265 medical practices in August 2025, MGMA found that 27% said no-show rates had increased in 2025, 60% that they had stayed the same and 13% that they had decreased166. For dental offices, Henry Schein One† (practice software) reports an average no-show rate of 4% for practices of one to seven locations49 and 3% for dental groups144, with no sample or definition printed.

At University of Michigan dental school clinics, 94.9% of patients who were emailed a reminder and 92.2% of those who were texted attended their appointment142. A peer-reviewed or ADA overall no-show rate for US dental practices is NF; the nearest evidence is these neighbors, which give ~3% to 8% (from the two ends below; likely low). The low end is Henry Schein One†'s 3% for dental groups144, and the high end is the dental school's 7.8% missed among patients who were texted (100% - 92.2%)142. It errs low for an average private practice because Henry Schein One†'s customers are practices that buy benchmarking software and the dental school's patients were reminded.

None of this is specific to a dental or legal practice, and none of it tested an AI voice agent: the Cochrane review was published in 2013, and the clinic study used live calls chosen by a risk model. We think the three together say that missed appointments are a large and measurable loss in health care, and that a practice should measure its own. A reminder placed by an AI voice is a call the practice initiates, which differs legally from an AI that answers.

Answering service and AI receptionist: the search data

Table 3 lists the front-desk phrases we looked up in Google Ads keyword data for November 2026. As Chapter 5 explains, an owner or a vendor would type these, not a patient or client; search volume shows interest and not purchase, and cost per click is what advertisers pay and not what a client is worth.

PhraseSearches a month, USCost per clickSource
ai receptionist74,000$43.6627
ai voice agent33,100$49.0727
ai answering service4,400$58.4827
medical answering service5,400$35.9227
legal answering service720$214.5127
dental answering service110$181.6827
ai medical receptionist140$80.7627
ai dental receptionist170$152.0827
legal intake software140$106.4427
law firm intake70$92.4127
speed to lead880$18.5327
lead response time50$68.0227

Table 3. Front-desk search phrases, from our Google Ads keyword data for November 2026.

Adding a profession collapses the volume. The generic phrase "ai receptionist" is 435 times "ai dental receptionist" and 529 times "ai medical receptionist" (74,000 / 170 and 74,000 / 140). The answering-service phrases differ widely too: "medical answering service" is 49 times "dental answering service" and 7.5 times "legal answering service" (from Table 3). We looked up no answering-service phrase for accounting.

Dentistry does not follow the pattern that Chapter 5 found in law and accounting, where the leading AI phrase is searched far more than the leading marketing phrase. "Dental ai" averages 880 searches a month and "dental marketing" 1,30027, and the dental front-desk phrases are small, at 110 and 170 searches a month27. We take "speed to lead" and "lead response time" to be one idea under two names, the speed of the first response described above. The first is searched 17.6 times as often as the second (880 / 50).

What an AI receptionist is, and what it may and must not do

An AI receptionist is software that answers a practice's telephone in a synthesized voice, converses with the caller, and acts within a narrow set of permissions. We describe our own, Technical Marketing's AI voice agent, priced at $299 a month with 250 call minutes included, and $0.39 a minute after that56. It answers questions, books, qualifies and hands off to the office team. It tells every caller that it is an AI assistant, and the practice approves the script it uses.

A health care practice needs no program to buy a marketing plan, because a plan carries no patient information57. An agent that answers patient calls does touch it, so for dental and medical offices we sell the agent inside a HIPAA-ready program, which needs a consultation first57. The program is a compliance layer with its own price, and the agent is added to it57. Four front-desk add-ons run on the same layer: reminders and rebooking, recall and reactivation, waitlist backfill, and intake with insurance capture57. For a health care practice we frame the agent and the program as service (speed of answer and more hands at the front desk), not as a source of new patients, which are the product of the plan. "HIPAA-ready" is our phrase for what we build toward. Compliance depends on a signed business associate agreement and on how each tool is set up, and we apply the word "compliant" to no tool, our own included.

Every row of Table 4 follows one design rule: the assistant acts on the practice's own facts, inside permissions the practice sets, and a named person receives every hand-off. The limits in the "may do" and "must not" columns are ours, drawn for dental, medical and legal settings from the rules we publish for our service and for accounting from the intake design in Chapter 9. The last column names the texts that bear on them.

SettingMay doMust notTexts that bear on the limits
Dental officeSay first that it is an AI assistant; book cleanings, exams and consults into open times; answer hours, location, new-patient steps and plans accepted from the practice's own list; take a name, callback number and reason; ring a named on-call person for the practice's urgent listDiagnose or say what a symptom means; give aftercare, medication or sedation instructions; quote a treatment cost or promise what insurance will pay; put health information in a text message; run on a service with no signed business associate agreementHHS on business associates55; satisfactory assurances in a written contract148; the ADA code on advertising7
Medical clinicSay first that it is an AI assistant; offer visit times the calendar shows as open; take a name, callback number and short reason; answer office questions from the clinic's own facts; transfer an urgent call to a named on-call person, and tell an emergency caller to dial 911Diagnose, triage or comment on symptoms; give medication, dosing or test-result information; put health information in a text message; run on a service with no signed business associate agreementThe same business associate texts; California section 1339.75 on AI communications about patient clinical information149
Law firmSay first that it is an AI assistant and not a lawyer or a member of staff; collect only the intake fields the firm chose; offer consultation times from the attorney's calendar; answer what the website covers; transfer an urgent call to a named lawyer or staff memberGive legal advice or say whether a caller has a case; say the firm will take the matter; ask for more detail than the firm chose to collect; quote a fee or promise an outcome the firm has not published; confirm who the firm's clients are or discuss an existing matterRules 5.3, 1.18, 1.6(c) and 7.16,219,220,222; Opinion 512221; Florida Opinion 24-1225
Accounting firmSay first that it is an AI assistant; ask the kind of business, the services wanted, the size and the timing; offer a consultation time or a place on a waiting list; pass the answers to a named personTake Social Security numbers, income figures or documents; discuss any person's return or say whether someone is a client; give tax advice; disclose confidential client informationSection 7216 and its regulations269,271,273; the AICPA confidentiality rule8; the FTC Safeguards Rule275

Table 4. What a front-desk AI assistant may do and must not do, by setting, and the texts that bear on the limits.

What each limit rests on

For the dental and medical rows, the contract limit rests on the business associate rules. HHS describes a business associate as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information55. A subcontractor that does so for a business associate is itself one55. A covered entity may disclose protected health information to a business associate if it obtains satisfactory assurance that the associate will appropriately safeguard it, and the assurance must be documented in a written contract148. A caller who gives a name and a symptom may be handing over information of that kind, and whether the practice is a covered entity and the call involves protected health information is for its counsel.

An AI receptionist can sit on a telephone carrier and call a separate language-model provider, so the chain of agreements is as long as the path the call takes (Chapter 10). In our check of 28 tools, 15 sign a business associate agreement only on certain plans54. So the limit is that the assistant does not run on a service without a signed agreement.

The clinical limits rest on how far the disclosure statutes reach once an assistant says something about health. California Health and Safety Code section 1339.75 (AB 3030) covers a health facility, clinic, physician's office or group practice office that uses generative AI to generate written or verbal patient communications pertaining to patient clinical information. Those communications must include a disclaimer and instructions for reaching a human149. Patient clinical information does not include appointment scheduling, billing or other clerical or business matters149, and the duty does not apply to a communication that a licensed or certified health care provider reads and reviews149. On the text, for the facilities the statute names, an assistant that only schedules sits outside the duty, and one that talks about symptoms or results appears to sit inside it unless a provider reviews the message. A live call has no step at which a provider reads what the assistant says first, so in our design the assistant stays on the scheduling side of the line.

The text cited here names health facilities, clinics, physician's offices and group practice offices and does not mention dental offices, so whether it reaches one is not settled by the text. For the dental row no text cited here bars an assistant from answering a clinical question, so the limit is our design rule. We think judging how urgent a symptom is, or advising what to do, is clinical work that belongs to the clinicians.

For the dental price limit, the nearest text is section 5.F of the ADA code, which says no dentist shall advertise or solicit patients in any form of communication in a manner that is false or misleading in any material respect7. It is an ethics code and not a statute, and whether a price given on a call falls within it is for counsel. Our rule is simpler: the assistant does not quote a treatment cost or promise what insurance will pay.

For the law row, ABA Model Rule 5.3 requires a partner or lawyer with comparable managerial authority to make reasonable efforts to ensure the firm has measures giving reasonable assurance that a nonlawyer's conduct is compatible with the lawyer's obligations220. ABA Formal Opinion 512 says lawyers have additional supervisory obligations insofar as they rely on others outside the firm to employ generative AI tools221. Rule 1.18 treats a person who consults a lawyer about forming a relationship as a prospective client219. It bars the lawyer from using or revealing what was learned even when no relationship ensues, subject to the rule's exceptions219. That bears on the limit to the fields the firm chose.

Florida's advisory Opinion 24-1 warns that website chatbots used for intake present the risk that a prospective-client relationship, or even a lawyer-client relationship, has been created without the lawyer's knowledge225. It also says that generative AI chatbots that communicate with clients or third parties must comply with restrictions on lawyer advertising225. They must include a disclaimer that they are an AI program and not a lawyer or employee of the firm225. That bears on the rule that the assistant never says the firm will take a matter. Rule 7.1 bars a false or misleading communication about a lawyer's services6, which bears on the fee and outcome limit. Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of information relating to a representation222, which bears on the limit on discussing who the firm's clients are.

The FTC announced a complaint against DoNotPay in September 2024, with a settlement that would require a payment of $193,000238. The complaint alleged that the company did not test whether its chatbot's output was equal to the level of a human lawyer238. It is an allegation about claims made to consumers, and we cite it only as a reason this paper does not say that an assistant gives legal advice.

For the accounting row, the evidence is thinner. Section 7216 makes it a misdemeanor for a tax return preparer to knowingly or recklessly disclose or use information furnished for a return other than to prepare the return269. The regulations bar disclosure or use before the taxpayer's written consent unless the statute or the regulation on permissible disclosures authorizes it273. The regulations define a preparer to include a person engaged in providing auxiliary services in connection with preparing returns, including one who develops software used to prepare or file them271.

The AICPA rule bars disclosing confidential client information without the client's specific consent8. Its interpretation on service providers asks for a confidentiality contract or the client's specific consent before information goes to a third-party provider8. The FTC lists tax preparation firms among the financial institutions the Safeguards Rule covers275. Chapter 9 explains that the sources do not say whether what a prospect types or says to an intake system is information furnished for a return. We therefore design the assistant to collect none. The limit on tax advice is our design rule. The nearest text is the IRS Office of Professional Responsibility's June 2026 alert on a practitioner's own use of AI, which says that practitioners cannot rely solely on AI and that human scrutiny is essential286. That alert applies existing duties and is not a new rule.

The law of the call: recording, AI voices and disclosure

This section describes what the texts cited here say about a telephone call. It decides nothing. Chapter 10 compares the disclosure and recording texts across the four professions.

Calls the practice receives and calls it places

The Federal Communications Commission confirmed in 2024 that the Telephone Consumer Protection Act's restrictions on an "artificial or prerecorded voice" encompass current AI technologies that generate human voices186. Calls that use such technologies require the prior express consent of the called party to initiate them, absent an emergency purpose or exemption186. The ruling concerns calls initiated with an AI-generated voice, and that is the outbound case: a reminder, a recall call or a follow-up placed by a practice's voice agent.

On the inbound case, the FCC's September 2024 notice of proposed rulemaking on AI-generated calls said that the Act's prohibition on an artificial or prerecorded voice extends only to outbound calls made or initiated by the caller187. It added that "The TCPA's requirements do not extend to technologies used to answer inbound calls"187. That is a proposal and not a ruling. We do not say that the 2024 ruling reaches an AI agent answering a call that the patient or client placed.

The statute speaks of calls that are made or initiated. It makes it unlawful to make a call using an artificial or prerecorded voice, other than an emergency call or one made with the prior express consent of the called party, to the categories of lines the paragraph lists306. For residential lines it makes it unlawful "to initiate any telephone call" to deliver a message in such a voice without that consent306. This paper cites no source that applies either provision to a call the patient or client placed.

The statute also sets the stakes of the outbound case. A person may sue in state court to recover actual monetary loss or $500 in damages for each violation, whichever is greater306. A court that finds a willful or knowing violation may increase the award to not more than 3 times that amount306. This is why this paper treats an AI voice that places reminder calls as a separate decision from an AI that answers (Chapter 13).

Recording and transcripts

Federal law lets a private person intercept a communication where the person is a party to it or one of the parties has given prior consent, unless the purpose is criminal or tortious192. That rule is a floor, and state statutes can be stricter. Washington makes it unlawful to intercept or record a private communication transmitted by telephone or other device without first obtaining the consent of all the participants179. California Penal Code section 632 punishes a person who records a confidential communication without the consent of all parties, with a fine not exceeding $2,500 per violation149. Florida and Maryland make interception lawful where all the parties have given prior consent134,307, and Massachusetts defines interception as secretly hearing or recording a communication without the prior authority of all parties308.

Whether an AI service's transcript counts as a recording is a question for counsel. A voice agent that keeps audio or text of calls raises it, so a practice that uses one should know, in writing, whether recording is on, who can hear or read the result, and what the caller was told.

Telling the caller that it is an AI

No single text governs the disclosure, and the texts differ more in their triggers than in what they ask.

Utah's general duty applies to a supplier that uses generative AI to interact with an individual in connection with a consumer transaction183. It applies only if the individual asks or otherwise prompts the supplier about whether AI is being used, and the request must be clear and unambiguous183. In the text cited, Utah's separate duty for regulated occupations states no ask-first condition. An individual providing services in a regulated occupation must prominently disclose when an individual receiving services is interacting with generative AI in the provision of regulated services, if the use constitutes a high-risk interaction183. The statute defines that to include the collection of sensitive personal information such as health data or financial data184. It also includes personalized recommendations, advice or information that could reasonably be relied upon to make significant personal decisions, including financial, legal and medical advice or services184. The disclosure is given verbally at the start of a verbal interaction183. The page we read does not say which professions are regulated occupations184.

California's section 1339.75, described above, asks for the disclaimer verbally at the start and the end for audio communications149.

Texas's H.B. 149 took effect January 1, 2026150. Its general disclosure section is addressed to governmental agencies. A separate subsection requires the provider of a health care service or treatment in which an AI system is used to give the disclosure no later than the date the service or treatment is first provided, or as soon as reasonably possible in an emergency150.

California's bot law makes it unlawful to use a bot to communicate or interact with another person in California online, with the intent to mislead the person about its artificial identity in order to incentivize a purchase or sale or to influence a vote149. A person who discloses clearly and conspicuously that it is a bot is not liable under that section149. The source does not answer whether a telephone voice agent is covered. Colorado's reenacted act, effective January 1, 2027, requires a notice before a covered automated decision-making technology is used to materially influence a consequential decision303. That notice is about the use of the technology in a decision affecting the consumer. Florida's advisory Opinion 24-1 addresses lawyers225. The AMA's policy, which is not law, says AI use should be clearly disclosed at the beginning of an interaction where patients engage directly with AI185.

Our rule does not depend on which of these applies. Callers and site visitors are always told they are talking to an AI assistant, at the start. For audio contacts that touch clinical matters, the design in Chapter 10 repeats the statement at the end. This goes beyond what Utah's general duty requires.

Hand-off design: when a person takes over

Hand-off is the moment when the assistant's work ends and a person's begins, and it has three design parts: when it happens, to whom, and what the person receives. Chapter 10 states the principle as the fifth design requirement, which is to keep a named human responsible.

The triggers are of two kinds: the caller's own (a request for a person, repeated misunderstanding, distress) and a list the practice writes of what must reach a person at once. Our published examples are a knocked-out tooth or heavy bleeding after a procedure in a dental office, and an emergency or urgent clinical concern in a clinic, with a direction to dial 911 in an emergency. In a law firm they are a hearing, an arrest or a filing deadline. Each trigger has a named person. When no one answers, the assistant takes a message and sends it with the reason. A caller who asks for a person gets one, or a message that a person will see.

What the person receives matters as much as the transfer. A summary should carry the caller's name, callback number, reason and the urgency the assistant assigned, and the person should be able to see what the caller said and not only the assistant's account of it. NIST's generative AI profile lists Human-AI Configuration among its risks, covering automation bias and over-reliance on a generative AI system190. A summary the receiving person accepts without looking is that risk in small form. Where what the caller said is kept, the recording statutes above come into play.

The hand-off also moves the delay and does not remove it. Suppose an assistant answers at once and takes a message, and the practice returns it 42 hours later, which, as reported, is the average response time among the companies in the Oldroyd audit that responded within 30 days217. The assistant has then shortened only the ring. The hand-off therefore has its own count, the time from the message to the first human reply, which is a form of the first-reply time that Chapter 11 lists.

Callers vary in how they take to the machine, and the evidence is mixed. A natural field experiment at a large telecommunications company introduced a voice-based AI system to replace the interactive voice response system324. It temporarily increased machine-service duration and customers' demand for human service, but persistently reduced customer complaints324.

In a Gartner survey of 5,728 customers in December 2023, 64% said they would prefer that companies didn't use AI in their customer service325. In the same survey, 53% would consider switching to a competitor if they found out a company was going to use AI for customer service325. Neither the experiment nor the survey concerns a practice's patients or clients, and we use them for one point: some callers will want a person, and the path to one has to be short and always open. The Consumer Financial Protection Bureau's 2023 report on chatbots in consumer finance said that several customers had filed complaints about getting stuck in "doom loops" with chatbots, often because the issue fell outside the chatbot's limited capabilities326. The passage cited gives no rate.

Entrepreneur reported that Klarna's chief executive told Bloomberg in May 2025 that the company was hiring human customer service workers again, saying it was critical to be clear to customers that there will always be a human if they want one327. That is a news report of an interview, and it states a design principle this section reached by argument.

What patients and practices say about AI in health care

The surveys that touch AI in health care ask about different things, and none asks patients about an AI receptionist. A Pew Research Center survey of 11,004 US adults in December 2022 found that 60% would be uncomfortable with their provider relying on AI in their own health care and that 39% would be comfortable328. In the same survey 57% said that AI used to diagnose disease and recommend treatments would make the patient-provider relationship worse, and 13% said better328. Both questions concern clinical AI, not AI that answers the phone or schedules. A Talker Research survey of 2,000 US adults who had visited a doctor in the past year, commissioned by PatientPoint, found that 53% believed it would become standard practice in the near future to schedule appointments without speaking to a human329. That is an expectation about the future, not a preference.

Practice leaders answer different questions. In an MGMA poll of April 2025 with 375 applicable responses, about one in five medical group practices (19%) used some version of a chatbot or virtual assistant for patient communication188. In a February 2026 poll with 177 applicable responses, leaders named scheduling (31%) and calls (27%) as the areas where they were focusing AI and automation to improve access189. In a March 2026 poll with 294 applicable responses on the time-intensive tasks staff handle on the phones, eligibility and prior authorization came first at 45% of responses and scheduling second at 31%197. The MGMA polls are of self-selected members. The AMA's 2026 survey found 81% of 1,692 physicians reporting awareness or use of AI in their practice, up from 66% in 2024330, and that is awareness or use, with no front-desk category in the uses it lists.

An independent survey of patients or clients about AI receptionists or voice agents in any of the four professions is NF; the nearest measures are Pew's discomfort with AI in clinical care, Talker's forecast about scheduling and the Gartner finding on customers in general325. We do not estimate a preference from them: they measure clinical comfort, a forecast and a general attitude to customer service, and no arithmetic joins the three.

Quality and error: what the accuracy studies measured

A randomized or field study of an AI agent answering inbound calls or scheduling in a dental, medical, legal or accounting setting is NF, and so is a peer-reviewed containment rate for inbound calls. The nearest evidence is one peer-reviewed study of outbound calls and the vendor reports below. At Mount Sinai in 2025, an AI voice assistant made pre-procedure calls to patients, 1,606 calls to 1,431 patients across two phases. In the first phase it completed 86.4% of calls47. Of all its calls in that phase, 36.6% were fully automated and needed no human intervention47, and an AI hallucination led to an incomplete call in 6 calls, 0.7%47. The calls were outbound, at one cath lab, with no control group, so the study does not measure a front desk.

Relatient† (patient scheduling) said in an August 2025 release that its voice product answered 62% of inbound calls, of which 72% were either autonomously handled or escalated within the organizations48. Notable†'s customer story for one health system says its voice assistant handled 40% to 60% of inbound insurance inquiries331. Both are marketing material with no sampling described, and "handled or escalated" is not containment.

The accuracy studies in Table 5 are of other tasks, and we do not carry a result from its task to another.

StudyTask measuredResultWhat the task is not
Kini and co-authors, 2026 (peer reviewed)Outbound pre-procedure calls by an AI voice assistant at Mount Sinai, 1,606 callsFirst phase: 86.4% of calls completed; hallucination led to an incomplete call in 0.7%47An inbound line; a control group
Dahl, Magesh, Suzgun and Ho, 2024 preprintSpecific, verifiable questions about random federal court casesIn the preprint, legal hallucinations ran from 58% of the time with ChatGPT 4 to 88% with Llama 2230A call to a firm's intake line
Magesh and co-authors, 2024Queries to AI legal research tools from LexisNexis and Thomson ReutersEach hallucinated between 17% and 33% of the time231A conversation with a prospective client
Kung and co-authors, 2023The three United States Medical Licensing Examination examsChatGPT at or near the 60% passing threshold on all three332A conversation with a patient
Ayers and co-authors, 2023195 public patient questions posted to an online forum, judged by licensed health professionalsChatbot responses preferred to the verified physician's in 78.6% of 585 evaluations142A practice's own patients, or an unsupervised reply
Wood and co-authors, 202328,085 accounting assessment and test-bank questionsStudents averaged 76.7%; ChatGPT as of January 2023, 47.5% without partial credit and 56.5% with it281Advice to a client
Eulerich and co-authors, 2024The CPA, CMA, CIA and EA examsChatGPT 3.5 averaged 53.1% and passed none; after moving to ChatGPT 4 with further improvements it passed all, averaging 85.1%282Work for a client
Liu, Zhang and Liang, 2023An audit of commercial generative search enginesOn average 51.5% of generated sentences fully supported by citations and 74.5% of citations supporting their sentence158A voice conversation

Table 5. Accuracy studies, each with the task it measured.

