technicalmarketing.ai Free AI audit

Report · November 2026

Does AI recommend the same businesses twice? 945 answers, asked five at a time.

We asked ChatGPT, Gemini and Perplexity the same 63 questions, such as "Who is a good accountant for a small business in Atlanta, GA?", five times each across 16 US cities and four professions. A business named in one answer was in another answer from the same assistant 55% of the time. Across its five answers an assistant named 13.5 different businesses per question, and 2.3 of them every time.

Key findings

  1. A second ask often drops you. A business in one answer was in another answer from the same assistant 55% of the time.
  2. The assistants differ widely. Perplexity repeated a business 78.9% of the time, Gemini 50.5% and ChatGPT 35.6%.
  3. 47.1% of businesses were named in only one of five answers, and 17% in all five.
  4. One answer shows 46% of the list. An assistant named 6.2 businesses in its first answer and 13.5 across five.
  5. First place held through all five answers in 30% of sets. On average an assistant led with 2.5 different businesses.
  6. Disagreement between assistants is larger than an assistant's own variation. Two answers from one assistant shared 43.2% of businesses, and two from different assistants 9.7%.
  7. Repeats did not bring the three together. After five asks each, 4% of all businesses named had been named by all three.

The same question, five times.

We put each of 63 questions to each assistant five times, with the same wording and the same model, on the same day. That makes 189 sets of five answers. A single answer named 6.3 businesses on average. Across its five answers to one question an assistant named 13.5 different businesses, and 2.3 of them appeared every time.

The table counts each business once per assistant and question, by how many of that assistant's five answers named it.

Named inBusinessesShareChatGPTGeminiPerplexity
All five answers43417%3.8%10.5%41.5%
Four of five2208.6%6.3%8.4%11.6%
Three of five27110.6%8.8%11.2%12%
Two of five42516.6%19.1%17.7%12.2%
One of five1,20447.1%61.9%52.2%22.7%

Named once. Named again?

Take a business that appears in one answer. We checked how often it also appears in another answer to the same question from the same assistant. Across the three assistants it did 55% of the time. For ChatGPT it was 35.6%, so a business in one ChatGPT answer was more likely to be missing from the next than to be in it. For Perplexity it was 78.9%.

AssistantModelBusinesses per answerDifferent businesses across five answersIn every answerNamed again in another answerShared by two answers
ChatGPTgpt-6-luna4.713.80.535.6%24.3%
Geminigemini-3.8-flash6.315.31.650.5%36.8%
Perplexitysonar7.711.54.878.9%68.4%

"Shared by two answers" is the businesses in both answers as a share of the businesses in either. The list also keeps growing with each ask. After its first answer an assistant had named 6.2 businesses on average, and after five it had named 13.5, so the first answer showed 46% of the businesses the assistant went on to name.

Different businesses named so far, per assistant and question
Answers so farAll three, averageChatGPTGeminiPerplexity
16.24.76.37.7
28.87.89.39.3
310.710.211.710.3
412.31213.810.9
513.513.815.311.5

The first business named.

The first name in an answer gets the most attention. An assistant put the same business first in all five answers for 57 of the 189 sets (30%). In 68% of sets one business led most of the answers. On average an assistant led with 2.5 different businesses across its five answers.

AssistantSame first pick in all five answersOne business first in most answersDifferent first picks, average
ChatGPT10%33%3.5
Gemini25%78%2.4
Perplexity56%94%1.6

Asking again against asking another assistant.

Two answers from the same assistant shared 43.2% of the businesses either one named. Two answers from different assistants shared 9.7%. So the gap between assistants that our first study found does not come from asking once. An assistant's own answers overlap about four times as much as answers from two different assistants.

Answers comparedBusinesses in both, of all in either
ChatGPT with ChatGPT24.3%
Gemini with Gemini36.8%
Perplexity with Perplexity68.4%
ChatGPT with Gemini10.3%
ChatGPT with Perplexity8.7%
Gemini with Perplexity10.1%

More asks did not bring the three together. In a single round, with one answer from each assistant, 2% of the businesses named were named by all three (the first study, across 23 kinds of business, found 4.1%), and in 74% of rounds they had no business in common. Counting every business an assistant named at any point in its five answers, 83 of 2,085 (4%) were named by all three.

What is countedBusinessesNamed by all threeShareNo business in common
One answer from each assistant5,0611032%74% of rounds
Businesses an assistant named in at least three of its five answers783192.4%50 of 63 questions
Every business named in any of the five answers2,085834%20 of 63 questions

By profession.

Answers about family law attorneys changed least: a business in one answer was in another 63% of the time. Answers about primary care doctors changed most (45.5%).

ProfessionSets of answersBusinesses per answerDifferent across five answersIn every answerNamed again in another answerSame first pick in all five
Family law attorneys456.412.2363%44%
Accountants486.312.32.560.4%31%
Dentists486.314.32.251.6%27%
Primary care doctors48615.21.645.5%19%

By city.

Each city has 12 sets of answers at most (four professions, three assistants), so read a single row as a small sample.