The tasks differ from a front-desk call in how an error is counted. An examination has a key, and a forum question has no caller waiting and no booking at stake. A search answer is judged by whether its citations support it, and a call by whether the caller's need was met and the practice's facts were stated correctly. The studies indicate that error rates depend heavily on the task and the model, and that the rates measured on legal questions and accounting exams were not negligible. They do not show a rate for any of the four settings.

In the one voice task for which this paper cites an accuracy figure, Nation's Restaurant News reported that a 2022 report from BTIG showed McDonald's voice ordering accuracy in the low 80% range, against a goal of 95% or better333. It separately reported that McDonald's confirmed it would end its test with IBM333. It is trade press summarizing a report whose method this paper does not cite, and in the words cited it does not tie the end of the test to accuracy. A voice order for food is not a practice's call.

In a study of consultants, Dell'Acqua and co-authors found that for a task inside the technological frontier those using AI completed tasks 25.1% more quickly and produced results of more than 40% higher quality263. For a task outside the frontier they were 19 percentage points less likely to produce correct solutions263.

How a narrow scope and a hand-off limit the damage

The accuracy studies show that errors occur and that the task decides how often. NIST's generative AI profile defines confabulation as the production of confidently stated but erroneous or false content, by which users may be misled or deceived190. The design question is therefore how far a confident error can travel.

The first limit is on what the assistant may say. If it states only hours, location, accepted plans and open appointment times, drawn from the practice's own approved facts, every statement can be checked against a source, and the practice can test the line itself. Retrieval from approved text does not remove error: Magesh and co-authors found that the retrieval-based legal research tools they tested hallucinated between 17% and 33% of the time, despite vendor claims of eliminating errors231. Those tools answer legal research questions, a wider task than a short list of a practice's own facts. We think a short list of permitted statements reduces what can go wrong, and the hand-off handles whatever the list leaves out.

The second limit is on what the assistant may do. OWASP defines excessive agency as the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM, regardless of what is causing the malfunction312. A caller can say anything to a voice agent. An assistant whose only tools book into open slots, take a message and transfer a call cannot, whatever it is told, issue a refund, accept a matter or change a record, because it has no tool that does so. The reported case of the National Eating Disorders Association is an example of a bot acting outside its program. CBS News reported that the association said the current version of its Tessa chatbot, running its Body Positive program, may have given information that was harmful and unrelated to the program334. It reported that the association took the program down pending an investigation334. The statement is reported secondhand.

The third limit is that a person remains answerable. In Moffatt v. Air Canada, a tribunal in British Columbia found the airline liable for negligent misrepresentation because its website chatbot told a customer that he could apply for bereavement fares retroactively191. The tribunal rejected the suggestion that the chatbot was a separate legal entity, saying that the airline is responsible for all the information on its website, whether it comes from a static page or a chatbot191. It is a decision of a provincial tribunal in Canada, and we do not offer it as authority in the United States. In Utah, it is not a defense to the violation of a statute administered and enforced by the Division of Consumer Protection that generative AI made the violative statement, undertook the violative act, or was used in furtherance of the violation294.

Brynjolfsson, Li and Raymond studied 5,179 customer support agents in the staggered introduction of a generative AI conversational assistant and found that access increased productivity, measured by issues resolved per hour, by 14% on average335. It measured an assistant helping human chat-support agents, not an AI agent replacing them, and not voice calls. A 2026 arXiv preprint reports a randomized field experiment on Alibaba's Taobao platform. Deploying an agentic AI system for AI-eligible customer chats reduced average chat duration and had limited effects on retrial rates, but substantially lowered ratings for those chats336. Neither tells a practice what an assistant would do on its line.

The FTC alleged in a 2025 complaint that Air AI advertised a product that can replace human customer service representatives337. This paper cites no finding on the allegation; we cite it only as a reason it makes no claim that an assistant replaces staff.

What unanswered calls and missed visits are worth: two chains and a service case

The cases below multiply cited links and leave one open: the share of unanswered callers that an agent turns into a conversation. No study has measured that share for a front desk, so each case is solved for the share the agent must reach to cover its price, and no value is supposed for it.

The agent is $299 a month in every profession, its published price with 250 call minutes included, which is about 62 four-minute calls (250 / 4)56. A minute after that costs $0.3956, and every case below stays inside the included minutes.

A dental or medical office also needs HIPAA-ready Essentials around it, which is $699 a month and $125 a month for its $1,500 setup spread over 12 months57, so $824 a month in the first year and $699 after it. The program is a compliance layer and does not include the agent57. Every tool that handles patient information runs inside it: the agent on patient calls, the site assistant, appointment forms, texts to patients and call tracking57, and so do the front-desk add-ons (reminders, recall, waitlist backfill and intake), which have their own prices57. Its fee is therefore shared by everything on it, and no single gain is asked to carry it. The agent and Essentials together are $1,123 a month in the first year ($299 + $824), which is the narrow case of a practice that buys a program for the agent and runs nothing else on it.

The plan, agent and Essentials together are $2,122 in the first year ($999 + $1,123) and $1,997 after it ($2,122 - $125), with the plan at $99956. The plan produces demand and is set against new patients. The agent answers that demand and is set against answered calls and front-desk hours. The program and the reminders add-on, $973 a month together in the first year ($824 + $149 for reminders57), are set against kept visits, because reminders are what the two no-show studies measured.

Dental: calls that hang up

A dental office has the best-documented chain because Patient Prism† has published its call counts as a sequential funnel: the 34 is a share of the 69 calls that connect, and the 21 a share of the 3424. The callers are new and existing patients together, so a booking is valued at two ends, one visit at the low end and a new patient's first year at the high end. The truth for a practice lies between them, and its own phone log says where. Table 6 gives the links.

LinkFigureSource
Calls a month per location116Patient Prism†: 11,552,668 calls at 8,280 locations in 2025, divided by 12 (calc., above), not a measured average per office24
Callers who hang up before reaching anyone31%, so 36 a month; 69 in 100 get throughPatient Prism†24
Of the 69 who get through, calls that are a real chance to book34, or 49%; the other 35 are not bookablePatient Prism†'s own classification24
Chances that end in a booking21 of 34, or 62%Same source
Bookings a month among the hang-ups, if they resemble callers who get through and every one were answered11.0 (36 x 21 / 69)calc.
A booking, low end: one visit$282ADA billings per dentist, $965,66021, divided by a visit count whose source is not named, 3,429.5153, so a rough figure
A booking, high end: a new patient's first year$887AHRQ MEPS118
Value that hangs up each month$3,089 to $9,730calc.
New patients a month, average practice30Henry Schein One†49

Table 6. The dental call chain, link by link. The first five links are Patient Prism†'s calls and its classification of them, the values are an association's billings and a household survey, and the last link is Henry Schein One†'s.

At 36 hang-ups a month, the agent at $299 is covered if it turns 3% of them into conversations when each booking is a new patient's first year (0.3 bookings, $299 / $887). When each booking is one visit, that is 10% (1.1 bookings, $299 / $282). A practice that buys the program for the agent alone and runs nothing else on it would need 12% and 36% (1.3 and 4.0 bookings, $1,123 / $887 and $1,123 / $282), which is the narrow case. Table 7 puts the open link at four values and shows the narrow case in its last column.

Share of hang-ups turned into conversationsCalls a monthBookingsGain, one visit eachGain, a new patient eachLess the agent ($299)Less the agent and a program bought for it alone ($1,123)
10%3.61.1$309$973$10 to $674-$814 to -$150
25%9.02.7$772$2,433$473 to $2,134-$351 to $1,310
50%18.05.5$1,544$4,865$1,245 to $4,566$421 to $3,742
75%27.08.2$2,317$7,298$2,018 to $6,999$1,194 to $6,175

Table 7. The open link at four values, dental (calc., Table 6). The four shares are illustrative, and the table is not a forecast. The last column is the narrow case of a practice that runs nothing else on the program.

The chain is set against the agent's own fee, $299 a month, because answered calls are the agent's gain. A dental office also needs the HIPAA-ready program around the agent57, but the program is a layer and not a tool: the site assistant, texts to patients and the front-desk add-ons run inside the same one57, so its fee is shared by everything on it and is not charged to answered calls alone. The plan is not charged here either, because its product is new demand, which the plan's own case counts. The plan alone, at $999 a month, is covered by 1.1 new patients a month ($999 / $887), which is 3.8% more than the average practice's 30 ($999 / $887 / 30).

The chain assumes that callers who hang up resemble callers who get through, which the source does not show. It counts none of the callers who ring back, which lowers every figure, most of all at the low end where the callers are existing patients. The 31 hang-ups per 100 count all callers and not new patients, and the 21 booked per 100 is a classification of calls and not a booking rate for all callers. The $887 is spending per person with any dental visit, from all payers, reported by households, so it is not what a practice collects. The nearest peer-reviewed figure for an AI voice agent finishing its calls is the 86.4% above, for outbound calls at one hospital47.

Medical: calls that reach no one, and visits that are missed

For a medical office the links are looser, and Table 8 uses two measures of a call that reaches no one. The Veterans Health Administration's call centers abandoned a mean of 12.0% of calls in October 2014159 and 8.3% by September 2016159, the later figure being the one Table 8 uses. Among Invoca†'s health care customers, 54% of calls were answered by a person26, so 46% were not (100% - 54%). Of calls answered by a person, 43% were leads and 45% of those converted on the call, by Invoca†'s own classification and across six health care fields26. Call volume is Call Box†'s 299 to 505 calls per office location in 2020198. A booking is valued at the same two ends as in the dental case: one primary care visit at a mean payment of $18661, and a new patient's first year at $300, which is derived in Chapter 11, Table 2.

Share of calls reaching no oneCalls a monthCalls missedBookings lost if they convert like answered callsValue, one visit each to a new patient eachShare the agent must hold to cover its fee ($299)The same, for a program bought for the agent alone ($1,123)
8.3% abandoned (VA call centers, 2016)299254.8$893 to $1,44121% to 33%78% to more than all of them
8.3% abandoned (VA call centers, 2016)505428.1$1,509 to $2,43312% to 20%46% to 74%
46% not answered by a person (Invoca†, health care, 2026)29913826.6$4,950 to $7,9854% to 6%14% to 23%
46% not answered by a person (Invoca†, health care, 2026)50523245.0$8,361 to $13,4872% to 4%8% to 13%

Table 8. The medical call chain (calc. from the links in the text). The conversion of missed calls assumes they would have converted as answered calls do. Neither miss rate is a private physician office. The last column is the narrow case of a practice that runs nothing else on the program.

Answered calls are the agent's gain, so Table 8 sets them against the agent's own fee. At the VA rate, an office with 299 calls a month covers $299 if the agent holds 21% to 33% of its 25 missed calls, and an office with 505 calls a month if it holds 12% to 20% of its 42 (calc., Table 8). At the 46% rate the shares are 4% to 6% and 2% to 4%. Every share is below the 86.4% of calls that the hospital study's agent completed47, and the highest, 33%, is below the 36.6% of calls the same agent finished with no person47. Neither the hospital's outbound calls nor the VA's call centers are a front desk answering calls, so the comparison shows how the required shares sit against those rates and not what an agent at a front desk would hold. In the narrow case of a program bought for the agent alone, the VA rows need 78% to more than all of the missed calls at 299 calls a month and 46% to 74% at 505 (calc., Table 8).

At each share in the fee column the agent's calls fit inside its 250 included minutes, so its fee is $29956. An agent that took every missed call in the last row, 232 calls at four minutes (928 minutes), would cost $563.42 ($299 + 678 minutes x $0.39), against calls worth $8,361 to $13,487.

Kept visits and front-desk hours are separate gains on top of this one. Table 9 sets the agent's $299 against front-desk hours, and the program and the reminders add-on, $973 a month together, against a physician's missed visits. Its inputs are 51.1 appointments a week for a primary care physician in Epic data195, which is 221 a month (51.1 x 52 / 12), and no-show rates of 18.8%162 to 23%142. The payment is a mean of $186 a primary care visit in 201661, and front-desk cost is $34.59 an hour for office and administrative support in health care, including benefits, in June 2026196. The effect of reminders is the 0.5 to 3 points from the reminders section, the VA before-and-after comparison and the safety-net clinic trial.

One physicianFour physicians
Appointments a month221886
Missed visits a month at 18.8% to 23%42 to 51167 to 204
What one point off the no-show rate is worth a month$412$1,647
Kept visits needed to cover the program and reminders ($973)5.2 (2.4 points off the no-show rate)5.2 (0.6 points off the no-show rate)
Kept visits needed to cover reminders alone, where the program is already in place ($149 / $186)0.80.8
Kept visits if no-shows fall by 0.5 to 3 points1.1 to 6.6 ($206 to $1,236)4.4 to 26.6 ($824 to $4,942)
The agent's fee in front-desk hours ($299 / $34.59)8.6 a month (about 25 minutes a working day)8.6 a month (about 25 minutes a working day)

Table 9. The program and reminders against kept visits, and the agent against front-desk hours, medical (calc. from the cited inputs). The 18.8%, which is a VA figure, and the 51.1, which comes from health-system data, are not measures of a small private office.

The agent's own fee, $299, is what 8.6 front-desk hours a month cost ($299 / $34.59), about 25 minutes a working day (8.6 x 60 minutes / 21 working days), and Table 8 gives the missed calls that cover it. Those hours are the fee and not the talk time that covers it: on talk time alone the fee is covered at about 18 hours of calls a month, because minutes past the 250 included are billed (1,080 minutes cost $622.70, $299 + 830 x $0.39, and release $622.62, 18 x $34.59). A solo primary care physician is the tightest case for the program. Covering the program and the reminders takes 2.4 points off the no-show rate for one physician, toward the top of the 0.5 to 3 points found, and 0.6 points for four. A solo office would therefore cover the program by running more than one tool on it: the agent's answered calls, reminders, recall, waitlist backfill and intake all sit on the same layer57.

One physician has 171 to 180 kept visits a month once 18.8% to 23% are missed, and they bring payments of $31,714 to $33,444 at $186 a visit (calc.). That is 6.3% to 6.7% for the whole stack of $2,122 a month, and 1.6% to 1.7% for four physicians. Covering it takes 11.4 kept visits, which is 22.4% to 27.4% of one physician's missed visits and 5.6% to 6.9% of four physicians'.

Counting new patients instead, the plan alone needs 3.3 a month. That is 15.6% more than the 21 to 22 new-patient visits a month that one physician has now at 12.5% of kept visits, and 3.9% more for four physicians' 85 to 90. The whole stack needs 7.1 a month on a first-year count, 33.2% more for one physician and 8.3% more for four, and 2.4 a month if patients stay three years. The payment per visit is the 2016 primary care mean, and the 2023 mean for all doctor visits is nearly twice that193.

The mean is what a practice collects per visit on average, but a few costly visits pull it above the typical visit, which the median shows. Table 10 repeats the counts at three payment levels, and the median is the tightest.

Payment per visitA patient's payments in a yearNew patients a month to cover the planKept visits a month to cover the program and reminders ($973)Kept visits to cover everything
Median primary care visit, 2016: $10761$1735.89.119.8
Mean primary care visit, 2016: $18661$3003.35.211.4
Mean visit to any medical doctor, 2023: $347193$5601.82.86.1

Table 10. The same counts at three payment levels (calc.). The plan's count, which is for the plan alone at $999 a month, is on a first-year count.

Payments are what insurers and patients paid and not what the practice keeps after the cost of the visit. A kept visit in a slot that is already staffed adds nearly all of its payment, and a new patient who needs added capacity adds less, so a practice should use its own margin. At the median payment, one physician needs 9.1 kept visits a month to cover the program and reminders. That is 4.1 points of a 221-appointment schedule (9.1 / 221), more than the 0.5 to 3 points that the two reminder studies found.

The hours an agent releases in a medical office are NF, and we do not estimate them, because the hours depend on call length, which no cited source gives for a front desk. The nearest evidence is the MGMA poll above, in which leaders named scheduling as a time-intensive phone task in 31% of responses197.

Law and accounting: no call count, so the other direction

With no source for calls a month at a law firm or an accounting firm, the chain runs the other way, from the price to the clients needed to cover it, with each client's value from Chapter 11, Table 2. The evidence on how many callers a firm loses is Clio†'s 2024 phone test above23. In Clio†'s 2019 consumer survey, 57% of those who had shopped for a lawyer contacted more than one firm212, and 42% said they would stop looking if they liked the first lawyer they spoke with212. A caller who reaches no one at a firm may therefore reach a competitor. For accounting, the share of calls missed is ~3 to 6 in 10, as above.

Cost: our price against the BLS wages

The cost comparison puts two prices in the same units and draws no conclusion from them. Our AI voice agent is priced at $299 a month with 250 call minutes included, and $0.39 a minute after that56. The Bureau of Labor Statistics reports a median hourly wage of $18.27 for receptionists in May 2025 and $21.53 for customer service representatives338,339. For health care offices, the Bureau also reports employer compensation of $34.59 an hour worked, wages and benefits together, for office and administrative support occupations in the health care and social assistance industry in June 2026196. Table 11 sets out the arithmetic. It supposes, for illustration, one person working 40 hours a week for 52 weeks, which is 2,080 hours.

QuantityArithmeticResult
Our AI voice agent for a year, with 250 minutes a month included$299 x 12$3,588
Receptionist median wage for 2,080 hours$18.27 x 2,080$38,001.60
Customer service representative median wage for 2,080 hours$21.53 x 2,080$44,782.40
Hours of receptionist wage that $299 buys$299 / $18.2716.4 hours
Hours of customer service representative wage that $299 buys$299 / $21.5313.9 hours
Hours of health care front-office compensation that the agent ($299) buys$299 / $34.598.6 hours
Hours of the same compensation that the agent and Essentials buy, the narrow case of a practice that runs nothing else on the program ($1,123)$1,123 / $34.5932.5 hours

Table 11. Arithmetic on our price and the BLS median hourly wages and loaded compensation. The wage rows are wages and not the cost of an employee.

The health care compensation figure includes benefits, which are 32.0% of it in the source196, but it is a broad occupation group and not receptionists alone. Minutes past the 250 included cost $0.39 each, and a practice should ask in writing how minutes are counted. For a dental or medical office, the HIPAA-ready program carries its own prices57, which are shared by every tool that runs on it57 and are in none of the rows of Table 11 above the last.

A median wage is the price of a job, and a job is a bundle of tasks; the voice agent is the price of one of them, within the limits of Table 4. The studies in the quality section give no containment rate or cost per contact for an inbound line, and we make no claim of savings here. What the cited figures allow is a comparison of the price with what it has to recover, and Table 12 makes it for each profession. Whether a practice's missed calls cost it more than the price depends on its own count of missed calls and its own value per client, which its log and records give.

ClientAgent on a plan ($299 a month)Plan plus agent ($1,298 a month)Agent and Essentials on a plan ($1,123 a month)Plan, agent and Essentials ($2,122 a month)
Primary care patient12.051.944.984.9
Dental patient4.017.615.228.7
DUI defense matter1.04.53.97.3
Auto injury case0.52.32.03.8
Advisory client at a CPA firm that sells advisory services0.20.90.81.4
Bookkeeping client, most common vendor band1.25.24.58.5

Table 12. New clients a year needed to cover each purchase, on first-year value alone (calc. from the stacks above and the client values in Chapter 11, Table 2). A health care row uses the two Essentials columns when the purchase touches patient information, and a law or accounting row uses the first two. For a health care practice the first column is what the agent costs, and the Essentials columns add the program that every tool handling patient information shares.

Table 12 counts first-year value only. A client who stays longer needs fewer new clients: counting three years, the full stack needs 28.3 new primary care patients a year ($2,122 x 12 / $900), which is 2.4 a month, and over 8 years 0.9 a month (calc.). It also sets each purchase against new clients alone, which understates a system whose agent also recovers missed calls and kept visits, as Tables 7 and 9 show.

For law, an hour billed at $300 brings in $245.52 after realization of 88% and collection of 93% ($300 x 88% x 93%)52,59. The lowest plan and the agent at $1,298 a month are therefore covered by 5.3 collected hours a month ($1,298 / $245.52). The average lawyer in Clio†'s data bills 38% of an eight-hour day, which is 3.0 hours, and has 5.0 hours unbilled52. A law firm is not short of hours, so we think the gain from more demand there is choosing better matters and setting a higher minimum fee more than filling time.

For accounting, 0.9 new advisory clients a year cover the lowest plan and the agent at a CPA firm that sells advisory services. That figure comes from a benchmark of 206 practices that chose to answer a survey about a service they sell well113, so it is the most favorable case for a small firm that does mostly tax returns. A firm that sells only base Form 1040 returns at the NATP average fee of $23660 needs about 66 such clients a year (12 x $1,298 / $236).

To test a front desk of any kind before relying on it, Chapter 13 gives the first step: call the practice's own line as a stranger would, in and out of hours, and write down what happens. Chapter 11 gives the counts that keep the answer current: calls received, calls answered, calls missed and the time to the first reply.

Chapter 13

Building the system: a first year

A first year as a sequence: framing, plumbing, finish, operating

This chapter is for you if you own or manage a dental practice, a clinic, a law firm or an accounting firm. It turns the argument of the earlier chapters into an order of work. Framing comes first, because nothing can be wired until you know what the system has to do. Plumbing comes second: the profiles, tracking, forms, calls, booking and vendor agreements through which a person finds you, reaches you and hands you information. Chapter 3 introduced the order of framing, plumbing and finish, and the difference between a brochure site and a working rep for your front office.

Read the months as a sequence, not a calendar of results. We give no timeline for when anything will be found, chosen or booked, and the sequence promises no outcome, because assistants change their answers and a practice's results depend on much that no system controls. Each step is tied to the rule or finding that makes it necessary, with a pointer to the chapter that established it. Nothing here is legal, medical, tax or financial advice.

PhaseMonthsWhat exists at the endEstablished in
Framing1 to 2A system map, the questions that bring work, a data inventory, a claims inventory, a dated baseline of what three assistants answerChapters 3, 4, 5 and 10
Plumbing2 to 4Consistent profiles and listings, first-party counting, forms, call handling, booking, a vendor ledger with agreements signedChapters 4, 6 to 10 and 12
Finish4 to 6Pages written from a claims register, a site that answers, books and hands off, a reviews programChapters 3, 5 to 9
Operating6 to 12A monthly dated check, a monthly report of counts, a change log, a rules calendarChapters 4, 10 and 11

Table 1. The first year at a glance.

Months 1 to 2: framing before any page is touched

Start with four documents and one baseline, none of which needs software. The documents are a map of the system, a list of the questions that bring work, and two inventories, one of the places where information enters your practice and one of the claims you make about yourself. The baseline is a dated record of what the AI assistants answer today. Together they tell you what the later months have to build.

The system map is the pipeline from Chapter 3 drawn for your practice. It starts where a person asks a question, whether to a search engine, an assistant, a friend or a directory. It ends where an appointment or a matter is kept and the person returns or refers someone else. Between those points mark every handoff and write the name of the person on your staff who owns it. The map shows where first contact lands and who catches it, since Chapter 12 treats first contact as the conversion point.

For the question list, write down what your front desk hears most often, taken from your call log, your intake forms and your inbox. The questions fall into three kinds: who and where (finding), whether you take this insurance or this kind of matter (choosing), and how to book (reaching).

Then run the baseline. Put the questions to ChatGPT, Gemini and Perplexity, starting from the wording of our study, "Who is a good dentist in Atlanta, GA? Name specific businesses."1, and add the specialty, insurance, emergency and language questions that Chapter 4 noted may behave differently. Record the date, every business named and every site cited. Our study found that 79.7% of the businesses named were named by only one of the three assistants1, which is why you ask all three. Each question was asked once in our study, and answers through an API are a baseline, not what every signed-in user sees1, so treat your own first pass as one dated draw.

The data inventory is the data map from Chapter 10: every place information enters (phone, form, chat, booking page, email, review inbox, listing messages), what could be said there, and every tool that can see it. The claims inventory lists what you say about yourself today, including specialties, results, comparisons, testimonials and prices. Do both before you build anything, because the pages and listings of months 4 to 6 are written from the claims register, and a claim that is already live is already speaking for you. Chapters 6 to 9 explain the profession's rules on both. Find out at this stage which state rules and which professional-body texts apply, and which of the disclosure texts in Table 4 of Chapter 10 reach you.

The data inventory sets the size of the plumbing work. Every tool that can see patient or client information becomes a row in the vendor ledger of the next phase, and every row needs an agreement or a consent checked. A two-location practice that counts its places of entry and its tools in month 1 learns then, and not in month 5, how many agreements it has to check. Our check of 28 software tools that small practices commonly use shows the range, according to the vendors' own pages: 9 sign a business associate agreement without a special tier, 15 only on certain plans and 4 offer none54.

Months 2 to 4: plumbing, from listings to vendor agreements

Plumbing is the longest phase, and it is where most of the rules bite.

Profiles and listings

Make the facts about your practice identical wherever an assistant may read them: name, address, phone number, hours, services and the people who work there. In our study the 25 most-cited websites drew 29% of all citations, and they were directories, review platforms, list sites, publishers and a forum, not any one practice's site1. We infer, as Chapter 4 does, that third-party profiles carry weight. The wording on each profile comes from the claims register, so a specialty claim on a listing is the one approved for your site. Before you pay for a listing or a lead service, check whether it is an advertisement or a recommendation under your profession's rules; Chapter 8 covers the lawyers' rule on paying for recommendations.

Tracking

Set up counting that does not depend on advertising tags. Count calls, form submissions and booking confirmations inside systems you hold an agreement with, and ask each new patient or client how they found you. Keep advertising-platform tags off the pages where a person states a condition, a matter or a return detail, which Chapter 10 calls action pages. Google Analytics and the Meta Pixel are two of the four tools in our check that offer no business associate agreement54. Chapter 7 explains the HHS bulletin and the court decision behind this, and Chapter 10 sets out the design. Whether HIPAA, Washington's My Health My Data Act or neither applies to a given page is for your counsel.

Forms

Every field on a form is a flow of information, so give each form three things: the fewest fields that do its job, a plain statement of what happens to what is entered, and a destination that appears in your vendor ledger.

For a law firm, add what the form does not create. ABA Model Rule 1.18 treats a person who consults a lawyer about forming a client-lawyer relationship as a prospective client and bars the lawyer from using or revealing what was learned even if no relationship follows, subject to the rule's exceptions219. The Model Rules are model texts; check which version of each rule applies to you. Florida's opinion warns that a website intake chatbot can create such a relationship without the lawyer's knowledge225. For an accounting firm, a tax return preparer may not disclose or use return information before obtaining the taxpayer's written consent unless section 7216 or 26 CFR 301.7216-2 authorizes it273. Where a form would send return information beyond what the regulations authorize, design the consent step before the form goes live. Chapter 8 treats the lawyers' duties and Chapter 9 the preparers'.

Calls

Phone calls are where demand arrives, so decide who answers, when, and what an assistant says. If an AI assistant answers some calls, settle before it goes live the opening statement that it is an AI assistant, whether and how calls are recorded, and how a person takes over. Washington, for instance, makes it unlawful to record a private communication without the consent of all the participants179. Chapter 10 compares the disclosure and recording texts, and Chapter 12 applies them to the call. Our AI voice agent is priced at $299 a month with 250 call minutes included, and $0.39 a minute after that56. It answers, books and hands off to the office team, and it tells every caller it is an AI assistant. For a dental or medical office it runs inside a HIPAA-ready program and is bought as a service: speed of answer and more hands at the front desk. Before any assistant takes a real call, call your own line as a stranger would, in and out of hours, and write down what happens.

Booking and reminders

Treat the booking tool and the reminder tool as vendors in the chain. HHS's tracking bulletin gives the example of an appointment made through a clinic's website, where a vendor receiving the appointment and the IP address is a business associate and an agreement is required12. Chapter 6 covers booking, reminders and the evidence on attendance, and Chapter 7 covers access.

If you later add an AI voice for reminder calls, that is a call you initiate. The FCC's 2024 ruling says that calls using AI-generated voices require the prior express consent of the called party to initiate them, absent an emergency purpose or exemption186. The FCC's own 2024 proposal said the Telephone Consumer Protection Act's requirements do not extend to technologies used to answer inbound calls187. That proposal is not a ruling, and whether it reaches a line that patients or clients call is for counsel. Reminder and rebooking calls are one of four front-desk add-ons that run inside a HIPAA-ready program57.

Agreements with vendors

Fill the vendor ledger and sign before information flows. For each tool, record its role, what it can see, the kind of agreement that covers it, the plan that carries the agreement and the date you checked. Of the 15 tools that sign a business associate agreement only on certain plans, 6 do not publish the price of the qualifying plan54. So you ask for the price, buy the qualifying plan and keep the evidence. HHS publishes sample agreement provisions. It says that using them is not required for compliance and may not alone be sufficient to form a binding contract under state law168, so the agreement itself still has to be read. For a law firm the instrument is supervision of the outside tool221, and for an accounting firm it is a confidentiality contract or the client's specific consent8.

Months 4 to 6: the finish, pages, design and a reviews program

With the plumbing in, the pages can be written and designed without changing where information goes. We think that is the reason for the order: a page designed first invites a form placed wherever it looks best, with no agreement yet behind its destination. By month 4 the ledger and the data map exist, and the claims inventory, reviewed by your professionals, becomes the claims register. The finish works inside all three.

Write each page to answer one question people ask, and draw its facts from the claims register. The pages most practices need are a page for each service, a profile for each dentist, physician, lawyer or accountant, a page for each location, a page of frequently asked questions and the booking or intake page. Chapter 5 explains why headings carry compound, specific phrases, and Chapters 6 to 9 set out each profession's edge on claims. For a law firm, every communication made under ABA Model Rule 7.2 must include the name and contact information of at least one lawyer or firm responsible for its content68, so the pages carry that name.

The design is the part of the job that looks like web design, and Chapter 3 gave it a job: the site answers, books, qualifies and hands off. Each page ends with one next step, which is to book, to call or to ask, and each next step lands in a path built in months 2 to 4. The hand-off goes to a named person on the office team. Test it as a stranger would, on a phone, and write down what you could not do.

The reviews program is the last piece. Ask every patient or client in the same way. The FTC's rule makes it an unfair or deceptive act to provide compensation or incentives conditioned on reviews expressing a particular sentiment, positive or negative84. No reward in the program depends on what a review says, and the program uses no fake reviews. The rule names reviews that misrepresent that the reviewer exists or had the experience described, and the FTC has said it covers AI-generated fake reviews83,210. The rule also prohibits using unfounded or groundless legal threats, intimidation or knowingly false public accusations to prevent a review or to have it removed85. Write the reply policy before the first reply goes out; the mistakes section below shows what replies in the enforcement record contained.

Months 6 to 12: operating, with a monthly check and a rules calendar

By month 6 the system exists, and the remaining months are for running it.

The monthly check repeats the baseline from month 1: the same questions, the same wording, the same three assistants, on about the same day each month, with the names and cited sites recorded. The record matters more than any single answer, because one check is one draw from a process we have not shown to be stable1. Chapter 11 sets out the minimum plan. A change in a month's answers is not evidence that your work caused it, and a name appearing is not a measure of the quality of care or advice you give.

The monthly report counts outcomes by source: calls, forms, bookings and new patients or matters, plus what people say when asked how they found you. Chapter 11's worked example of the small-numbers problem shows how to read differences between months: at the average of 30 new patients a month that Henry Schein One† (software) reports for dental practices49, a month anywhere from about 24.5 to 35.5 sits within one step of the average (calc., Chapter 11). The report therefore shows both numbers and a note, not an explanation.

The change log keeps the other documents accurate. Any new tool, any plan change, any new form field, any new claim and any change to what the assistant says goes through the register, the ledger or the map first, is approved by its owner and is dated. Vendor positions need a scheduled recheck, because 15 of the 28 tools sign a business associate agreement only on certain plans54, and a vendor page you read in month 3 may read differently in month 9.

The rules calendar lists the dated changes to watch. The items below are the ones in our evidence, and the list is not complete.

  • HHS proposed Security Rule changes on January 6, 2025; the Unified Agenda entry we read lists final action as 07/00/2027, so the proposal was not a final rule on that entry170.
  • Colorado's reenacted statute on automated decision-making technology takes effect on January 1, 2027, apart from a few sections that take effect on passage303.
  • HHS's tracking bulletin page says the court vacated the guidance only to the extent it provides that HIPAA obligations are triggered where an IP address is connected with a visit to an unauthenticated public page about specific health conditions or providers12. Check that page for later changes.
  • The State Bar of California's 2026 guidance on generative AI replaced its November 2023 version, so check which version you are reading226.
  • HHS publishes inflation-adjusted civil money penalty amounts; the table we read was published on January 28, 2026, and you should check for a later adjustment129.

Once a year, repeat the data inventory and the claims inventory, and re-verify every vendor.

What to buy, what to build and what to leave alone

We think you should buy what needs a contract, uptime and a team to run it: telephony and answering, scheduling and reminders, a customer relationship tool, call tracking. Build the documents. The claims register, the vendor ledger, the data map, the responsibility map, the disclosure script and the rules calendar are documents, not software, and a dated spreadsheet with a named owner for each serves.

Our plan prices are one reference point for the buying decision, and this chapter gives no other provider's prices. Table 2 shows ours with twelve months of arithmetic beside each, and Table 3 shows the stacks a practice would buy, by scope. What each plan contains is outside the scope of this paper, and the HIPAA-ready programs begin with a consultation. A marketing plan alone needs no HIPAA program, because a plan carries no patient information57. The least a dental or medical practice can buy is therefore a plan, at $999 a month56. A program is a compliance layer added to a plan, and it does not include the voice agent57. Every tool that handles patient information runs inside it, the agent on patient calls among them, and the front-desk add-ons run inside it at their own prices57, so its fee is shared by everything on it and no single gain is asked to carry it. A law or accounting firm needs no program and buys the agent alone or with the lowest plan.

ItemPriceTwelve months (calc.)Source
Visibility plan, first location$999 a month with a $499 onboarding fee, or $9,990 for a 12-month term paid in full, two months free (saves $1,998), with the fee waived$11,988 plus $499 onboarding (12 x $999), or $9,99056
Growth plan, first location$1,999 a month with a $999 onboarding fee, or $19,990 for a 12-month term paid in full, two months free (saves $3,998), with the fee waived$23,988 plus $999 onboarding (12 x $1,999), or $19,99056
Engineered plan, first location$2,999 a month with a $3,000 setup fee, or $29,990 for a 12-month term paid in full, two months free (saves $5,998), with the setup fee at $1,400$35,988 plus $3,000 setup (12 x $2,999), or $29,990 plus $1,400 setup56
Each additional location, on any plan$299 a month$3,588 (12 x $299)56
AI voice agent (AI receptionist)$299 a month with 250 call minutes included, then $0.39 a minute$3,588 (12 x $299), plus $0.39 a minute past 250 a month56
Managed marketing led by Sezen Güner$1,800 a month on every term, with no term discount$21,600 (12 x $1,800)56
Ads management, an add-on12% of ad spend when we deploy and track; 20% when we also build the tests and the deployment strategy; $500 monthly minimum; content we create needs managed marketing too12% or 20% of the year's ad spend, at least $6,000 (12 x $500)56
HIPAA-ready program, Essentials, per location$1,500 setup plus $699 a month$9,888 in the first year57
HIPAA-ready program, Integrated, per location$3,500 setup plus $1,199 a month$17,888 in the first year57
Front-desk add-on: reminders and rebooking$149 a month, plus usage$1,788 plus usage57
Front-desk add-on: recall and reactivation$199 a month, plus usage$2,388 plus usage57
Front-desk add-on: waitlist backfill$99 a month, plus usage$1,188 plus usage57
Front-desk add-on: intake with insurance capture$149 a month plus $250 setup, plus usage$1,788 plus $250 setup, plus usage57
All four front-desk add-ons$449 a month, plus usage$5,388 plus usage57

Table 2. Technical Marketing's prices as one reference point, with first-year arithmetic. A 12-month term paid in full commits to a full year, and a monthly price over a year is twelve times the month. The HIPAA-ready programs are priced after a compliance consultation, and their rows include neither a plan nor the voice agent. The add-ons run inside a HIPAA-ready program, are priced per location, and the source does not say whether the intake setup also applies to the bundle of all four.

What is boughtWho buys itMonthlyTwelve months (calc.)
A plan alone: the marketingAny profession; a dental or medical practice needs no HIPAA program for it$999$11,988
The agent added to a plan the firm already hasA firm with a plan; law and accounting$299$3,588
The lowest plan plus the agentLaw and accounting$1,298$15,576
Essentials alone, the compliance layer that the agent and every other tool handling patient information runs inside, first yearDental and medical$824$9,888
Essentials after the first yearDental and medical$699$8,388
The agent plus Essentials, added to a plan, first yearDental and medical, the narrow case of a practice that runs nothing else on the program$1,123$13,476
The lowest plan, the agent and Essentials, first yearDental and medical, the full stack$2,122$25,464
The same after the first yearDental and medical$1,997$23,964

Table 3. The stacks by scope (calc., Table 2, on monthly billing, before the add-ons and before a plan's one-time onboarding fee). Every stack that includes the agent prices it at $299, with 250 call minutes included (about 62 four-minute calls, 250 / 4), and the worked cases stay inside those minutes; a month past them costs $0.39 a minute. The first-year rows spread the $1,500 Essentials setup over 12 months ($1,500 / 12).

Add the four front-desk add-ons at $449 a month to the full stack and the monthly cost is $2,571 in the first year ($2,122 + $449) and $2,446 after it ($1,997 + $449), before any usage the add-ons carry and before the intake setup if it applies. Chapter 12 works out what each fee has to recover: it sets the agent against answered calls and front-desk hours, the program and the reminders add-on against kept visits, and the plan against new clients. The plan is covered by 1.1 new dental patients a month, or by 3.3 new primary care patients a month on a first-year count (Chapter 14 gives the figures and the tightest case).

Suppose a two-location practice chose the Growth plan on monthly billing. The plan is $1,999 a month for the first location and the additional location is $299 a month56, so $2,298 a month ($1,999 + $299), or $27,576 over twelve months (12 x $2,298), plus the $999 onboarding fee. On a 12-month term paid in full the plan is $19,990 a year, two months free, with the onboarding fee waived56. The additional location is priced by the month, so the yearly total carries it at 12 x $299: $23,578 ($19,990 + $3,588). The arithmetic says what the plan costs, not what it will return, and we do not promise a return.

Leave alone what our evidence argues against. In our design, advertising-platform tags stay off action pages, nothing is paid for as a recommendation, no review is bought, and a machine does not answer clinical or legal questions without a human review. California's disclaimer duty is lifted for a generative AI message that a licensed or certified human provider has read and reviewed149. Formal Opinion 512 says that submitting an AI tool's output without an appropriate degree of independent verification could violate the duty of competence221. Chapter 12 weighs the evidence on the quality of AI answers.

Be wary of promises about AI. The FTC alleged in its complaint against Air AI that the company claimed its product could replace human customer service representatives337. It also alleged that consumers, many of them small business owners, lost as much as $250,000337. Those are allegations in a complaint, not findings. As for promised placement in AI answers, the authors of the paper that introduced generative engine optimization describe generative engines as black boxes that leave content creators little to no control over when and how their content is displayed31. Google says that no specific optimization is required for AI Overviews and AI Mode32. Our own study asked each question once and gives no basis for promising a place in any assistant's answer1.

Who is responsible for what inside the practice

Table 4 assigns each duty in this chapter to a role, and gives the rule or finding that makes it a named duty. The roles are generic, and a small practice will put several on one person. Chapter 10 gave the reason each duty needs a name: a practice can supervise only parties it can name.

DutyHeld byWhy it is a named duty
Approves every public claimA licensed professional in the practiceThe professional texts address the professional; for lawyers, communications must name a responsible lawyer or firm68
Keeps the claims register currentThe marketing lead, or the office managerFor lawyers, misleading truthful statements are prohibited (Model Rule 7.1, comment 2), so omissions need an owner204
Owns the data map and the vendor ledgerThe compliance lead, or the owner where there is noneDisclosure to a business associate requires satisfactory assurance documented in a written contract148
Reviews any AI-drafted message on clinical or legal mattersA licensed professionalThe California disclaimer duty is lifted for messages a licensed or certified provider reads and reviews149; Formal Opinion 512 says a lawyer's reliance on AI output without an appropriate degree of independent verification could violate the duty of competence221
Takes the hand-offs from the AI assistantNamed front-office staff, by shiftCalifornia's duty includes instructions for reaching a human149
Oversees the vendor and reads the monthly report and the assistants' answersThe owner or managing partnerFor lawyers, a partner or lawyer with managerial authority must make reasonable efforts to ensure measures giving reasonable assurance that nonlawyers retained by or associated with the lawyer behave compatibly with the lawyer's obligations (Model Rule 5.3)220
Keeps the rules calendar and takes rule questions to a lawyerCounsel or the compliance leadGuidance and rules change; the proposed Security Rule is not final170
Builds and operates integrations under agreementThe vendor, by named peopleA subcontractor that handles PHI for a business associate is itself a business associate, so each layer and each person is named55

Table 4. Duties inside the practice and who holds them.

A British Columbia tribunal found that Air Canada did not take reasonable care to ensure its chatbot was accurate and that the customer's reliance on it was reasonable191. That is a decision of a Canadian tribunal, not United States law. Utah's code provides by statute, within its scope, that it is no defense that generative AI made the violative statement294. The duty to approve what the assistant may say is therefore the first row of the table.

The mistakes that cost the most, by profession

These are the mistakes for which our sources show an enforcement action, a sanction, a decision or a stated penalty. We do not know how often any of them occurs, and none is a prediction about your practice. Each is described as the source describes it, and Chapters 6 to 9 give the profession's full rules.

Dentistry

The best-documented mistake is a review reply that discloses patient information. The Elite Dental resolution agreement recites a complaint that the practice, replying to a patient's Yelp review, disclosed her last name, details of her treatment plan, insurance and cost information122. The agreement also says OCR's review of the practice's Yelp page discovered that it had disclosed protected health information of other patients when responding to their reviews without valid authorizations122. The practice agreed to pay $10,000 and adopt a corrective action plan to settle potential violations121. The patient's complaint is an allegation, as the agreement recites it. New Vision Dental paid $23,000 and agreed to a corrective action plan to resolve an investigation of disclosures in response to online reviews125. Its complaint is likewise an allegation123.

In a third matter OCR imposed a $50,000 civil money penalty on a North Carolina dental practice that had disclosed a patient's PHI on a webpage in response to a negative review127. The practice also did not respond to OCR's data request or an administrative subpoena127. That was a penalty, not a settlement.

A second dental mistake is an announcement of specialization that the practice does not meet. The ADA code says that dentists who announce specialization should use "specialist in" and must devote a sufficient portion of their practice to the announced specialty to maintain expertise7. It is an ethics code, not law, and which state rules apply to specialist claims in your state is a separate question.

Medicine

The mistake most specific to a clinic is an advertising pixel on a booking page. In HHS's example, a vendor that receives the appointment and the IP address is a business associate and an agreement is required12. HHS adds that a banner asking users to accept cookies is not a valid HIPAA authorization12. The FTC's complaint against Cerebral charged that it provided sensitive information of nearly 3.2 million consumers to third parties by using or integrating tracking tools176, and the order was announced as proposed176.

The review reply is a medical mistake too. Manasa Health Center, a New Jersey psychiatric provider, paid $30,000 and agreed to a corrective action plan to resolve potential violations over a response to a patient's negative online review126. The agreement says it impermissibly disclosed the protected health information of four patients in response to negative reviews on Google Reviews124. A third is an AI-written patient message about clinical information with no disclaimer and no way to reach a human, sent without a licensed provider's review. That is the situation California's section 1339.75 addresses for physician offices and clinics149.