CitySets of answersNamed again in another answerShared by two answersSame first pick in all five
Portland, OR1261.1%49.4%17%
Boston, MA1260.2%47.2%25%
Las Vegas, NV1259%47.3%42%
Austin, TX1258.7%47.8%33%
Houston, TX1258.4%45.5%25%
San Diego, CA1257.3%46.2%25%
Denver, CO1257.1%44.8%33%
Dallas, TX1255.3%42.9%33%
New York, NY1254.5%41%33%
Seattle, WA1252.9%41%33%
Charlotte, NC952.6%42.4%44%
Chicago, IL1252.2%40.2%25%
Los Angeles, CA1251.9%41%42%
Miami, FL1251.9%39.7%8%
Phoenix, AZ1251.6%40.4%42%
Atlanta, GA1245.1%34%25%

The websites behind each answer.

Every answer came with source links. Two of Perplexity's answers to the same question shared 96.5% of the websites cited, and 92.4% of the sites it cited for a question were cited in all five answers. ChatGPT read a different set nearly every time: two of its answers shared 30.6% of their sites. Perplexity's lists of businesses still moved, though: two of its answers shared 68.4% of the businesses named. The same sources did not produce the same list.

AssistantWebsites in both answers, of all in eitherWebsites cited in all five answers, of all cited for the questionBusinesses in both answers, of all in either
Perplexity96.5%92.4%68.4%
Gemini39.3%13%36.8%
ChatGPT30.6%5.3%24.3%

What this means for a practice.

This part is our reading of the numbers above, not a measurement.

One ask is not a measurement

A business in one ChatGPT answer was in another 35.6% of the time. Ask a question several times before you conclude that an assistant does or does not name you.

Record a rate

Write down how many of five answers named you, for each assistant, each month. Going from one in five to four in five is a change. A single appearance could be either.

Each assistant is still its own channel

After five asks each, 4% of the businesses named had been named by all three. Repeating the question does not make one assistant stand in for another.

First place moves

An assistant kept one first pick through all five answers in 30% of sets. A screenshot of first place records one answer.

How we did this, and its limits.

Method

  • Each question had the form "Who is a good accountant for a small business in Atlanta, GA? Name specific businesses.", with the city and the profession changed: four professions in 16 cities.
  • Each question went to each assistant five times as separate requests with no conversation history, through each assistant's API with web search turned on (ChatGPT: gpt-6-luna, Gemini: gemini-3.8-flash and Perplexity: sonar), using DataForSEO's LLM Responses API. These are the models and the wording of the first study.
  • All 959 answers were collected on November 11, 2026 within about 20 minutes, and cost $16.77 in API fees. One question (family law attorneys in Charlotte, NC) is left out because one of its 15 answers did not come back, which leaves 63 questions and 945 answers.
  • We recorded the business names in each answer, up to eight per answer, and the websites each answer cited. Within a question a business counts as the same one when the spelling matches or one name is the other with words added at the end.
  • "Named again" takes each business in an answer and asks whether another answer in the same set names it, averaged over every pair of answers. "Shared by two answers" is the businesses in both divided by the businesses in either, averaged the same way.

Limits

  • The five answers to a question were collected minutes apart. Answers a week or a month apart may differ more, because the pages an assistant reads change.
  • We do not know why Perplexity cited nearly the same websites each time. If its search results are held for a short period, a repeat on a later day could look less stable.
  • We kept at most eight names per answer. An assistant that lists more than eight in a changing order looks less stable here than it is.
  • Names are matched by spelling. A business listed under two different names counts as two, which lowers every overlap figure.
  • The apps personalize by account and location. An answer through the API is a clean baseline, not what every person sees.
  • These are four professions in 16 large cities, on one day, with these models. Other trades, smaller towns and later models may behave differently.
  • Assistants can be wrong. Being named is not our endorsement, and not being named says nothing about a business's quality.

Cite this report

Technical Marketing. "Does AI recommend the same businesses twice? 945 answers, asked five at a time." November 11, 2026. https://technicalmarketing.ai/reports/ai-answer-consistency/

You may quote figures and tables from this report with a link to this page. The data behind it: stability by question and assistant (CSV), how many answers named each business (CSV). Found an error? Tell us and we will correct it and note the change here.

Questions people ask about this.

Does ChatGPT give the same answer every time?

No. In our November 2026 study we asked ChatGPT each of 63 local questions five times. Two of its answers to the same question shared 24.3% of the businesses named, and 3.8% of the businesses it named for a question appeared in all five answers.

Which AI assistant gives the most consistent recommendations?

Of the three we tested, Perplexity. A business in one of its answers was in another 78.9% of the time, against 50.5% for Gemini and 35.6% for ChatGPT. Consistent does not mean correct: we did not judge whether the businesses named were good choices.

How many times should I ask an AI assistant whether it recommends my business?

More than once. After its first answer an assistant had named 6.2 businesses on average, and after five it had named 13.5, so the first answer showed 46% of them. Five asks per assistant gave us a rate we could compare. We did not test whether more than five changes the picture.

Do AI assistants agree more when you ask them several times?

Not in our data. With one answer from each assistant, 2% of the businesses named were named by all three. Counting every business each assistant named across five answers, it was 4%. Two answers from different assistants shared 9.7% of their businesses, against 43.2% for two answers from the same assistant.

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.