Law

Paying for a recommendation is the first. ABA Model Rule 7.2(b) bars a lawyer from compensating or giving anything of value to a person for recommending the lawyer's services, with listed exceptions68. Comment [5] to Rule 7.2 says that to comply with Rule 7.1 a lawyer must not pay a lead generator that states, implies or creates a reasonable impression that it is recommending the lawyer, is making the referral without payment from the lawyer, or has analyzed the person's legal problems in choosing the lawyer205.

A reply to a negative review that discloses client information is the second. Formal Opinion 496 says a negative online review alone does not meet the requirements for disclosing client information in self-defense under Model Rule 1.6(b)(5)208. It adds that as a best practice lawyers should consider not responding208. The third is an intake chatbot that may create a prospective-client relationship the lawyer does not know about225. The fourth is unverified AI output in a filing, which is outside the marketing system and inside the firm's AI policy. In Mata v. Avianca two lawyers and their firm were ordered to pay a $5,000 penalty for filing non-existent cases generated by ChatGPT234. In Wadsworth v. Walmart motions in limine cited nine cases, of which eight did not exist, and the court revoked one attorney's pro hac vice admission and fined him $3,000235.

Accounting

The mistake most specific to a preparer is sending return information to a tool without the written consent the statute or regulations require273. A related regulation defines "tax return preparer" to include a person providing auxiliary services in connection with preparing returns, including one who develops software used to prepare or file them271. Under section 7216, a preparer who knowingly or recklessly discloses or uses information furnished for a return, other than to prepare the return, is guilty of a misdemeanor269. The punishment is a fine of not more than $1,000 ($100,000 for a disclosure or use to which section 6713(b) applies), imprisonment of up to 1 year, or both269. Section 6713 adds a civil penalty of $250 for each disclosure or use, capped at $10,000 for a calendar year270. The AICPA's route for third-party service providers is a confidentiality contract or specific client consent8, and the IRS alert says outsourced or third-party AI tools should be vetted286.

Two further mistakes follow from the advertising rules. The AICPA interpretation treats promotion that creates false or unjustified expectations of favorable results as deceptive8. It also tells a member to determine that a third party's promotional efforts comply, because members must not do through others what they are prohibited from doing themselves8. Separately, a firm with no written security plan has a gap the IRS names: tax and accounting professionals are considered financial institutions regardless of size, and the Safeguards Rule requires a written, accessible plan276.

In every profession

Four mistakes cut across all four professions. The first is recording calls without the consent the state requires. Washington requires the consent of all participants179, and California fines recording a confidential communication without the consent of all parties up to $2,500 per violation, with imprisonment also possible149. An answering service that records every call makes recording the default.

The second is buying reviews or incentives conditioned on sentiment: in warning letters of December 22, 2025, the FTC said violations of its review rule can bring civil penalties of up to $53,088 per violation139, and the FTC has said those amounts remain unchanged during 2026340. The third is letting an assistant say something no one approved, which is why the first row of Table 4 exists. The fourth is treating a vendor's badge as compliance: HHS states that it does not endorse or recognize private organizations' "certifications" regarding the Security Rule, and that they do not absolve covered entities of their legal obligations151. Chapter 14 turns to what this paper has not shown.

Chapter 14

Limits, objections and a research agenda

What this paper has not shown

This paper makes an argument and presents data of our own. It has not shown that technical marketing raises appointments, matters or revenue for any practice, and it compares no practices that built such a system with practices that did not. The evidence supports its premises: the three assistants we studied rarely agree on whom to name1; the professions' rules limit what a practice may say and do (Chapters 6 to 10); and measuring the return on marketing is hard even for large advertisers, since in field experiments with large retailers and brokerages the median confidence interval on the return on digital advertising was more than 100 percentage points wide103.

Counts of practice size come from BLS establishment data, which count locations and not firms. In the first quarter of 2024, 141,518 of 187,335 offices of lawyers, 44,800 of 62,828 offices of certified public accountants, 54,225 of 135,233 offices of dentists and 138,197 of 237,573 offices of physicians had fewer than 5 employees15,16,18,19,20, which is 75.5%, 71.3%, 40.1% and 58.2% (141,518 / 187,335 for the offices of lawyers, and the same division for the other three). A firm-level size distribution that could be opened is NF; the nearest evidence is these establishment counts, so a firm with several offices counts several times. Descriptions of rules are as of the dates our sources were read, several of them 2026-11-10.

The recommendation study

Chapter 4 states the limits of the study in full. Each question was asked once, on 2026-11-10. Names were matched across assistants by spelling, up to eight names were read per answer, and answers collected through an API are a baseline, not what every signed-in user sees1. Each assistant was represented by one model1, and every question had one wording, which asks for a good business and sets no standard1. The study has no row for physicians, its three-city rows rest on three questions each, and it did not query Google's AI Overviews or AI Mode1. Names and cited sites were counted separately1, so a name cannot be traced to the page behind it.

Three features of the design, asking each question once, matching names by spelling and reading at most eight names, tend to lower measured agreement, so we think the 3.5% for the 160 questions asked in every city1 is closer to a floor than to a central estimate. We cannot say how far. The study does not show that a named business is good or that a cited page caused a name to appear, and it says nothing about whether any answer led to an appointment or a matter.

The search data

Chapter 5 states the limits of the search data. The phrases are ones we chose, and an owner, a practice manager or a vendor would type them, not a patient or client. Volume shows interest, not purchase. One month of data shows level, not direction. Cost per click is what advertisers pay, not what a client is worth. Volumes sit on rounded bands, and close variants share a figure.

The leading AI phrase outruns the leading marketing phrase in legal and accounting, where "legal ai" averages 18,100 searches a month against 3,600 for "law firm seo", and "ai accounting" 6,600 against 1,300 for "cpa marketing"27. That is about five times in each (18,100 / 3,600 and 6,600 / 1,300), and the accounting ratio is a floor, because "cpa marketing" is also searched as an advertising term27. The pattern holds only weakly in medical, where the comparison rests on "hipaa compliant ai", a phrase that names a privacy rule and not a profession. In dentistry it does not hold: "dental ai" averages 880 searches a month against 1,300 for "dental marketing"27.

Objections to technical marketing, answered from the evidence

AI answers are still a small part of how professionals are chosen

The strongest form of this objection is that our thesis describes a shift the evidence has not reached. We concede much of it: what share of patients or clients use an assistant to choose a dentist, a doctor, a lawyer or an accountant is NF; the nearest evidence, below, is indirect, and some of it points the other way.

The surveys of use are about people in general. A Pew Research Center survey of 5,119 US adults in February 2026 found that 49% ever use AI chatbots, up from 33% in 2024, and 20% ever use them to get medical advice34. In a KFF poll fielded in February and March 2026, about a third (32%) of 1,343 adults said they had turned to AI chatbots in the past year for health information341. Pew's June 2026 survey found 34% of Americans use chatbots for at least one of eight health reasons, and 15% cited helping decide whether to see a doctor342. Bain's September 2025 survey of 1,500 consumers found that 56% mostly or always use search engines as their default search tool and 16% chatbots46. That consultancy survey points to a channel still smaller than search.

A consumer survey by the local search vendor BrightLocal† reported that 6% of consumers used ChatGPT or other AI tools to find local reviews in 2025, down from 9% in 202435. Its 2026 survey of 1,002 adults reports that use of ChatGPT and other generative AI tools for local recommendations rose from 6% in 2025 to 45%36. The two years are separate panels of about 1,000, the question wording is not printed, and a jump of more than seven times may partly be a change of wording, so we do not read the 45% as a share of choices made. A Clio† survey reported in 2019 found that 59% of clients sought a referral to an attorney and 57% searched for one on their own91. A 2023 BrightLocal† survey found 81% of consumers rating reviews as important or very important in the financial and legal sector86. These are vendor-reported surveys with samples we cannot check, and the 2019 figure predates the assistants studied here. They show referral, search and reviews as established channels and say little about assistants.

On the other side, Pew Research Center found AI summaries on 18% of participants' Google searches30, and ChatGPT's consumer product had been adopted by around 10% of the world's adult population by July 202533. Neither isolates the choice of a professional.

The share of choices of a practice that an assistant now influences is ~1% to 15% (from Conductor†'s AI referral share and a Pew survey; the low end likely low, the high end likely high). The low end is the 1.08% of website visits that came from AI referrals across 10 industries in data from Conductor† (SEO platform); its health care figure, 0.64%, is lower still42,43. The low end errs low because an answer that nobody clicks leaves no trace in traffic318. The high end is the 15% of Americans who use chatbots to help decide whether to see a doctor342. The high end errs high because it is a stated use by adults in general for one health decision and not the choice of a practice.

The paper's practical argument does not depend on the assistant channel being large today. The four premises of Chapter 1 hold however a patient first finds the practice, and referral, search and reviews need the same system: accurate listings, a site that answers the questions people ask and books, a counted intake and a design that respects the profession's rules. If the assistant channel stays small, we think preparing for it mostly costs work the practice needed anyway, and if it grows, the practice is less far behind. We should be held to the word "moving" in Chapter 1's thesis, and we would withdraw it if repeated measurement showed no growth in the use of assistants to choose professionals.

The assistants will converge, and the channel will collapse into one list

The objection has two versions. In the first, the assistants come to read the same sources and rank the same practices, so that the finding in Chapter 4 that each assistant is its own channel describes an early period. In the second, one assistant takes most of the market and the others stop mattering.

The evidence bears on the second version more than the first. Similarweb's worldwide traffic panel, whose method is not described, puts ChatGPT's share of generative AI website visits at about 76% in June 2025 and roughly 53% by May 2026, with Gemini rising from under 9% to around 27% to 28%343, which points away from concentration in one assistant. It does not decide the first version, because we asked each question once and have no second measurement to compare against1. The nearest repeat-asking evidence is SparkToro† and Gumshoe.ai†'s 2026 study, by volunteers who pasted answers by hand to prompts that were not about these professions, in which there was less than a 1 in 100 chance that ChatGPT or Google's AI, asked a prompt 100 times, would give the same list of brands in any two responses115. That points to variation inside one assistant and not to convergence.

In our study, 79.7% of the businesses named were named by one assistant only, and 88% of cited sites were cited by only one1. Gemini cited Reddit in 42% of its answers, ChatGPT in 2% and Perplexity in 0%1. Citations are also concentrated: the 25 most-cited sites drew 29% of all citations and were directories, review platforms, list sites, publishers and a forum1. If the assistants converge, we expect one point of convergence to be those same kinds of sites, which would make accurate and consistent facts there matter more.

If sources converge, the advice to check all three assistants becomes less useful, because one check would say more about the others, but the work the advice implies would not change: consistent facts, signals a reader can verify, and pages that answer specific questions. If a single assistant dominates, the economics of the list return, in which position carries weight: one study of ranking effects in the list era found links at the top of the screen especially likely to be clicked on mobile phones28. We concede that "each assistant is its own channel" is a reading of one run.

Platforms will sell the placement

Our study cannot say whether any assistant sells placement in its answers: it collected answers through APIs, not as a signed-in user sees them1, and did not classify cited sites as paid or unpaid. A maker's statement on how Gemini, Claude or Perplexity choose local businesses is NF; the nearest evidence is OpenAI's help page, which says that ChatGPT ranks search results using multiple factors and that placement is not guaranteed, and does not list the factors344. Search has an advertising history, and advertisers already pay high prices for professional phrases: in our data "legal answering service" averages $214.51 a click and "attorney marketing" $134.6627. If answers carried paid placement, a small practice would compete in an auction against larger bidders. We think intake and the count of outcomes (Chapters 11 and 12), what happens after the click, would then matter most, because the practice would be paying for each contact.

Paid placement would also meet the professions' rules. ABA Model Rule 7.2(b) bars a lawyer from compensating a person for recommending the lawyer's services, with listed exceptions that include the reasonable cost of permitted advertisements68. Comment [5] says a lawyer must not pay a lead generator that states, implies or creates a reasonable impression that it is recommending the lawyer205. How a state's rules would apply to a paid slot in an assistant's answer is a question for a lawyer's own counsel, and this paper cites no source that answers it. We do not know whether placement will be sold, and a paper written after it is might strike a different balance between earned and paid visibility.

Small practices cannot afford systems

The objection has two forms: the cost of the system and the cost of knowing whether it worked. On the first, cost pressure on practices is real. The ADA Health Policy Institute says expenses per dentist are rising more than revenue per dentist, which is driving the downward trend in general practice dentists' net incomes13. In an MGMA poll of 251 applicable responses in June 2026, 84% of medical groups reported higher year-to-date operating costs than at the same point in 2025, with an average increase of about 11% among the respondents156.

Our own prices are one reference point: plans at $999, $1,999 and $2,999 a month per location, or $9,990, $19,990 and $29,990 for a 12-month term paid in full (two months free), with a $3,000 setup fee on the top plan and a one-time onboarding fee of $499 and $999 on the two lower plans, all three waived on a 12-month term56, an AI voice agent at $299 a month with 250 call minutes included and $0.39 a minute after that56, and HIPAA-ready programs from $1,500 setup plus $699 a month per location after a consultation57, which are a compliance layer added to a plan and do not include the voice agent57.

A marketing plan alone needs no program for a health care practice57. We sell these, and whether they are affordable depends on what a new patient or client is worth to a practice. Chapter 11, Table 2 gives a cited first-year value for one client in each profession, and the cases below turn each price into the number of clients needed to cover it. The included minutes are about 62 four-minute calls (250 / 4), and the cases stay inside them, so the agent is priced at $299 in every stack that includes it. Chapter 13 sets out the stacks.

Four sets of figures (calc., Chapter 12) show how far each purchase is from covering itself. They use cited inputs, count first-year value unless a tenure is stated, and are not forecasts. For a dental office, the plan alone at $999 a month is covered by 1.1 new patients a month, which is 3.8% more than the 30 that Henry Schein One† reports as the average49. The full stack at $2,122 a month in its first year is 2.6% of one dentist's billings and 1.3% of two dentists', at $965,660 a year per general dentist21, and it is covered by 2.4 new patients a month if the practice keeps all of a patient's first-year spending of $887118, or 10.7 if it keeps only the 22.3% that is the average dentist's net income share of billings21.

For a primary care physician, the plan alone needs 3.3 new patients a month (calc.), 15.6% more than the 21 to 22 new-patient visits a month one physician has now and 3.9% more for four physicians, or 1.1 a month if patients stay three years.

For a law firm, the lowest plan and the agent at $1,298 a month are covered by 5.3 collected hours a month, 4.5 DUI matters a year or 2.3 injury cases a year. The average lawyer in Clio†'s data bills under four hours of an eight-hour day52.

For an accounting firm that sells advisory services, 0.9 new advisory clients a year cover the lowest plan and the agent, which is the high end because the benchmark is 206 self-selected practices113, while 2.6 to 5.2 bookkeeping clients a year cover it at the low end of the most common vendor band and a firm that sells only base returns at the NATP average fee needs about 66 clients a year60.

The tightest case is a solo primary care physician on the full health care stack. The stack is 6.3% to 6.7% of the payments for that physician's 171 to 180 kept visits a month (calc., Chapter 12). Covering it takes 11.4 kept visits a month, which is 22.4% to 27.4% of the 42 to 51 visits missed, or 7.1 new patients a month on a first-year count, which is 33.2% more than the 21 to 22 new-patient visits now. The agent's own fee is $299 a month, its published price, the cost of 8.6 front-desk hours at $34.59, about 25 minutes a working day (calc., Chapter 12). The program and the reminders add-on, $973 a month together in the first year, need 2.4 points off the no-show rate for that physician, near the top of the 0.5 to 3 points that two studies found162,163, and 0.6 points for four physicians (calc., Chapter 12).

At the median primary care visit payment of $10761, the same physician needs 9.1 kept visits for the program and reminders and 19.8 for the whole stack. For this office the case rests on running more than one tool on the compliance layer: the agent's answered calls, reminders, recall, waitlist backfill and intake. It is not typical of medicine: a four-physician practice covers the same stack with 5.6% to 6.9% of its missed visits, and the stack is about 1.6% to 1.7% of its payments (calc.).

The system loses money in one corner: a practice that is not growing and whose intake and scheduling are already sufficiently optimized. It would want no new patients or clients, lose few inquiries at first contact, miss few visits, have little staff time to release and lose few of the people it serves. The cited base rates put a typical practice far from that corner, though each counts a different event for a different group and none is a rate for the practice in question.

The mean no-show rate was 18.8% across ten VA clinics162. Patient Prism† (call analysis) reports that at dental practices the caller hung up before reaching an agent on 31 of every 100 patient calls24. Invoca† (call tracking) reports that a person answered 54% of health care calls, so 46% were not (calc.)26. In Clio†'s 2024 test, 40% of 500 law firms answered the call23. And nearly two in five dentists told the ADA Health Policy Institute that they could have treated more patients102, so for them a full schedule is not the limit.

A cost that is easy to miss appears in our check of tools. Of 28 software tools that small practices commonly use, 15 sign a business associate agreement only on certain plans, and 6 of those 15 do not publish the price of the plan that qualifies54. A practice that compares headline prices may undercount the cost of a set-up that can sit under such an agreement.

The second form concerns measurement. Lewis and Rao found that informative advertising experiments can easily require more than 10 million person-weeks, "making experiments costly and potentially infeasible for many firms"103. Their finding concerns digital advertising experiments by large advertisers, not a practice's telephone, but it states the difficulty. A small practice cannot run such experiments or expect precise estimates of what a channel returned. The paper's claim is smaller: a practice can count its calls, forms and bookings by source and learn what the counts allow, treating them as counts and not as estimates of causal return. We concede that for the smallest practices some of this may not pay; the tightest case and the corner above show where. Which practices fall short is a question for their own counts, and Chapter 11 says what to count.

The rules will change

They will, and several changes are already on record. A proposed rule to amend the HIPAA Security Rule was published on January 6, 2025169. It was not a final rule on the federal Unified Agenda entry read on 2026-11-10, which lists final action as July 2027170. In the proposal, HHS stated that electronic protected health information in AI training data, prediction models and algorithm data maintained by a regulated entity for covered functions is protected by the HIPAA Rules169.

On June 20, 2024, the U.S. District Court for the Northern District of Texas vacated part of HHS's online tracking guidance12, only to the extent it said HIPAA obligations are triggered where an online technology connects an individual's IP address with a visit to an unauthenticated public webpage addressing specific health conditions or healthcare providers12. The court added that its vacatur was not intended to limit the legal operability of other guidance in the document172.

Other changes come from states and agencies. The State Bar of California issued its generative AI guidance in November 2023, and a 2026 version, approved on May 14, 2026, replaces it226. In 2026 Colorado enacted SB 26-189, which repeals and reenacts, with amendments, part 17 of article 1 of title 6 of the Colorado Revised Statutes under the heading Automated Decision-Making Technology in Consequential Decisions303, and except for sections that take effect upon passage, the act takes effect on January 1, 2027303. On December 22, 2025, the FTC reopened and set aside its 2024 final consent order against Rytr, saying the complaint failed to meet the legal requirements of the FTC Act and that the order unduly burdens AI innovation345. These are five different kinds of change, and a practice cannot treat them alike.

We respond structurally and forecast no rule. Chapter 10 reduces the four rulebooks to requirements that recur: say only what is true and substantiated, know where every piece of personal information goes, have a contract with every party that touches it, tell people when they are dealing with a machine, and keep a human responsible. A design built on those survives most changes in the details of a rule. The details that do change, such as a disclosure wording, the terms of a vendor agreement or a tracking setting, are kept as named, dated settings that a responsible person reviews. We concede that a design that meets today's rules may need rework, and that the cost of rework falls on a small firm. Nothing in this paper makes a practice compliant.

AI recommendations may be wrong or biased

Our study cannot answer this. It did not check whether any named business exists, is open, accepts new patients or clients, or is any good, and low agreement among assistants does not show that any of them is wrong. The wider evidence on accuracy comes from studies of other tasks. Table 1 lists what each study measured.

StudyTask measuredResultSource
Liu, Zhang and Liang (2023)Whether sentences in answers from commercial generative search engines were supported by the citations attachedOn average 51.5% of generated sentences were fully supported by citations, and 74.5% of citations supported their associated sentence158
Dahl and co-authors (2024)Specific, verifiable questions about random federal court cases, put to general-purpose language modelsLegal hallucinations occurred 58% of the time with ChatGPT 4 and 88% with Llama 2230
Magesh and co-authors (2024)Evaluation of AI legal research tools from two providersEach tool hallucinated between 17% and 33% of the time231
Eulerich and co-authors (2024)The CPA, CMA, CIA and EA examsChatGPT 3.5 averaged 53.1% and passed none; with ChatGPT 4, few-shot training and reasoning and acting, the average was 85.1% and it passed all282
Kung and co-authors (2023)Questions from the three USMLE examsChatGPT performed at or near the 60% passing threshold on all three332
Ayers and co-authors (2023)195 public patient questions from an online forum, answered by verified physicians and by a chatbotLicensed health professionals preferred the chatbot's response in 78.6% of 585 evaluations142
Dell'Acqua and co-authors (2023)Consulting tasks inside and outside what the tool could doInside: 25.1% faster and over 40% higher quality. Outside: 19 percentage points less likely to produce correct solutions263

Table 1. What seven studies of AI accuracy and verifiability measured, and on what task.

No row of Table 1 asks an assistant to name a practitioner. The Liu study measures whether cited pages support a generated sentence, which is verifiability and not truth, and we do not know that its questions included requests for local professionals. The Eulerich row shows that a figure belongs to a setup, a task and a date, and the Dell'Acqua row that the same tool can help on one kind of task and hurt on another. We therefore carry no result from one task to another.

A different error is the practice's own. In Moffatt v. Air Canada, a tribunal found that the airline had not taken reasonable care to ensure its chatbot was accurate191. The case concerns an airline, and we draw no legal conclusion from it for a US practice, but it shows the question a tribunal asked about a business's own assistant. Chapter 12 treats that risk at the front desk.

This paper does not measure bias. Our study counts names. It does not test whether practices that are larger, more heavily reviewed or listed on more directories are favored, and it cannot, because it did not link names to pages. An assistant may inherit the weaknesses of the sources it cites. Physician ratings show how a source can carry little information: in an analysis of online ratings on RateMDs, the mean rating was 3.93 on a scale of 1 to 5, and 45.80% of physicians received a 579. If an assistant gave weight to ratings that cluster near the top, ratings would separate practices poorly. That is an inference: we do not know how any assistant weighs ratings.

The paper's argument does not require the assistants to be right. It requires that people use them, that they read sources, and that a practice's facts and verifiable signals be accurate where they read. If an assistant misstates a practice's hours or services, the remedy is the one for any wrong listing: correct the source, and know about the error, which is the purpose of the monthly check in Chapter 11. Whether a practice may describe an assistant's recommendation in its own advertising is a question for the advertising rules in Chapters 6 to 9 and for its counsel.

The argument comes from a seller

This is true. Technical Marketing sells marketing plans56, an AI voice agent56, managed marketing56 and HIPAA-ready programs57, and a seller has reason to find a problem its product solves. Chapter 1 declares the interest. The data are tabulated so that a reader can read them differently, the limits are stated, and parts of the argument run against a simple sales story: search volume is interest and not purchase, the study's figures give no basis for promising a place in any answer, and a small practice cannot measure the return on marketing precisely.

What would count against the thesis can be said in advance. Repeated panels could show that the assistants converge on one list, or that work on a practice's sources does not change what they name. Surveys of patients and clients could show that almost none use an assistant to choose a professional. Counted outcomes could show that practices that build such a system do no better on contacts and bookings than practices that do not.

What we do not know

What the second round of research found

Table 2 lists the findings that bear on the paper's open questions, with what each does not settle.

QuestionWhat was foundWhat it does not settle
How many people use an assistant49% of 5,119 US adults ever use AI chatbots, 20% for medical advice34; about a third of 1,343 adults used one for health information in the past year341Choosing a practice; stated use is not a call
How much traffic AI sends to websitesAbout 1 visit in 100 across 10 industries, 0.64% for health care, in Conductor†'s data42,43; under 1% on publisher sites in Microsoft Clarity†'s45An answer nobody clicks, or a practice's own site
Whether answers repeatLess than a 1 in 100 chance of the same list of brands in two responses in a volunteer study115Local professionals; an assistant's own variation
How often phones go unansweredDental calls hung up on before an agent, 31 in 100†24; health care calls answered by a person, 54%†26; VA calls abandoned, 8.3% in September 2016159; law firms that answered a test call, 40%23A rate for any practice; each counts a different event
What reminders do to no-shows0.5 points in a before-and-after comparison at one VA center, 3 points in one randomized trial162,163An AI voice reminder; a private office
What an AI voice agent achieves86.4% of outbound pre-procedure calls completed in the first phase of a Mount Sinai study47An inbound line; a control group
What one client is worthA cited first-year value for one client in each profession (Chapter 11, Table 2)A lifetime value; any practice's own figure

Table 2. What the second research round found on the paper's open questions, each with its limit.

What was searched for and not found

Table 3 lists what is NF, with the nearest evidence. Where a quantity could be estimated from neighbors, the estimate is in the chapter that needs it, marked ~.

Searched forNearest evidence
AI referral share or conversion for a dental, medical, legal or accounting practice's siteTraffic shares in other industries from Conductor† and Microsoft Clarity† (Chapter 11 estimates the share)
A repeat-asking study of AI recommendations for local businesses or practicesThe volunteer study above, which was not about these professions
Any maker statement on how Gemini, Claude or Perplexity choose local businessesOpenAI's help page for ChatGPT, which lists no factors344
An estimate of how far last-click or platform-reported attribution misstates results as suchThe gaps between observational and randomized estimates (Chapter 11 estimates it)
A randomized or field study of an AI agent answering inbound calls in these settings, or a peer-reviewed containment rate for inbound callsThe outbound Mount Sinai study above
Missed-call, call-volume or lead-conversion figures for accounting firms; call volume and inquiry-to-client conversion for law firmsLaw firm answer rates and health care call data (Chapter 12 estimates it for accounting)
The share of calls arriving after hours in any of the four professions, from a source that could be openedThe 89.4% in business hours from CallMyDoc† (answering service)160
A peer-reviewed or ADA overall no-show rate for US dental practices; a peer-reviewed US survey of how patients choose a dentistHenry Schein One†'s no-show averages and a dental school's attendance (Chapter 12 estimates it)
An independent survey of patients or clients about AI receptionists or voice agentsSurveys of discomfort with clinical AI and of expectations about scheduling (Chapter 12)
The effect of response time on booking in any of the four professionsA study of US companies in general, as reported (Chapter 12)
The share of new patients or clients by source taken from practice records, for any professionSurveys of opinion and stated behavior; one survey of 448 dental practices93
The firm-level size distribution of the four professionsEstablishment counts that count locations and not firms (above)

Table 3. What was searched for and not found, and the nearest evidence.

What only a practice's own records can settle

Some questions have no study that could answer them for a given practice, because the answer is in its own records. Table 4 names the record that settles each.

Question for one practiceThe record that settles it
How many calls it misses, hangs up on and receives after hoursIts telephone log (Chapters 11 and 12)
Whether a missed caller rings back or goes to a competitorIts call log, matched to bookings
What a new patient or client is worth to it, and what it keepsIts own payments, fees and margin (Chapter 11)
Where its new patients or clients come from, including an assistantIts source question at booking or intake (Chapter 11)
Its no-show rate, and how far reminders move itIts scheduling system, before and after a change
How many staff hours an agent releasesIts phone log and its staff's time
Whether and where the assistants name it, and what they sayIts monthly dated check (Chapter 11)
Whether its schedule is full, and what more demand would changeIts own calendar and waiting list

Table 4. Questions that only a practice's own records can settle.

Three open questions matter most: how assistants convert into appointments, how stable their answers are, and how they weigh reviews against credentials. Between an assistant naming a practice and an appointment lie several steps: the person reads or hears the answer, chooses, visits the practice's site or calls, reaches someone and books. A measure of any step after the answer, for a practice that an assistant names, is NF; the nearest evidence is the Pew click figures in Chapter 130, which are click rates for visits with and without a Google AI summary, taken across the dataset's searches, and not for any named practice. Our data contain answers and no people, so they cannot recover the steps. Without a conversion rate, what being named is worth is solved for and not supposed: the cases in Chapter 12 and in the objection about cost above give the number of bookings, kept visits or clients that a purchase has to recover, beside the base rates that are cited.

On stability, we expect assistants to change models and sources, and the model names in Chapter 4 to date. Two kinds of instability matter: variation inside one assistant from one run to the next, and drift over weeks as models and indexes change. We asked each question once and measured neither, and the volunteer study above suggests the first is large. Because the design tends to lower measured agreement, a repeated panel could find agreement higher than we report. It could also find that a practice's presence in an answer is volatile from week to week, which would change what a monthly check can tell it (Chapter 11).

Reviews and credentials are the third gap. In other markets reviews move demand: a one-star rise in a Yelp rating raised revenue by 5 to 9% for independent restaurants, and had no effect on chain restaurants346, and an extra half-star made restaurants sell out peak reservations 19 percentage points more often, with larger impacts when other information was scarcer77. Neither concerns an assistant or a professional. We think the second result fits a credence good, where other information is scarce by definition. What an assistant does with credentials such as licensure, board certification or a specialist designation is equally unmeasured. Our study described only the 25 most-cited sites by type1, so it cannot say whether credential registers are read at all.

A research agenda

The studies below are proposals, and none has been done.

Study 1: a repeated panel of assistant answers

A fixed set of questions would be put to each assistant many times a day for several months, through APIs and the consumer apps, signed in and signed out. The set would use several wordings for each kind of practice, including physicians, and add the specialty, insurance, emergency and language questions that patients and clients ask. Cities would be drawn at random. Every answer and citation would be kept, names matched by address and telephone number, and the questions and raw answers published so that others can repeat the work. The panel would measure agreement within and across assistants, drift over time and the concentration of cited sources, by profession.

Study 2: conversion with cooperating practices

Practices would agree to record the source of every contact, through call tracking and a source question at intake, with the assistant as a category, and to follow contacts to bookings over many months. Because one practice's counts are too small to read, results would be pooled across many practices and the analysis registered in advance (Chapter 11 explains the small-numbers problem). The design would collect counts and sources, not clinical, tax or legal details, and each profession's rules on information would be settled first. Self-reported source is imperfect, so it would be compared with tracked source wherever the rules allow tracking.

Study 3: an accuracy audit of professional recommendations

A sample of answers from the panel would be checked name by name. For each named business the audit would verify that it exists, that its license is current in the relevant register, and that its address, hours, services and acceptance of new patients or clients are as stated. It would also check, in the manner of Liu and co-authors' verifiability audit, whether the page cited for a statement supports it. Qualified reviewers in each profession would judge whether the description of a service is accurate. The audit would measure error rates with intervals, by assistant and by profession, and would not judge the quality of care or advice.

Study 4: how assistants weigh signals

An observational stage would relate whether and where a practice is named to its features: review count and rating, years in practice, specialist designation, the content of its site and its presence on directories, across many practices in many cities. A controlled stage would then change one feature at a time on the sites of cooperating practices, on a staggered schedule, and observe what changes in the answers. The study must not create fictitious practices, listings or reviews. A result from restaurants or books would not be carried over to professionals, or one assistant's to another's.

Study 5: how patients and clients ask

A survey of recent new patients and clients would ask how they chose their practice and whether they used an assistant, with the population and response rate reported. It would be paired with query data from cooperating practices' own search tools and call logs. A survey of owners would ask what they believe they are buying when they search for an AI receptionist or a legal answering service, which is the evidence Chapter 5 said would narrow its reading of the gap between marketing and AI phrases.

Study 6: front-desk audits and field evaluation

Secret-shopper calls and messages to practices in all four professions would measure whether contact is answered, how fast, and what the caller receives. They would extend the designs of the audits in Chapter 12, Table 1 (the calls to physician offices to schedule appointments and Clio†'s secret-shopper study of law firms22,23) to accounting firms and to private dental and physician offices drawn at random. Several of the audits so far were drawn from marketing lists, from a vendor's customers or from one specialty, and a random draw would turn them into rates. A field evaluation with cooperating practices would compare lines answered by staff, by a service and by an AI voice agent that identifies itself, on answered contacts, bookings and errors. Both designs would be cleared with counsel for recording, consent and AI voice rules before any call is placed (Chapter 10).

Conclusion: what the evidence supports

For 160 local questions asked in 16 cities, three assistants named 2,414 businesses, and 84 of them, 3.5%, were named by all three1. That is one asking of each question, and it shows disagreement, not error. It supports a narrow claim: a practice cannot infer its standing in one assistant from another's answer, and cannot treat being found as a single position on a single list.

Its thesis has three parts, and the evidence reaches them unevenly. That professional services are largely credence goods sold under speech rules by small firms rests on classic economics, on the professions' own codes, and on counts of practice size that are establishment counts for all four professions and not counts of firms. That the way they are found is moving toward written answers rests on indirect evidence: ~1% to 15% of choices, from Conductor†'s traffic share and a Pew survey42,342. That technical marketing is the discipline that fits the change is an argument. It follows from the premises, it has not been tested against an alternative, and we sell what it recommends.

We think what the evidence does support is a change in what a practice should be able to show. The assistants read different sources1, and the sites they cite most are directories, review platforms and list sites1. A practice's facts therefore have to be accurate and consistent where those readers look, its pages have to answer the questions people ask, its first contact has to work, its tools have to sit within the rules of its profession, and it has to count what happens. Each of these is ordinary work. The claim is that the work belongs together and should be measured as one system.

Until studies like those proposed here are done, a practice should read anything said about assistants, ours included, as a description of one date.

References

Before the paper cited a source, the page was fetched and the passage we rely on was matched against it, except where a reference is marked otherwise. The small numbers in the text point here.

  1. Technical Marketing. Who AI recommends locally: a study of local answers. Technical Marketing, 2026-11. technicalmarketing.ai/reports/ai-local-recommendations/ Our own data.
  2. Darby,Karni. Free Competition and the Optimal Amount of Fraud. University of Chicago Press, 1973. doi.org/10.1086/466756 Held through its record or abstract; passage not matched.
  3. Ellen Day and Hiram C. Barksdale Jr.. How firms select professional services. Elsevier / Industrial Marketing Management, 1992. doi.org/10.1016/0019-8501(92)90002-b Held through its record or abstract; passage not matched.
  4. Hanauer DA, Zheng K, Singer DC, Gebremariam A, Davis MM. Public Awareness, Perception, and Use of Online Physician Rating Sites. JAMA 2014;311(7):734-735 (author-hosted PDF), 2012. hai.ics.uci.edu/papers/jld130043.pdf
  5. U.S. Supreme Court (Blackmun, J.). Bates v. State Bar of Arizona, 433 U.S. 350. Legal Information Institute, Cornell Law School (text of the opinion), 1977-06-27. law.cornell.edu/supremecourt/text/433/350
  6. American Bar Association. Rule 7.1: Communications Concerning a Lawyer's Services (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  7. American Dental Association. ADA Principles of Ethics and Code of Professional Conduct: Veracity (section 5.F Advertising). American Dental Association, 2026-11-10. ada.org/about/principles/code-of-ethics/veracity
  8. American Institute of CPAs. AICPA Code of Professional Conduct, 1.600.001 Advertising and Other Forms of Solicitation Rule. AICPA (copy hosted by the Minnesota Board of Accountancy; text bears © 2026 AICPA), 2026. mn.gov/boa/assets/AICPA%20Code%20of%20Professional%20Conduct_tcm1233-7
  9. AMA Council on Ethical and Judicial Affairs. AMA Code of Medical Ethics, Opinion 9.6.1 Advertising & Publicity. American Medical Association, 2026-11-10. code-medical-ethics.ama-assn.org/ethics-opinions/advertising-publicity
  10. U.S. Department of Health and Human Services. 45 CFR 164.501 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.501
  11. HHS Office for Civil Rights. Marketing (HIPAA Privacy Rule guidance). U.S. Department of Health and Human Services, 2026-11-10. hhs.gov/hipaa/for-professionals/privacy/guidance/marketing/index.html
  12. HHS Office for Civil Rights. Use of Online Tracking Technologies by HIPAA Covered Entities and Business Associates. U.S. Department of Health and Human Services, 2024-06-26. hhs.gov/hipaa/for-professionals/privacy/guidance/hipaa-online-tracking
  13. ADA Health Policy Institute. The U.S. Dentist Workforce - 2026 update. American Dental Association, 2025. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  14. American Medical Association (Policy Research Perspectives). Physician Practice Characteristics in 2024: Private Practices Account for Less Than Half of Physicians in Most Specialties. American Medical Association, 2024. ama-assn.org/system/files/2024-prp-pp-characteristics.pdf
  15. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 54111 Offices of lawyers (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/54111.csv
  16. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter by establishment size class 1, all industries and areas. U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/size/1.csv
  17. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter by establishment size class 2, all industries and areas. U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/size/2.csv
  18. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 541211 Offices of certified public accountants (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/541211.csv
  19. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 6212 Offices of dentists (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/6212.csv
  20. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 6211 Offices of physicians (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/6211.csv
  21. ADA Health Policy Institute. Dental practice research: Dental practice FAQs. American Dental Association, 2023. ada.org/resources/research/health-policy-institute/dental-practice-res Held through its record or abstract; passage not matched.
  22. AMN Healthcare (formerly Merritt Hawkins). 2025 Survey of Physician Appointment Wait Times and Medicare and Medicaid Acceptance Rates. AMN Healthcare, 2025. amnhealthcare.com/siteassets/amn-insights/physician/ps-2025-physician-
  23. Clio. Highlights from the Legal Trends Report: the Legal Industry in 2024. Clio (a legal practice-management software vendor), 2024. clio.com/blog/highlights-from-2024-legal-trends-report/ Secondary source.
  24. Patient Prism. The Dental Patient Access Report. Patient Prism, 2025. patientprism.com/report/dental-patient-access-report/ Secondary source.
  25. Jay Geier. A Goldmine of New Patient Opportunity. Inside Dentistry, Dentalaegis (Volume 11, Issue 2, September 2015, pp. 4-5), 2014-2015. dentalaegis.com/id/special-issues/2015/09/a-goldmine-of-new-patient-op Secondary source.
  26. Invoca. The Healthcare Lead Conversion Benchmarks Report 2026. Invoca, 2026. invoca.com/reports/the-invoca-healthcare-lead-conversion-benchmarks-re Secondary source.
  27. Technical Marketing. What service businesses search for: 2026 demand data. Technical Marketing, 2026-11. technicalmarketing.ai/reports/service-business-search-demand/ Our own data.
  28. Anindya Ghose, Avi Goldfarb, and Sang Pil Han. How Is the Mobile Internet Different? Search Costs and Local Activities. Information Systems Research / INFORMS, 2013. doi.org/10.1287/isre.1120.0453
  29. Brian Dean. Here's What We Learned About Organic Click-Through Rate. Backlinko, 2023. backlinko.com/google-ctr-stats Secondary source.
  30. Pew Research Center. Do people click on links in Google AI summaries?. Pew Research Center, 2025. pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to
  31. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande. GEO: Generative Engine Optimization. arXiv, 2023. arxiv.org/abs/2311.09735
  32. Google Search Central. AI features and your website. Google, 2025. developers.google.com/search/docs/appearance/ai-features
  33. Chatterji et al.. How People Use ChatGPT. National Bureau of Economic Research, 2025. doi.org/10.3386/w34255
  34. Pew Research Center. Americans and AI 2026: Chatbots, Smart Devices and Views on Impact. Pew Research Center, 2026. pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-sma
  35. BrightLocal Research Team. Local Consumer Review Survey 2025. BrightLocal, 2025. brightlocal.com/research/local-consumer-review-survey-2025 Secondary source.
  36. Rosie Murphy. Local Consumer Review Survey 2026: Star Ratings Keep Rising, Old Reviews Don't Cut It. BrightLocal, 2026. brightlocal.com/research/local-consumer-review-survey/ Secondary source.
  37. Fishkin. 2024 Zero-Click Search Study: For every 1,000 EU Google Searches, only 374 clicks go to the Open Web. In the US, it’s 360.. SparkToro, 2024. sparktoro.com/blog/2024-zero-click-search-study-for-every-1000-us-goog Secondary source.
  38. Miriam Ellis. New Research: The Prevalence of AI Overviews in Local Search. Whitespark, 2025. whitespark.ca/blog/case-study-the-prevalence-of-ai-overviews-in-local- Secondary source.
  39. Khatiwoda. Large Firms With at Least 20 Employees Biggest AI Users. U.S. Census Bureau, 2026. census.gov/library/stories/2026/05/ai-use-businesses.html
  40. ADA Health Policy Institute. Dentists AI Usage and Attitudes. American Dental Association, 2026. ada.org/resources/research/health-policy-institute/dental-practice-res
  41. Mark Calaguas. 2024 Artificial Intelligence TechReport. American Bar Association, Law Practice Division (ABA TechReport), 2024. americanbar.org/groups/law_practice/resources/tech-report/2024/2024-ar
  42. Conductor. The 2026 AEO / GEO Benchmarks Report. Conductor, 2025. conductor.com/academy/aeo-geo-benchmarks-report/ Secondary source.
  43. Conductor. Health Care Industry: 2026 AEO / GEO Benchmarks. Conductor, 2025. conductor.com/academy/health-care-aeo-geo-benchmarks Secondary source.
  44. Conductor. Financials Industry: 2026 AEO / GEO Benchmarks. Conductor, 2025. conductor.com/academy/financials-aeo-geo-benchmarks/ Secondary source.
  45. Microsoft Clarity. AI Traffic Converts at 3x the Rate of Other Channels (Study). Microsoft, 2025. clarity.microsoft.com/blog/ai-traffic-converts-at-3x-the-rate-of-other Secondary source.
  46. Bain & Company, Global Consumer Lab. Why Consumers Choose an AI Chatbot Over a Search Engine (snap chart). Bain & Company, 2025. bain.com/insights/why-consumers-choose-an-ai-chatbot-over-a-search-eng Secondary source.
  47. Annapoorna Kini et al.. Conversational artificial intelligence for pre-procedural patient preparation: implementation, validation and patient satisfaction. npj Digital Medicine, 2025. pmc.ncbi.nlm.nih.gov/articles/PMC13710839
  48. Relatient. Relatient Voice AI Deflects More Than 60% of Routine Appointment Scheduling Calls. Relatient (via Business Wire), 2025. relatient.com/resources/press-release/relatient-voice-ai-deflects-more Secondary source.
  49. Henry Schein One. 2026 Catalyst Index: Dental Practice Benchmarking Report. Henry Schein One, 2026. henryscheinone.com/insights/ebook/2026-catalyst-index/ Secondary source.
  50. Santo L, Kang K. National Ambulatory Medical Care Survey: 2019 National Summary Tables. National Center for Health Statistics, 2019. stacks.cdc.gov/view/cdc/123251/cdc_123251_DS1.pdf
  51. Daniel Hood. Half of firms' clients have been with them 8+ years. Accounting Today (Arizent), 2019. accountingtoday.com/list/half-of-firms-clients-have-been-with-them-8-y Secondary source.
  52. Clio (Themis Solutions Inc.). 2025 Legal Trends Report. Clio; PDF copy hosted by the Alaska Bar Association, 2025. alaskabar.org/wp-content/uploads/2025-Legal-Trends-Report.pdf Secondary source.
  53. Vivek Pandya. AI Traffic Drives 83% Higher Revenue Per Visit. Adobe, 2026. business.adobe.com/blog/ai-traffic-higher-revenue-per-visit-retail-hol Secondary source.
  54. Technical Marketing. Which software will sign a HIPAA BAA? 28 common tools, checked. Technical Marketing, 2026-11. technicalmarketing.ai/reports/hipaa-baa-availability/ Our own data.
  55. HHS Office for Civil Rights. Business Associates Frequently Asked Questions (FAQs). U.S. Department of Health and Human Services, 2026. hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/i
  56. Technical Marketing. Pricing. Technical Marketing, 2026-11. technicalmarketing.ai/pricing/ Our own data.
  57. Technical Marketing. HIPAA-ready programs. Technical Marketing, 2026-11. technicalmarketing.ai/hipaa/ Our own data.
  58. George J. Stigler. The Economics of Information. Journal of Political Economy (University of Chicago Press), 1959-02. doi.org/10.1086/258464
  59. State Bar of Michigan. 2023 Economics of Law Survey. State Bar of Michigan, 2023. michbar.org/file/pmrc/pdfs/EoL_Report23.pdf
  60. National Association of Tax Professionals. NATP TAXPRO guide to the 2025 Fee Study (TAXPROPages). National Association of Tax Professionals, 2025. dam.natptax.com/m/39796eb5ee56a576/original/TAXPROPages.pdf
  61. Steven R. Machlin and Emily M. Mitchell. Statistical Brief #517: Expenses for Office-Based Physician Visits by Specialty and Insurance Type, 2016. Agency for Healthcare Research and Quality (MEPS), 2016. meps.ahrq.gov/data_files/publications/st517/stat517.shtml
  62. Phillip Nelson. Information and Consumer Behavior. Journal of Political Economy (University of Chicago Press), 1970. doi.org/10.1086/259630 Held through its record or abstract; passage not matched.
  63. Makoto Nakayama; Norma Sutcliffe; Yun Wan. Has the web transformed experience goods into search goods?. Electronic Markets (Springer), 2010. doi.org/10.1007/s12525-010-0041-z Secondary source.
  64. Uwe Dulleck; Rudolf Kerschbamer. On Doctors, Mechanics, and Computer Specialists: The Economics of Credence Goods. Journal of Economic Literature (American Economic Association), 2006. doi.org/10.1257/002205106776162717
  65. Zeithaml. How Consumer Evaluation Processes Differ Between Goods and Services. American Marketing Association, 1981. Held through its record or abstract; passage not matched.
  66. Mike Bendixen; Graham Gault. Consumer evaluation of perceived risks for goods and services. South African Journal of Business Management, 1981. sajbm.org/index.php/sajbm/article/download/827/761 Secondary source.
  67. George A. Akerlof. The Market for "Lemons": Quality Uncertainty and the Market Mechanism. The Quarterly Journal of Economics (Oxford University Press), 1970. doi.org/10.2307/1879431
  68. American Bar Association. Rule 7.2: Communications Concerning a Lawyer's Services: Specific Rules (ABA Model Rules). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  69. Michael Spence. Job Market Signaling. The Quarterly Journal of Economics (Oxford University Press), 1973. doi.org/10.2307/1882010
  70. David A. Hanauer; Kai Zheng; Dianne C. Singer; Achamyeleh Gebremariam; Matthew M. Davis. Public Awareness, Perception, and Use of Online Physician Rating Sites. JAMA, 2012-09. doi.org/10.1001/jama.2013.283194
  71. A. Parasuraman; Valarie A. Zeithaml; Leonard L. Berry. A Conceptual Model of Service Quality and Its Implications for Future Research. Journal of Marketing (American Marketing Association), 1985. doi.org/10.1177/002224298504900403 Held through its record or abstract; passage not matched.
  72. Aurelio Mauri; Roberta Minazzi; Simonetta Muccio. A Review of Literature on the Gaps Model on Service Quality: A 3-Decades Period: 1985-2013. International Business Research (Canadian Center of Science and Education), 2013. doi.org/10.5539/ibr.v6n12p134 Secondary source.
  73. A. Parasuraman; Valarie A. Zeithaml; Leonard L. Berry. SERVQUAL: A Multiple-Item Scale for Measuring Customer Perceptions of Service Quality. Marketing Science Institute, 1986. msi.org/working-paper/servqual-a-multipleitem-scale-for-measuring-cust
  74. Parasuraman,Zeithaml,Berry. SERVQUAL: A Multiple-Item Scale for Measuring Consumer Perceptions of Service Quality. Journal of Retailing, 1988. Held through its record or abstract; passage not matched.
  75. Michael Luca. Reviews, Reputation, and Revenue: The Case of Yelp.Com. Harvard Business School Working Paper 12-016, 2011. econpapers.repec.org/RePEc:hbs:wpaper:12-016
  76. Sarah Yang (UC Berkeley Media Relations). Yelp reviews boost restaurant business. Berkeley News, University of California, Berkeley, 2012. news.berkeley.edu/2012/09/04/yelp-reviews-boost-restaurant-business/ Secondary source.
  77. Anderson and Magruder. Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database. Oxford University Press (OUP), 2012. doi.org/10.1111/j.1468-0297.2012.02512.x
  78. Judith A. Chevalier; Dina Mayzlin. The Effect of Word of Mouth on Sales: Online Book Reviews. Journal of Marketing Research (American Marketing Association), 2006. doi.org/10.1509/jmkr.43.3.345
  79. Gao, McCullough, Agarwal, Jha. A Changing Landscape of Physician Quality Reporting: Analysis of Patients’ Online Ratings of Their Physicians Over a 5-Year Period. JMIR Publications Inc., 2012. doi.org/10.2196/jmir.2003
  80. Ye Lin; Y Alicia Hong; Bradley S Henson; Robert D Stevenson; Simon Hong; Tianchu Lyu; Chen Liang. Assessing Patient Experience and Healthcare Quality of Dental Care Using Patient Online Reviews in the United States: Mixed Methods Study. Journal of Medical Internet Research, 2020. doi.org/10.2196/18652
  81. Adam E. Block, Medha Reddy and Ella Corkum. The Impact of 1-Star Physician Ratings on New Patient Volume. The American Journal of Managed Care, 2015-01-01 to 2018-04-25. ajmc.com/view/the-impact-of-1-star-physician-ratings-on-new-patient-vo
  82. Judith A. Chevalier; Dina Mayzlin. The Effect of Word of Mouth on Sales: Online Book Reviews. National Bureau of Economic Research (NBER Working Paper 10148), 2003. nber.org/papers/w10148.pdf
  83. Federal Trade Commission. 16 CFR 465.2 Fake or false consumer reviews, consumer testimonials, or celebrity testimonials. Legal Information Institute (copy of the e-CFR text), 2024-10-21. law.cornell.edu/cfr/text/16/465.2
  84. Federal Trade Commission. 16 CFR 465.4 Buying positive or negative consumer reviews. Legal Information Institute (copy of the e-CFR text), 2024-10-21. law.cornell.edu/cfr/text/16/465.4
  85. Federal Trade Commission. 16 CFR 465.7 Review suppression. Legal Information Institute (copy of the e-CFR text), 2024-10-21. law.cornell.edu/cfr/text/16/465.7
  86. BrightLocal. Local Consumer Review Survey 2023. BrightLocal, 2023. brightlocal.com/research/local-consumer-review-survey-2023/ Secondary source.
  87. FindLaw (Thomson Reuters). The 2024 U.S. Consumer Legal Needs Survey: What attorneys need to know. FindLaw, 2024. findlaw.com/lawyer-marketing/blog/the-2024-u-s-consumer-legal-needs-su Secondary source.
  88. Susannah Fox and Maeve Duggan. Health Online 2013. Pew Research Center, 2012. pewresearch.org/internet/2013/01/15/health-online-2013/
  89. MSDC Staff. Report: Digital Medical Experience For Patients a Major Determinant in Booking an Appointment. Medical Society of the District of Columbia, 2021. msdc.org/detail/news/2021/12/08/report-digital-medical-experience-for- Secondary source.
  90. ABA Journal. How People Find Lawyers: Referrals Are Popular, Blogs Not So Much, Poll Finds. ABA Journal, 2010-09. abajournal.com/news/article/how_people_find_lawyers_referrals_are_popu Secondary source.
  91. Illinois Supreme Court Commission on Professionalism / Clio. Clio Legal Trends Report Finds Law Firms Aren't Responsive to Potential Clients. Illinois Supreme Court Commission on Professionalism, 2019. 2civility.org/clio-legal-trends-report-finds-law-firms-arent-responsiv Secondary source.
  92. Ana Maria Cristina Tâncu; Victor Lorin Purcărea; Mihaela Pantea; Marina Imre. Key Factors Correlations in Selecting Dental Services. Journal of Medicine and Life, 2019. doi.org/10.25122/jml-2019-0023
  93. MouthWatch. Dental Practice Marketing in 2026: What 448 Practices Say. MouthWatch, 2026. mouthwatch.com/resources/dental-practice-marketing-2026 Secondary source.
  94. IronMonk (press release via PR Newswire). New Survey Reveals Americans Choose Dentists Based on Online Reviews, Proximity and Rank in the Search Results. BioSpace (PR Newswire release dated Aug. 27, 2019), 2019. biospace.com/new-survey-reveals-americans-choose-dentists-based-on-onl Secondary source.
  95. Press Ganey. High standards, high stakes: How healthcare provider reviews drive patient choice. Press Ganey, 2023. pressganey.com/?p=19081 Secondary source.
  96. CPA.com. Not all leads are treated equally. CPA.com, 2024. cpa.com/sites/cpa/files/2024-11/Not_all_Leads_are_Treated_Equally.pdf Secondary source.
  97. Jason Bramwell. What Do Businesses Actually Value When Choosing or Staying with an Accounting Firm?. CPA Practice Advisor (reporting TaxDome's 2025 Niche Business Accounting Report), 2025. cpapracticeadvisor.com/2025/10/29/what-do-businesses-actually-value-wh Secondary source.
  98. Internal Revenue Service (survey by Pacific Consulting Group). Comprehensive Taxpayer Attitude Survey, Executive Report (January 2022), IRS Publication 5296. Internal Revenue Service, 2021. irs.gov/pub/irs-prior/p5296--2022.pdf
  99. Johannes Gluckler and Thomas Armbruster. Bridging Uncertainty in Management Consulting: The Mechanisms of Trust and Networked Reputation. Organization Studies / SAGE Publications, 2003. doi.org/10.1177/0170840603024002346 Held through its record or abstract; passage not matched.
  100. Google. Tips to improve your local ranking on Google (Google Business Profile Help). Google Business Profile Help, 2026-11. support.google.com/business/answer/7091?hl=en
  101. ADA Health Policy Institute. The State of the U.S. Dental Economy 4th Quarter 2025 Update. American Dental Association, Q4 2025. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  102. ADA Health Policy Institute. The State of the U.S. Dental Economy 3rd Quarter 2026 Update. American Dental Association, Q3 2026. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  103. Randall A. Lewis; Justin M. Rao. The Unfavorable Economics of Measuring the Returns to Advertising. The Quarterly Journal of Economics (Oxford University Press), 2015. doi.org/10.1093/qje/qjv023
  104. Brett R. Gordon; Florian Zettelmeyer; Neha Bhargava; Dan Chapsky. A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook. Marketing Science (INFORMS), 2019. doi.org/10.1287/mksc.2018.1135
  105. Thomas Blake; Chris Nosko; Steven Tadelis. Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment. Econometrica (Econometric Society), 2015. doi.org/10.3982/ECTA12423
  106. Erik Brynjolfsson; Lorin M. Hitt; Heekyung Hellen Kim. Strength in Numbers: How Does Data-Driven Decisionmaking Affect Firm Performance?. International Conference on Information Systems (ICIS) 2011 Proceedings (AIS Electronic Library), 2011. doi.org/10.2139/ssrn.1819486
  107. Nicholas Bloom; John Van Reenen. Measuring and Explaining Management Practices Across Firms and Countries. LSE Centre for Economic Performance Discussion Paper (published in The Quarterly Journal of Economics, 2007), 2007. doi.org/10.1162/qjec.2007.122.4.1351
  108. Erik Brynjolfsson; Kristina McElheran. The Rapid Adoption of Data-Driven Decision-Making. American Economic Review: Papers and Proceedings (American Economic Association), 2005-2010. doi.org/10.1257/aer.p20161016
  109. Sunil Gupta; Donald R. Lehmann; Jennifer Ames Stuart. Valuing Customers. Journal of Marketing Research (American Marketing Association), 2004. doi.org/10.1509/jmkr.41.1.7.25084
  110. Reichheld. Zero defections: quality comes to services.. Harvard Business Review, 1990. eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=1010708
  111. Leonard L. Berry. Relationship marketing of services-growing interest, emerging perspectives. Journal of the Academy of Marketing Science (Springer), 1995. doi.org/10.1177/009207039502300402
  112. James W. Mold, George E. Fryer, A. Michelle Roberts. When Do Older Patients Change Primary Care Physicians?. Journal of the American Board of Family Practice, 1999-2001. jabfm.org/content/17/6/453.long
  113. CPA.com and AICPA Private Companies Practice Section (PCPS). Findings and insights from the 2024 CPA.com & AICPA PCPS client advisory services (CAS) benchmark survey. CPA.com / AICPA, 2024. cpa.com/sites/cpa/files/2024-12/2024-CAS-Benchmark-Survey.pdf
  114. NATP Media Relations Team. 2025 NATP Fee Study - Media Highlights. National Association of Tax Professionals, 2025. dam.natptax.com/m/5f112786acbd1130/original/2025_NATP_Fee_Study_MediaH
  115. Rand Fishkin. NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility. SparkToro, 2025. sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recom Secondary source.
  116. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Dentists. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/healthcare/dentists.htm
  117. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 6212 Offices of dentists (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/6212.csv
  118. R. Henry Olaisen and Richard J. Manski. Dental Utilization and Expenditures, U.S. Civilian Noninstitutionalized Population Aged 2 and Older, 2019-2021 (Statistical Brief #555). Agency for Healthcare Research and Quality, 2021. meps.ahrq.gov/data_files/publications/st555/stat555.pdf
  119. National Center for Health Statistics. FastStats: Oral and Dental Health. Centers for Disease Control and Prevention, 2025. cdc.gov/nchs/fastats/dental.htm
  120. National Center for Health Statistics. Dental Care Utilization Among Adults Aged 18-64: United States, 2019 and 2020 (NCHS Data Brief No. 435). Centers for Disease Control and Prevention, 2020. cdc.gov/nchs/products/databriefs/db435.htm
  121. HHS Office for Civil Rights. Dental Practice Pays $10,000 to Settle Social Media Disclosures of Patients’ Protected Health Information. U.S. Department of Health and Human Services, 2019-10-02. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/elit
  122. HHS Office for Civil Rights and Elite Dental Associates. Resolution Agreement and Corrective Action Plan, Elite Dental Associates (HHS TN 16-240465). U.S. Department of Health and Human Services, 2016-06-04. hhs.gov/sites/default/files/elite-dental-ra-cap.pdf
  123. HHS Office for Civil Rights and New Vision Dental. Resolution Agreement and Corrective Action Plan, New Vision Dental (OCR Transaction Number 18-288980). U.S. Department of Health and Human Services, 2017-11-29. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/new-
  124. HHS Office for Civil Rights and Manasa Health Center. Resolution Agreement and Corrective Action Plan, Manasa Health Center, LLC (HHS TN 20-380176). U.S. Department of Health and Human Services, 2023-03-28. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/mana
  125. HHS Office for Civil Rights. HHS Civil Rights Office Enters Settlement with Dental Practice Over Disclosures of Patients’ Protected Health Information. U.S. Department of Health and Human Services, 2022-12-14. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/new-
  126. HHS Office for Civil Rights. Manasa Health Center resolution agreement (HHS HIPAA enforcement page). U.S. Department of Health and Human Services, 2023-06-05. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/mana
  127. HHS Office for Civil Rights. Dr. U. Phillip Igbinadolor, D.M.D. & Associates, P.A. (UPI) HIPAA Enforcement Action. U.S. Department of Health and Human Services, 2022-03-28. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/upi/
  128. U.S. Department of Health and Human Services. Notification of Enforcement Discretion Regarding HIPAA Civil Money Penalties (84 FR 18151). Federal Register (govinfo.gov), 2019-04-30. govinfo.gov/content/pkg/FR-2019-04-30/pdf/2019-08530.pdf
  129. U.S. Department of Health and Human Services. Annual Civil Monetary Penalties Inflation Adjustment (Federal Register document 2026-01688). Federal Register, 2025 adjustment, published 2026-01-28. public-inspection.federalregister.gov/2026-01688.pdf
  130. California Legislature. California Business and Professions Code section 651. FindLaw (reproduction of the California Code), 2026-01-01. codes.findlaw.com/ca/business-and-professions-code/bpc-sect-651/
  131. California Legislature. California Business and Professions Code section 1680. FindLaw (reproduction of the California Code), 2026-01-01. codes.findlaw.com/ca/business-and-professions-code/bpc-sect-1680/
  132. Texas State Board of Dental Examiners. 22 Tex. Admin. Code section 108.54 - Advertising of Specialties. Cornell Law School Legal Information Institute (reproduction of the Texas Administrative Code), 2021-04-01. law.cornell.edu/regulations/texas/22-Tex-Admin-Code-SS-108-54
  133. Texas State Board of Dental Examiners. 22 Tex. Admin. Code section 108.57 - False, Misleading or Deceptive Advertising. Cornell Law School Legal Information Institute (reproduction of the Texas Administrative Code), 2013-05-01. law.cornell.edu/regulations/texas/22-Tex-Admin-Code-SS-108-57
  134. Florida Legislature. Florida Statutes 466.019 Advertising by dentists. The Florida Legislature (Online Sunshine), 2024. flrules.org/gateway/statute.asp?id=466.019
  135. New York State Board of Regents. N.Y. Comp. Codes R. & Regs. Tit. 8 section 29.5 - Special provisions for the professions of dentistry and dental hygiene. Cornell Law School Legal Information Institute (reproduction of the New York Codes, Rules and Regulations), 2026-11-10. law.cornell.edu/regulations/new-york/8-NYCRR-29.5
  136. New York State Board of Regents. N.Y. Comp. Codes R. & Regs. Tit. 8 section 29.1 - General provisions. Cornell Law School Legal Information Institute (reproduction of the New York Codes, Rules and Regulations), 2026-11-10. law.cornell.edu/regulations/new-york/8-NYCRR-29.1
  137. Washington Dental Quality Assurance Commission. WAC 246-817-420 Specialty representation. Washington State Legislature (Washington Administrative Code), 2021-07-03. lawfilesext.leg.wa.gov/Law/WAC/WAC%20246%20%20TITLE/WAC%20246%20-817%2
  138. Federal Trade Commission. Trade Regulation Rule on the Use of Consumer Reviews and Testimonials (89 FR 68034). Federal Register (govinfo.gov), 2024-10-21. govinfo.gov/content/pkg/FR-2024-08-22/pdf/2024-18519.pdf
  139. Federal Trade Commission. FTC Warns 10 Companies About Possible Violations of the Agency's New Consumer Review Rule. Federal Trade Commission, 2025-12-22. ftc.gov/news-events/news/press-releases/2025/12/ftc-warns-10-companies
  140. PatientNews. Mastering Dental Phone Conversion. PatientNews, undated. patientnews.com/article/mastering-dental-phone-conversion Secondary source.
  141. Reddy S, Speer M, Saxon M, Ziegler M, Dedolph Z, Qaasim S. Evaluating network adequacy of oral health services for children on Medicaid in Arizona. AIMS Public Health (PMC8755958), 2021. pmc.ncbi.nlm.nih.gov/articles/PMC8755958/
  142. Gurol-Urganci I, de Jongh T, Vodopivec-Jamsek V, Atun R, Car J. Mobile phone messaging reminders for attendance at healthcare appointments. Cochrane Database of Systematic Reviews (abstract via Europe PMC), 2013. ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:24310741%20AN
  143. Jennifer S Holtzman, Kathryn A Atchison et al.. The association between oral health literacy and failed appointments in adults attending a university based general dental clinic. Community Dentistry and Oral Epidemiology (via PubMed Central), 2014. pmc.ncbi.nlm.nih.gov/articles/PMC4024362 Secondary source.
  144. Henry Schein One. What Are Good No-Show and Cancellation Rates for a DSO or Multi-Location Dental Practice?. Henry Schein One, 2026. henryscheinone.com/insights/blogs/dso-dental-practice-no-show-cancella Secondary source.
  145. ADA Standards Committee on Dental Informatics. ADA SCDI White Paper No. 1106, Dentistry - Overview of Artificial and Augmented Intelligence Uses in Dentistry. American Dental Association, 2022-12-30. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/p
  146. ADA Council on Dental Practice. 2026 Council on Dental Practice (annual report to the House of Delegates). American Dental Association, 2026. ada.org/-/media/project/ada-organization/ada/ada-org/files/about/annua
  147. ADA News. ADA responds to HHS request for information on AI adoption in dentistry. American Dental Association, 2026-02. adanews.ada.org/ada-news/2026/february/ada-responds-to-hhs-request-for
  148. U.S. Department of Health and Human Services. 45 CFR 164.502 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.502
  149. California Legislature. California Health and Safety Code section 1339.75 (Chapter 2.13, Artificial Intelligence in Health Care Services). California Legislative Information (leginfo.legislature.ca.gov), 2024. leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=HS
  150. Texas Legislature. H.B. No. 149 (enrolled), Texas Responsible Artificial Intelligence Governance Act. Texas Legislature Online (capitol.texas.gov), 2026-01-01. capitol.texas.gov/tlodocs/89R/billtext/html/HB00149F.htm
  151. HHS Office for Civil Rights. Are we required to “certify” our organization’s compliance with the standards of the Security Rule?. U.S. Department of Health and Human Services, 2013-07-26. hhs.gov/hipaa/for-professionals/faq/are-we-required-to-certify-our-org
  152. U.S. Department of Health and Human Services. 45 CFR 164.508 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.508
  153. Statista Research Department. Average number of yearly patient visits per U.S. dentist from 1990 to 2025. Statista, 2025. statista.com/statistics/965365/patient-visits-among-all-dentists-us-pe Secondary source.
  154. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Physicians and Surgeons. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/healthcare/physicians-and-surgeons.htm
  155. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 6211 Offices of physicians (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/6211.csv
  156. Chris Harrop, MGMA. Operating costs keep climbing for medical practices in 2026 (MGMA Stat). Medical Group Management Association, 2026-06-23. mgma.com/mgma-stat/operating-costs-keep-climbing-for-medical-practices
  157. MGMA. Does your margin have breathing room? Insights from MGMA DataDive benchmarks (MGMA Stat). Medical Group Management Association, 2011-2024. mgma.com/mgma-stat/does-your-margin-have-breathing-room-mgma-stat-data
  158. Liu, Zhang, and Liang. Evaluating Verifiability in Generative Search Engines. Association for Computational Linguistics, 2023. arxiv.org/abs/2304.09848
  159. Kevin N Griffith et al.. Call Center Performance Affects Patient Perceptions of Access and Satisfaction. American Journal of Managed Care, 2014-10. pmc.ncbi.nlm.nih.gov/articles/PMC8177735/
  160. Dr. Shahinaz Soliman, M.D. (CallMyDoc). State of Patient Phone Communication 2026: Data from 29 Million Calls. CallMyDoc (Sphinx Medical Technologies, Inc.), 2025. blog.callmydoc.com/new-blog-1/patient-phone-communication-2026 Secondary source.
  161. Bisgaier. Auditing Access to Specialty Care for Children with Public Insurance. Massachusetts Medical Society, 2011. doi.org/10.1056/NEJMsa1013285
  162. Parviz Kheirkhah, Qianmei Feng, Lauren M Travis, Shahriar Tavakoli-Tabasi, Amir Sharafkhaneh. Prevalence, predictors and economic consequences of no-shows. BMC Health Services Research, 1997-2008. pmc.ncbi.nlm.nih.gov/articles/PMC4714455/
  163. Tarabichi Y, Higginbotham J, Riley N, Kaelber DC, Watts B. Reducing Disparities in No Show Rates Using Predictive Model-Driven Live Appointment Reminders for At-Risk Patients: a Randomized Controlled Quality Improvement Initiative. Journal of General Internal Medicine (PMC10150669), 2022. pmc.ncbi.nlm.nih.gov/articles/PMC10150669
  164. McLean SM, Booth A, Gee M, Salway S, Cobb M, Bhanbhro S, Nancarrow SA. Appointment reminder systems are effective but not optimal: results of a systematic review and evidence synthesis employing realist principles. Patient Preference and Adherence, 2016. pmc.ncbi.nlm.nih.gov/articles/PMC4831598/
  165. Majed Al-Turbag, Mary Mooney, Margarita Corry. A systematic review and meta-analysis of appointment reminders for enhancing hospital attendance. Journal of Hospital Management and Health Policy, 2026. jhmhp.amegroups.org/article/view/10215/html
  166. MGMA. Patient no-shows in 2025. Medical Group Management Association (MGMA Stat), 2025-08-12. mgma.com/mgma-stat/patient-no-shows-in-2025
  167. U.S. Department of Health and Human Services. 45 CFR 164.504 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.504
  168. HHS Office for Civil Rights. Business Associate Contracts: Sample Business Associate Agreement Provisions. U.S. Department of Health and Human Services, 2013. hhs.gov/hipaa/for-professionals/covered-entities/sample-business-assoc
  169. HHS Office for Civil Rights. HIPAA Security Rule To Strengthen the Cybersecurity of Electronic Protected Health Information (notice of proposed rulemaking, 90 FR 898). Federal Register (govinfo.gov), 2025-01-06. govinfo.gov/content/pkg/FR-2025-01-06/pdf/2024-30983.pdf
  170. HHS Office for Civil Rights. HHS/OCR RIN 0945-AA22: HIPAA Security Rule to Strengthen the Cybersecurity of Electronic Protected Health Information (Unified Agenda entry). Reginfo.gov (OIRA / GSA), 2026-11-10. reginfo.gov/public/do/eAgendaViewRule?pubId=202510&RIN=0945-AA22
  171. HHS Office for Civil Rights. Use of Online Tracking Technologies by HIPAA Covered Entities and Business Associates (archived copy, January 2023). Internet Archive copy of hhs.gov, 2022-12-01. web.archive.org/web/20230115000000/https://www.hhs.gov/hipaa/for-profe
  172. United States District Court for the Northern District of Texas. American Hospital Association v. Becerra, No. 4:23-cv-01110-P (N.D. Tex.), Opinion and Order, ECF No. 67. Court order, as hosted by the American Hospital Association, 2024-06-20. aha.org/system/files/media/file/2024/06/opinion-order-in-aha-et-al-v-x
  173. Federal Trade Commission. FTC Enforcement Action to Bar GoodRx from Sharing Consumers’ Sensitive Health Info for Advertising. Federal Trade Commission, 2023-02-01. ftc.gov/news-events/news/press-releases/2023/02/ftc-enforcement-action
  174. Federal Trade Commission. FTC to Ban BetterHelp from Revealing Consumers’ Data, Including Sensitive Mental Health Information, to Facebook and Others for Targeted Advertising. Federal Trade Commission, 2023-03-02. ftc.gov/news-events/news/press-releases/2023/03/ftc-ban-betterhelp-rev
  175. Federal Trade Commission. Ovulation Tracking App Premom Will be Barred from Sharing Health Data for Advertising Under Proposed FTC Order. Federal Trade Commission, 2023-05-17. ftc.gov/news-events/news/press-releases/2023/05/ovulation-tracking-app
  176. Federal Trade Commission. Proposed FTC Order will Prohibit Telehealth Firm Cerebral from Using or Disclosing Sensitive Data for Advertising Purposes, and Require it to Pay $7 Million. Federal Trade Commission, 2024-04-15. ftc.gov/news-events/news/press-releases/2024/04/proposed-ftc-order-wil
  177. Federal Trade Commission. Health Breach Notification Rule (89 FR 47028, final rule). Federal Register (govinfo.gov), 2024-07-29. govinfo.gov/content/pkg/FR-2024-05-30/pdf/2024-10855.pdf
  178. Federal Trade Commission. FTC Finalizes Changes to the Health Breach Notification Rule. Federal Trade Commission, 2024-04-26. ftc.gov/news-events/news/press-releases/2024/04/ftc-finalizes-changes-
  179. Washington State Legislature. Chapter 19.373 RCW, Washington My Health My Data Act (RCW 19.373.010(8)). Washington State Legislature, 2023. app.leg.wa.gov/rcw/default.aspx?cite=19.373&full=true
  180. Federal Trade Commission. Health Products Compliance Guidance. Federal Trade Commission, 2022-12. ftc.gov/business-guidance/resources/health-products-compliance-guidanc
  181. California Legislature. California Business and Professions Code section 2271. FindLaw (reproduction of the California Code), 2026-01-01. codes.findlaw.com/ca/business-and-professions-code/bpc-sect-2271/
  182. Texas Medical Board. 22 Tex. Admin. Code section 164.3 - Prohibited Acts or Omissions in Advertising. Cornell Law School Legal Information Institute (reproduction of the Texas Administrative Code), 2025-01-09. law.cornell.edu/regulations/texas/22-Tex-Admin-Code-SS-164-3
  183. Utah Legislature. Utah Code 13-77-103 Required disclosures. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S103_2025050720250507.html
  184. Utah Legislature. Utah Code 13-77-101 Definitions. Utah State Legislature (le.utah.gov), 2026-05-06. le.utah.gov/xcode/Title13/Chapter77/C13-77-S101_2026050620260506.html
  185. American Medical Association. Augmented Intelligence Development, Deployment, and Use in Health Care (AMA policy). American Medical Association, 2024-11-12. ama-assn.org/system/files/ama-ai-principles.pdf
  186. Federal Communications Commission. Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts, Declaratory Ruling, FCC 24-17. Federal Communications Commission, 2024-02-08. docs.fcc.gov/public/attachments/FCC-24-17A1.pdf
  187. Federal Communications Commission. Implications of Artificial Intelligence Technologies on Protecting Consumers From Unwanted Robocalls and Robotexts (notice of proposed rulemaking), 89 Fed. Reg. 73321. Federal Register / U.S. Government Publishing Office, 2024-09-10. govinfo.gov/content/pkg/FR-2024-09-10/html/2024-19028.htm
  188. MGMA. Sizing up the market for AI chatbots and virtual assistants in medical practices in 2025. Medical Group Management Association (MGMA Stat), 2025-04-08. mgma.com/mgma-stat/sizing-up-the-market-for-ai-chatbots-virtual-assist
  189. MGMA. Automatic for the people: AI for front office access. Medical Group Management Association (MGMA Stat), 2026-02-10. mgma.com/mgma-stat/automatic-for-the-people-ai-for-front-office-access
  190. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. National Institute of Standards and Technology, 2024-07. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
  191. Christopher C. Rivers, Tribunal Member, Civil Resolution Tribunal (British Columbia). Moffatt v. Air Canada, 2024 BCCRT 149. Civil Resolution Tribunal of British Columbia, 2024-02-14. decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
  192. United States Congress. 18 U.S.C. 2511: Interception and disclosure of wire, oral, or electronic communications prohibited. Legal Information Institute, Cornell Law School (US Code text), 2026. law.cornell.edu/uscode/text/18/2511
  193. Agency for Healthcare Research and Quality, Center for Financing, Access, and Cost Trends. MEPS HC-248G: 2023 Office-Based Medical Provider Visits. Agency for Healthcare Research and Quality (MEPS), 2023. meps.ahrq.gov/data_stats/download_data/pufs/h248g/h248gdoc.shtml
  194. National Center for Health Statistics. FastStats - Ambulatory Care Use and Physician office visits. Centers for Disease Control and Prevention, 2019. cdc.gov/nchs/fastats/physician-visits.htm
  195. Lisa S Rotenstein, Nate Apathy, Susan Edgman-Levitan, Bruce Landon. Comparison of Work Patterns Between Physicians and Advanced Practice Practitioners in Primary Care and Specialty Practice Settings. JAMA Network Open, 2021. ncbi.nlm.nih.gov/pmc/articles/PMC10265293/
  196. U.S. Bureau of Labor Statistics. Table 4. Employer Costs for Employee Compensation for private industry workers by occupational and industry group [June 2026]. U.S. Bureau of Labor Statistics, 2026-06. bls.gov/news.release/ecec.t04.htm
  197. MGMA. Phones are still a backlog costing medical practices time. Medical Group Management Association (MGMA Stat), 2026-03-10. mgma.com/mgma-stat/phones-are-still-a-backlog-costing-medical-practice
  198. Call Box. The Top 5 Phone Trends of 2020 and How Healthcare Practices Are Taking Action: Data Study. Call Box, 2020-02. assets-cdn.carwars.com/websites/callbox/resources/dental/CB+Healthcare Secondary source.
  199. American Bar Association. ABA Profile of the Legal Profession. American Bar Association, 2025. americanbar.org/news/profile-legal-profession/
  200. American Bar Association. ABA Profile of the Legal Profession 2024 Demographics. American Bar Association, 2024-01-01. americanbar.org/news/profile-legal-profession/demographics/
  201. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Lawyers. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/legal/lawyers.htm
  202. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 54111 Offices of lawyers (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/54111.csv
  203. U.S. Census Bureau. Quarterly Selected Services Estimates, Fourth Quarter 2025 (advance), Table A4. U.S. Census Bureau, 2025. www2.census.gov/services/qss/2025/qssq4-25advpr.pdf
  204. American Bar Association. Comment on Rule 7.1 (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  205. American Bar Association. Comment on Rule 7.2 (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  206. American Bar Association. Rule 7.3 Solicitation of Clients (ABA Model Rules of Professional Conduct). American Bar Association, 2019. americanbar.org/groups/professional_responsibility/publications/model_
  207. BrightLocal. Local Consumer Review Survey 2024. BrightLocal, 2024. brightlocal.com/research/local-consumer-review-survey-2024/ Secondary source.
  208. ABA Standing Committee on Ethics and Professional Responsibility. ABA Formal Opinion 496: Responding to Online Criticism. American Bar Association, 2021-01-13. americanbar.org/content/dam/aba/administrative/professional_responsibi
  209. Federal Trade Commission. The Consumer Reviews and Testimonials Rule: Questions and Answers. Federal Trade Commission, 2024-10-21. ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule
  210. Federal Trade Commission. Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials. Federal Trade Commission, 2024-08-14. ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commissi
  211. Clio. AI-powered legal practices surge: Clio's latest Legal Trends Report reveals major shift (press release). Clio (a legal practice-management software vendor), 2024. clio.com/about/press/clio-latest-legal-trends-report/ Secondary source.
  212. Clio (Themis Solutions Inc.). Legal Trends Report 2019. Clio, as hosted by the District of Columbia Bar, 2018-2019. dcbar.org/getmedia/a7bbe4b0-eec7-4fc7-8c85-485853c235ab/Clio-2019-Lega Secondary source.
  213. North Carolina Bar Association (NC Lawyer / CPM column). Secret Shop Your Law Firm. North Carolina Bar Association, 2016. ncbar.org/?p=10300 Secondary source.
  214. Hennessey Digital. 2025 Lead Form Response Times & Trends. Hennessey Digital, 2025. hennessey.com/2025-lead-form-response-time-study/ Secondary source.
  215. Insurance Research Council. Why Do Claimants Hire Attorneys?. Insurance Research Council, 2016. insurance-research.org/news/why-do-claimants-hire-attorneys
  216. Christi Crane (Ohio State News), on research led by Simon Haeder and Wendy Xu. Having insurance doesn't mean you can get a mental health appointment. The Ohio State University, 2025-10 to 2026-03. news.osu.edu/having-insurance-doesnt-mean-you-can-get-a-mental-health-
  217. SmartCompany. Are you letting online leads go cold? (quoting Oldroyd, McElheran and Elkington, The Short Life of Online Sales Leads, Harvard Business Review, March 2011). SmartCompany (Australia), 2011. smartcompany.com.au/marketing/20110704-are-you-letting-online-leads-go Secondary source.
  218. Oldroyd. The Short Life of Online Sales Leads. Harvard Business Review, 2011. thedenmangroupselling.wordpress.com/wp-content/uploads/2011/03/the-sho
  219. American Bar Association. Rule 1.18: Duties to Prospective Client (ABA Model Rules of Professional Conduct). American Bar Association, 2020. americanbar.org/groups/professional_responsibility/publications/model_
  220. American Bar Association. Rule 5.3: Responsibilities Regarding Nonlawyer Assistance (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  221. ABA Standing Committee on Ethics and Professional Responsibility. ABA Formal Opinion 512: Generative Artificial Intelligence Tools. American Bar Association, 2024-07-29. americanbar.org/content/dam/aba/administrative/professional_responsibi
  222. American Bar Association. Rule 1.6: Confidentiality of Information (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  223. American Bar Association. Comment on Rule 1.1 (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  224. Jim Ash. Board of Governors adopts ethics guidelines for generative AI use. The Florida Bar News, 2024-01-19. floridabar.org/the-florida-bar-news/board-of-governors-adopts-ethics-g
  225. The Florida Bar. Florida Bar Ethics Opinion 24-1 (January 19, 2024). The Florida Bar (copy hosted by University at Buffalo School of Law CLE materials), 2024-01-19. law.buffalo.edu/content/dam/law/content/cle/24mar27-materials/fl-bar-e
  226. State Bar of California, Committee on Professional Responsibility and Conduct. Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (Executive Summary). State Bar of California, 2026-05-14. calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guida
  227. Chief Judge David C. Godbey, U.S. District Court for the Northern District of Texas. Special Order No. 2-98 (adopting new rules LR 7.2(f) and LCrR 47.2(e)). U.S. District Court for the Northern District of Texas, 2024-09-03. txnd.uscourts.gov/sites/default/files/orders/2-98.pdf
  228. U.S. District Court for the Northern District of Texas. Local Civil Rules, N.D. Tex., LR 7.2(f) Disclosure of Use of Generative Artificial Intelligence. U.S. District Court for the Northern District of Texas, 2024-09-03. txnd.uscourts.gov/sites/default/files/documents/CIVRULES.pdf
  229. U.S. District Court for the Northern District of Texas. Judge Brantley Starr: Specific Requirements (Certificate regarding Artificial Intelligence). U.S. District Court for the Northern District of Texas, 2026-11-10. txnd.uscourts.gov/judge/judge-brantley-starr/
  230. Matthew Dahl, Varun Magesh, Mirac Suzgun, Daniel E. Ho. Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models. arXiv (v2, last revised 21 Jun 2024), 2024. arxiv.org/abs/2401.01301
  231. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, Daniel E. Ho. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv (submitted 30 May 2024), 2024. arxiv.org/abs/2405.20362
  232. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, Daniel E. Ho. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv (published in the Journal of Empirical Legal Studies, 2025), 2024. arxiv.org/pdf/2405.20362
  233. Vals AI. VLAIR - Legal Research. Vals AI, 2025. vals.ai/industry-reports/vlair-10-14-25 Secondary source.
  234. Judge P. Kevin Castel, U.S. District Court, S.D.N.Y.. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (Doc. 54). U.S. District Court for the Southern District of New York (RECAP/CourtListener copy of the filed order), 2023-06-22. storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.
  235. U.S. District Court for the District of Wyoming. Wadsworth v. Walmart Inc., No. 2:23-CV-118-KHR, Order on Sanctions and Other Disciplinary Action (Doc. 181). U.S. District Court for the District of Wyoming (RECAP/CourtListener copy of the filed order), 2025-01-22. storage.courtlistener.com/recap/gov.uscourts.wyd.64014/gov.uscourts.wy
  236. U.S. District Court for the Northern District of Alabama. Johnson v. Dunn, No. 2:21-cv-1701-AMM, Sanctions Order (Doc. 204). U.S. District Court for the Northern District of Alabama (Docket Alarm copy of the filed order), 2025-07-23. docketalarm.com/cases/Alabama_Northern_District_Court/2--21-cv-01701/J
  237. Damien Charlotin. AI Hallucination Cases Database. Damien Charlotin (personal research database), 2026-11-10. damiencharlotin.com/hallucinations/ Secondary source.
  238. Federal Trade Commission. FTC Announces Crackdown on Deceptive AI Claims and Schemes. Federal Trade Commission, 2024-09-25. ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdow
  239. Federal Trade Commission. FTC Finalizes Order with DoNotPay (FTC press release page). Federal Trade Commission, 2025-01-16. ftc.gov/node/87474
  240. Bob Ambrogi. Thomson Reuters Survey: Over 95% of Legal Professionals Expect Gen AI to Become Central to Workflow Within Five Years. LawSites (LawNext), reporting a Thomson Reuters Institute survey, 2025. lawnext.com/2025/04/thomson-reuters-survey-over-95-of-legal-profession Secondary source.
  241. Clio (Themis Solutions Inc.). 2023 Legal Trends for Solo Law Firms. Clio, 2022. clio.com/wp-content/uploads/2022/10/2023-Legal-Trends-for-Solo-Law-Fir Secondary source.
  242. Washington State Office of Public Defense. Private Representation Costs - Misdemeanors. Washington State Office of Public Defense, 2020. opd.wa.gov/sites/default/files/2023-08/00915-2020_MPrivateRates.pdf
  243. Insurance Research Council. Insurance Research Council Study Finds Rising Medical Costs and Attorney Involvement Are Driving Up Auto Injury Claim Costs. Insurance Research Council (The Institutes), 2022. insurance-research.org/news/insurance-research-council-study-finds-ris
  244. State Bar of California, Standing Committee on Professional Responsibility and Conduct. Arbitration Advisory No. 2025-01: Determination of a Reasonable Fee. State Bar of California, 2025. calbar.ca.gov/sites/default/files/portals/0/documents/mfa/2025/2025-01
  245. U.S. Bankruptcy Court, Eastern District of Wisconsin. Presumptively Reasonable ("No-Look") Fees in Chapter 13 Cases. U.S. Bankruptcy Court, Eastern District of Wisconsin (wieb.uscourts.gov), 2024. wieb.uscourts.gov/sites/default/files/forms/Presumptively%20Reasonable
  246. U.S. Bankruptcy Court, Central District of California. The Central Guide Supplement 3015-1(v): Presumptively Reasonable Attorney's Fees in Chapter 13 Cases. U.S. Bankruptcy Court, Central District of California (cacb.uscourts.gov), 2024. cacb.uscourts.gov/sites/cacb/files/documents/the-central-guide/TCGSupp
  247. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Accountants and Auditors. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/business-and-financial/accountants-and-auditors.htm
  248. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541211 Offices of certified public accountants (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541211.csv
  249. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541213 Tax preparation services (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541213.csv
  250. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541219 Other accounting services (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541219.csv
  251. National Association of State Boards of Accountancy (NASBA). How Many CPAs Are There?. NASBA, 2026-08-27. nasba.org/licensure/how-many-cpas/
  252. Internal Revenue Service. Tax Professional Management Office federal tax return preparer statistics. Internal Revenue Service, 2026-09-01. irs.gov/tax-professionals/tax-professional-management-office-federal-t
  253. AICPA. Accounting Firms Say Hiring Outlook Is Strong, AICPA Report Finds (news release on 2025 Trends). AICPA & CIMA, 2023-24. aicpa-cima.com/news/article/accounting-firms-report-strong-hiring-outl
  254. AICPA. Pool of Accounting Graduates Continues to Shrink in U.S., AICPA Report Finds. AICPA & CIMA, 2021-22. aicpa-cima.com/news/article/pool-of-accounting-graduates-continues-to-
  255. AICPA. Diversity Hiring of Graduates into the Accounting Profession is Increasing: AICPA Trends Report. AICPA & CIMA, 2019-20. aicpa-cima.com/news/article/diversity-hiring-of-graduates-into-the-acc
  256. Sweeney, J. T.; Summers, S. L.. The Effect of the Busy Season Workload on Public Accountants' Job Burnout. Behavioral Research in Accounting (American Accounting Association), 2002. doi.org/10.2308/bria.2002.14.1.223
  257. Persellin, J. S.; Schmidt, J. J.; Vandervelde, S. D.; Wilkins, M. S.. Auditor Perceptions of Audit Workloads, Audit Quality, and Job Satisfaction. Accounting Horizons (American Accounting Association), 2019. doi.org/10.2308/acch-52488
  258. CPA Practice Advisor. Senior Accountants Crushed by Busy Season Stress. CPA Practice Advisor, 2025. cpapracticeadvisor.com/2025/08/26/senior-accountants-crushed-by-busy-s Secondary source.
  259. American Institute of CPAs. CPA Firms Report Steady Growth in Revenue and Profit, AICPA Research Finds. AICPA, FY2024. aicpa-cima.com/news/article/cpa-firms-report-steady-growth-in-revenue-
  260. Internal Revenue Service. Filing season statistics for week ending April 3, 2026. Internal Revenue Service, 2026-04-03. irs.gov/newsroom/filing-season-statistics-for-week-ending-april-3-2026
  261. Internal Revenue Service. Filing season statistics for week ending May 8, 2026. Internal Revenue Service, 2026-05-08. irs.gov/newsroom/filing-season-statistics-for-week-ending-may-8-2026
  262. Noy and Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. doi.org/10.1126/science.adh2586
  263. Dell'Acqua et al.. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School, 2023. mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf
  264. U.S. Department of the Treasury. 31 CFR 10.30 - Solicitation. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/31/10.30
  265. Texas State Board of Public Accountancy. 22 Tex. Admin. Code section 501.82 - Advertising. Cornell Law School Legal Information Institute (reproduction of the Texas Administrative Code), 2026-11-10. law.cornell.edu/regulations/texas/22-Tex-Admin-Code-SS-501-82
  266. U.S. Securities and Exchange Commission. Investment Adviser Marketing: A Small Entity Compliance Guide. U.S. Securities and Exchange Commission, 2022-11-04. sec.gov/resources-small-businesses/small-business-compliance-guides/in
  267. Financial Industry Regulatory Authority. FINRA Rule 2210: Communications with the Public. FINRA, 2026-11-10. finra.org/rules-guidance/rulebooks/finra-rules/2210
  268. Financial Industry Regulatory Authority. Regulatory Notice 24-09: FINRA Reminds Members of Regulatory Obligations When Using Generative Artificial Intelligence and Large Language Models. FINRA, 2024-06-27. finra.org/rules-guidance/notices/24-09
  269. United States Congress (Internal Revenue Code). 26 U.S. Code 7216 - Disclosure or use of information by preparers of returns. Legal Information Institute, Cornell Law School (U.S. Code text), 2026-11-10. law.cornell.edu/uscode/text/26/7216
  270. United States Congress (Internal Revenue Code). 26 U.S. Code 6713 - Disclosure or use of information by preparers of returns. Legal Information Institute, Cornell Law School (U.S. Code text), 2026-11-10. law.cornell.edu/uscode/text/26/6713
  271. U.S. Department of the Treasury / Internal Revenue Service. 26 CFR 301.7216-1 - Penalty for disclosure or use of tax return information. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/26/301.7216-1
  272. U.S. Department of the Treasury / Internal Revenue Service. 26 CFR 301.7216-2 - Permissible disclosures or uses without consent of the taxpayer. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/26/301.7216-2
  273. U.S. Department of the Treasury / Internal Revenue Service. 26 CFR 301.7216-3 - Disclosure or use permitted only with the taxpayer's consent. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/26/301.7216-3
  274. Kelly D. Mullins (AICPA Professional Ethics Division). Accounting ethics in the age of AI. Journal of Accountancy (AICPA & CIMA), 2026-08-05. journalofaccountancy.com/news/2026/aug/accounting-ethics-in-the-age-of
  275. Federal Trade Commission. FTC Safeguards Rule: What Your Business Needs to Know. Federal Trade Commission, 2026-11-10. ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-busi
  276. Internal Revenue Service. Publication 5708: Creating a Written Information Security Plan for your Tax & Accounting Practice (Rev. 8-2024). Internal Revenue Service, 2024-08. irs.gov/pub/irs-pdf/p5708.pdf
  277. Federal Trade Commission. 16 CFR 314.4 - Elements. Legal Information Institute, Cornell Law School (e-CFR text), 2024. law.cornell.edu/cfr/text/16/314.4
  278. Federal Trade Commission. 16 CFR 314.5 - Effective date. Legal Information Institute, Cornell Law School (e-CFR text), 2024-05-13. law.cornell.edu/cfr/text/16/314.5
  279. U.S. Department of the Treasury. 31 CFR 10.22 - Diligence as to accuracy. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/31/10.22
  280. U.S. Department of the Treasury. 31 CFR 10.35 - Competence. Legal Information Institute, Cornell Law School (e-CFR text), 2014-06-12. law.cornell.edu/cfr/text/31/10.35
  281. David A. Wood and co-authors (Issues in Accounting Education 38(4): 81-108). The ChatGPT Artificial Intelligence Chatbot: How Well Does It Answer Accounting Assessment Questions?. Issues in Accounting Education (American Accounting Association); abstract on the Penn State research portal, 2023-01. pure.psu.edu/en/publications/the-chatgpt-artificial-intelligence-chatb
  282. Marc Eulerich, Aida Sanatizadeh, Hamid Vakilzadeh, David A. Wood. Is it all hype? ChatGPT's performance and disruptive potential in the accounting and auditing industries. Review of Accounting Studies (Springer), 2024. link.springer.com/10.1007/s11142-024-09833-9
  283. Sina Gogani-Khiabani, Ashutosh Trivedi, ShinPing Chyi, Saeid Tizpaz-Niari. Performance of LLMs on VITA test: potential for AI-assisted tax returns for low income taxpayers. Artificial Intelligence and Law (Springer); record in the NSF Public Access Repository, 2025. par.nsf.gov/biblio/10649839
  284. Intuit QuickBooks. Accountants embrace AI and strategic advisory services to fuel growth, despite tech and talent challenges, according to 2025 Intuit QuickBooks Survey. Intuit (accounting software vendor), 2025-04. quickbooks.intuit.com/r/news/intuit-quickbooks-tech-survey/ Secondary source.
  285. Karbon. State of AI in Accounting 2025 Report. Karbon (practice management software vendor), 2025. karbonhq.com/resources/state-of-ai-accounting-2025/ Secondary source.
  286. Internal Revenue Service, Office of Professional Responsibility. Introductory Guidelines for Responsible AI Use in Federal Tax Practice (OPR Alerts, Issue Number 2026-19). Internal Revenue Service (GovDelivery bulletin), 2026-06-24. content.govdelivery.com/accounts/USIRS/bulletins/41d6e70
  287. J. Michael Reese, J.D., LL.M.. Should I disclose my use of gen AI to clients?. Journal of Accountancy (AICPA & CIMA), 2025-04-01. journalofaccountancy.com/issues/2025/apr/should-i-disclose-my-use-of-g Secondary source.
  288. AICPA and CIMA. U.S. Accounting Firms Show Strong Growth in Profit and Revenue, AICPA and CIMA. AICPA & CIMA, FY2022. aicpa-cima.com/news/article/u-s-accounting-firms-show-strong-growth-in
  289. CPA Practice Advisor. Ignition Report Shows Shift in Pricing for Accounting Firms. CPA Practice Advisor (reporting Ignition's benchmark), 2025. cpapracticeadvisor.com/2025/08/21/ignition-report-shows-shift-in-prici Secondary source.
  290. Ignition. 2025 U.S. Accounting and Tax Pricing Benchmark. Ignition (Ignition Software), 2025. ignitionapp.com/us/2025-accounting-tax-pricing-benchmark Secondary source.
  291. The Florida Bar, Professional Ethics Committee. Florida Bar Ethics Opinion 24-1. The Florida Bar, 2024-01-19. www-media.floridabar.org/uploads/2024/01/FL-Bar-Ethics-Op-24-1.pdf
  292. Professional Ethics Committee for the State Bar of Texas. Opinion 685. Texas Center for Legal Ethics, 2020-01. legalethicstexas.com/resources/opinions/opinion-685/
  293. North Carolina State Bar. 2018 Formal Ethics Opinion 1: Participation in Website Directories and Rating Systems that include Third Party Reviews. North Carolina State Bar, 2018-07-27. ncbar.gov/for-lawyers/ethics-and-governing-rules/ethics-opinions/opini
  294. Utah Legislature. Utah Code 13-77-102 Liability for violation of consumer protection law. Utah State Legislature (le.utah.gov), 2026-05-06. le.utah.gov/xcode/Title13/Chapter77/C13-77-S102_2026050620260506.html
  295. State Bar of California / California Supreme Court. California Rules of Professional Conduct, Current Rules 2026, rule 7.1 (Communications Concerning a Lawyer's Services), Comment [4]. State Bar of California, 2026. calbar.ca.gov/Portals/0/documents/rules/Rules-of-Professional-Conduct.
  296. Supreme Court of Texas / State Bar of Texas. Texas Disciplinary Rules of Professional Conduct (Including Amendments Effective January 31, 2022), Rule 7.04. State Bar of Texas, 2022-01-31. texasbar.com/AM/Template.cfm?Section=Home&ContentID=27271&Template=/CM
  297. Supreme Court of Florida / The Florida Bar. Rules Regulating The Florida Bar, Chapter 4 (RRTFB October 22, 2026), Rule 4-7.19 Evaluation of Advertisements. The Florida Bar, 2026-10-22. www-media.floridabar.org/uploads/2026/11/2027_04-OCT-Chapter-4-RRTFB.p
  298. Washington Supreme Court. Washington Court Rules, RPC 7.2 [RESERVED]. Washington Courts, 2021-01-26. courts.wa.gov/court_rules/pdf/RPC/GA_RPC_07_02_00.pdf
  299. Washington Supreme Court. Washington Court Rules, RPC 7.3 Solicitation of Clients. Washington Courts, 2025-04-29. courts.wa.gov/court_rules/pdf/RPC/GA_RPC_07_03_00.pdf
  300. Washington Supreme Court. Washington Court Rules, RPC 7.1 Communications Concerning a Lawyer's Services, Comment [3]. Washington Courts, 2016-09-01. courts.wa.gov/court_rules/pdf/RPC/GA_RPC_07_01_00.pdf
  301. New York State Unified Court System (Appellate Divisions); reproduced by Cornell Legal Information Institute. N.Y. Comp. Codes R. & Regs. Tit. 22 § 1200.7.1 - Advertising. Cornell Law School Legal Information Institute, 2026-06-01 and earlier. law.cornell.edu/regulations/new-york/22-NYCRR-1200.7.1
  302. New York State Bar Association. Ethical Marketing of Your Law Practice: Understanding New York's New Rules on Lawyer Advertising. New York State Bar Association, 2026-06-01. nysba.org/events/ethical-marketing-of-your-law-practice-understanding- Secondary source.
  303. Colorado General Assembly. Colorado Session Laws 2026, Chapter 131 (SB26-189), Automated Decision-Making Technology. Colorado General Assembly (leg.colorado.gov), 2026-05-14. leg.colorado.gov/laws/session-laws/SB26-189/131/download
  304. Utah Legislature. Utah Code 13-77-105 Enforcement. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S105_2025050720250507.html
  305. Utah Legislature. Utah Code 13-77-104 Safe harbor. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S104_2025050720250507.html
  306. United States Congress. 47 U.S.C. 227: Restrictions on use of telephone equipment. Legal Information Institute, Cornell Law School (US Code text), 2026. law.cornell.edu/uscode/text/47/227
  307. Maryland General Assembly. Maryland Code, Courts and Judicial Proceedings, section 10-402. Maryland General Assembly, 2026-11-10. mgaleg.maryland.gov/mgawebsite/Laws/StatuteText?article=gcj&section=10
  308. Massachusetts General Court. Massachusetts General Laws, Part IV, Title I, Chapter 272, Section 99. Massachusetts Legislature, 2026-11-10. malegislature.gov/Laws/GeneralLaws/PartIV/TitleI/Chapter272/Section99
  309. Professional Ethics Committee for the State Bar of Texas. Opinion 705. Texas Center for Legal Ethics, 2025-02. legalethicstexas.com/resources/opinions/opinion-705/
  310. NYSBA Task Force on Artificial Intelligence. Report and Recommendations of the New York State Bar Association Task Force on Artificial Intelligence. New York State Bar Association, 2024-04-06. nysba.org/wp-content/uploads/2022/03/2024-April-Report-and-Recommendat
  311. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. National Institute of Standards and Technology, 2023-01. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
  312. OWASP GenAI Security Project. LLM06:2025 Excessive Agency. OWASP Foundation, 2025. genai.owasp.org/llmrisk/llm062025-excessive-agency/
  313. OWASP GenAI Security Project. LLM01:2025 Prompt Injection. OWASP Foundation, 2025. genai.owasp.org/llmrisk/llm01-prompt-injection/
  314. Randall A. Lewis, Justin M. Rao, David H. Reiley. Here, there, and everywhere: correlated online behaviors can lead to overestimates of the effects of advertising. Proceedings of the 20th International World Wide Web Conference (WWW 2011), 2011. gwern.net/doc/economics/advertising/2011-lewis.pdf
  315. Brett R. Gordon, Robert Moakler, Florian Zettelmeyer. Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement. Marketing Science, 2023. econpapers.repec.org/RePEc:inm:ormksc:v:42:y:2023:i:4:p:768-793
  316. Ron Berman. Beyond the Last Touch: Attribution in Online Advertising. Marketing Science (INFORMS), 2018. doi.org/10.1287/mksc.2018.1104
  317. Frederick F. Reichheld; W. Earl Sasser, Jr.. Zero Defections: Quality Comes to Services. Harvard Business Review (excerpt hosted by Bain & Company), 1990. bain.com/insights/zero-defections-quality-comes-to-services-harvard-bu Secondary source.
  318. OpenAI. Setting up your business on ChatGPT. OpenAI Help Center, 2026. help.openai.com/en/articles/20001635-setting-up-your-business-on-chatg
  319. James B. Oldroyd; Kristina McElheran; David Elkington. The Short Life of Online Sales Leads. Harvard Business Review, 2011. hbr.org/2011/03/the-short-life-of-online-sales-leads
  320. Karin V. Rhodes; Genevieve M. Kenney; Ari B. Friedman; Brendan Saloner; Charlotte C. Lawson; Daniel Chearo; Douglas Wissoker; Daniel Polsky. Primary care access for new patients on the eve of health care reform. JAMA Internal Medicine, 2012-11/2013-04. doi.org/10.1001/jamainternmed.2014.20
  321. Charles W. LeHew; Yi-Fan Chen; Patrick D. Smith; Brittaney B. Hill; Flavia F. Lamberghini; Alejandra Valencia. Access to restorative oral health care for children living in Illinois with Medicaid vs private dental insurance: An audit study. The Journal of the American Dental Association, 2023. doi.org/10.1016/j.adaj.2023.07.018
  322. Eve N. Dexter, Scott Fields, Rebecca E. Rdesinski, Bhavaya Sachdeva, Daisuke Yamashita and Miguel Marino. Patient-Provider Communication: Does Electronic Messaging Reduce Incoming Telephone Calls?. Journal of the American Board of Family Medicine 2016;29(5):613-621, 2009-2014. jabfm.org/content/29/5/613.full.pdf
  323. Spruce Health. Doctors Are Losing More Than an Hour a Day to Phone Call Overhead. Spruce Health, various. sprucehealth.com/blog/doctors-losing-hour-day-phone-call-overhead/ Secondary source.
  324. Lingli Wang, Ni Huang, Yili Hong, Luning Liu, Xunhua Guo, Guoqing Chen. Voice-based AI in call center customer service: A natural field experiment. Production and Operations Management (Wiley), via Crossref, 2023. doi.org/10.1111/poms.13953
  325. Gartner, Inc.. Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI For Customer Service. Gartner, Inc., 2023-12. gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds
  326. Consumer Financial Protection Bureau. Chatbots in consumer finance. Consumer Financial Protection Bureau, 2023-06-06. consumerfinance.gov/data-research/research-reports/chatbots-in-consume
  327. Entrepreneur. Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO. Entrepreneur Media, 2025-05. entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-mo Secondary source.
  328. Alec Tyson, Giancarlo Pasquini, Alison Spencer and Cary Funk. 60% of Americans Would Be Uncomfortable With Provider Relying on AI in Their Own Health Care. Pew Research Center, 2022-12. pewresearch.org/science/2023/02/22/60-of-americans-would-be-uncomforta
  329. Talker Research (commissioned by PatientPoint). Is your doctor stuck in 2011? (survey of 2,000 US adults on doctor visits). Talker Research, 2025-07. talkerresearch.com/dr-past-patients-say-doctor-visits-feel-outdated/ Secondary source.
  330. American Medical Association, Center for Digital Health and AI. 2026 Physician Survey on Augmented Intelligence. American Medical Association, 2026. ama-assn.org/system/files/physician-ai-sentiment-report.pdf
  331. Notable. 24/7 answers: Southwest General's success with HOPE AI voice. Notable Health, undated (page read 2026-11-10). notablehealth.com/customer-stories/southwest-general-health-center-hop Secondary source.
  332. Kung et al.. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digital Health, 2023. doi.org/10.1371/journal.pdig.0000198
  333. Nation's Restaurant News. McDonald's is ending its AI drive-thru test with IBM. Nation's Restaurant News, 2022. nrn.com/quick-service/mcdonald-s-is-ending-its-ai-drive-thru-test-with Secondary source.
  334. CBS News (MoneyWatch). Eating disorder helpline shuts down AI chatbot that gave bad advice. CBS News, 2023-05. cbsnews.com/news/eating-disorder-helpline-chatbot-disabled Secondary source.
  335. Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond. Generative AI at Work. National Bureau of Economic Research (Working Paper 31161), 2023. nber.org/papers/w31161
  336. arXiv:2605.14830. Agentic AI and Human-in-the-Loop Interventions: Field Experimental Evidence from Alibaba's Customer Service Operations. arXiv (preprint), 2026-05. arxiv.org/abs/2605.14830
  337. Federal Trade Commission. FTC Sues to Stop Air AI from Using Deceptive Claims about Business Growth, Earnings Potential, and Refund Guarantees to Bilk Millions from Small Businesses. Federal Trade Commission, 2025-08-25. ftc.gov/news-events/news/press-releases/2025/08/ftc-sues-stop-air-ai-u
  338. U.S. Bureau of Labor Statistics. Receptionists, Occupational Outlook Handbook. U.S. Bureau of Labor Statistics, 2025-05. bls.gov/ooh/office-and-administrative-support/receptionists.htm
  339. U.S. Bureau of Labor Statistics. Customer Service Representatives, Occupational Outlook Handbook. U.S. Bureau of Labor Statistics, 2025-05. bls.gov/ooh/office-and-administrative-support/customer-service-represe
  340. Federal Trade Commission. Civil Penalty Inflation Adjustments (Federal Trade Commission notice, document 2026-18853). Federal Register (public inspection copy), 2026. public-inspection.federalregister.gov/2026-18853.pdf
  341. KFF. Poll: 1 in 3 Adults Are Turning to AI Chatbots for Health Information, Equaling the Share Who Use Social Media for Health. KFF, 2026. kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-
  342. Pew Research Center. From Diagnoses to Treatments, Why Americans Use AI Chatbots for Health. Pew Research Center, 2026. pewresearch.org/wp-content/uploads/sites/20/2026/08/PS_2026.8.25_AI-ch
  343. Maayan Zohar Basteker. AI Search Stats 2026: Market Share, Referral, and Citation Data. Similarweb, 2026. aisearch.similarweb.com/blog/gen-ai-stats/ Secondary source.
  344. OpenAI. Searching the web with ChatGPT. OpenAI Help Center, 2026. help.openai.com/en/articles/9237897-chatgpt-search
  345. Federal Trade Commission. FTC Reopens and Sets Aside Rytr Final Order in Response to the Trump Administration's AI Action Plan. Federal Trade Commission, 2025-12-22. ftc.gov/news-events/news/press-releases/2025/12/ftc-reopens-sets-aside
  346. Luca. Reviews, Reputation, and Revenue: The Case of Yelp.Com. Elsevier BV, 2016. doi.org/10.2139/ssrn.1928601

Cite this paper

Mike Golliet. "Technical marketing for professional practices: how dental, medical, legal and accounting firms are found, chosen and reached when AI answers the question." Technical Marketing, November 10, 2026. https://technicalmarketing.ai/papers/technical-marketing-for-professional-practices/

You may quote from this paper with a link to this page. It describes rules and research in general terms and is not legal, medical, tax or financial advice. Found an error? Tell us and we will correct it and note the change here.

Short answers from the paper.

What is technical marketing?

In this paper, technical marketing is the design, construction and measurement of a firm's acquisition, intake, compliance and reporting as a single system, in which each stage has an owner, an input, an output and a count, so that a change at one stage can be traced to its effect on the next. The definition is ours. The phrase is searched about 210 times a month in the United States27, too rarely to assume a shared meaning. Chapter 3 compares it with creative marketing and with SEO, which it contains as one channel.

Full answer

Do AI assistants agree on which dentist, lawyer or accountant to recommend?

Mostly not, in our study. We put 199 local questions to ChatGPT, Gemini and Perplexity across 16 US cities on 2026-11-101. Of the 2,975 businesses they named, 4.1% were named by all three and 79.7% by only one1. For dentists, 3 of the 266 businesses named (1.1%) were named by all three, and for accountants none of 272 were1. Each question was asked once, and answers came through each assistant's API, which is a baseline and not what every signed-in user sees1. Chapter 4 gives the method and its limits.

How does a practice get more patients or clients?

The paper's argument is that being visible is one stage of three. A practice is found, then chosen, then reached, and a gain at one stage is lost if the next stage drops it. Chapter 3 sets out that pipeline, the four profession chapters apply it, and Chapter 12 treats the front desk, where a practice that has been found and chosen can still lose the patient or client. The published figures on unanswered calls come mostly from vendors and measure different things: in one vendor's data on dental offices, 31 of every 100 callers hung up before reaching anyone24, and in a test of 500 law firms, 40% answered the call23.

Is there such a thing as HIPAA compliant AI?

People search for that phrase about 6,600 times a month27, and the paper applies it to no tool. HHS describes a business associate as a person or organization that performs functions for a covered entity that involve creating, receiving, maintaining or transmitting protected health information55, and it says no Security Rule standard requires a covered entity to certify its compliance151. In our check of 28 tools small practices commonly use, 9 will sign a business associate agreement without a special tier, 15 only on certain plans and 4 offer none, according to the vendors' own pages54. This describes the rules in general terms. The practice's own counsel or compliance lead decides how they apply.

Do the same marketing rules apply to dentists, doctors, lawyers and accountants?

No. Each profession answers to its own mix of ethics codes, state rules and federal law, and health care adds the HIPAA rules on patient information. Chapter 10 sets the four side by side and reduces them to a short list of design requirements that one system can meet. It describes what the texts say and is not advice on how they apply to a given practice.

Find out what AI says about you.

A free 20-minute call. We ask the assistants your customers' questions and show you where you come up today.

No pitch unless you ask for one.