Study #1 · 615 AI answers
Do AI answers change every time you ask?
The short answer
Almost always. We measured 615 AI answers across two B2B markets in September 2026. Asking the same engine the same question three times returned the identical list of brands 2 times out of 205. A single answer names about 6 brands. The same question, across 5 engines and 3 runs, surfaces about 45.
Published September 20, 2026 · measurements taken September 15 and 17, 2026 · written by Thibault Genon-Catalot, founder
What we measured
Two markets, two continents, 615 recorded answers.
Two audits run in September 2026 for two B2B companies: theTribe, a custom software company in Nantes, France, and Tessera, an event ticketing platform in Singapore. 2 markets that share nothing, in 2 languages. Every answer was taken through the engines' developer interfaces, timestamped and stored with the sources each engine cited.
| Study A | Study B | |
|---|---|---|
| Market | Custom software development, France | Event ticketing, Singapore |
| Company audited | theTribe, Nantes | Tessera, Singapore |
| Buyer questions | 21 | 20 |
| Engines | 5 | 5 |
| Runs per question per engine | 3 | 3 |
| AI answers recorded | 315 | 300 |
| Distinct brands named | 686 | 544 |
| Total namings | 1,951 | 1,928 |
The five engines: ChatGPT, Claude, Perplexity, Gemini and Google AI Overview. 615 answers in total, 1,226 distinct brands named across the two markets, 3,879 namings. This study reports how the engines behave. It says nothing about how visible either company is: that number belongs to them and stays in their own report.
The picture in one chart
Almost every brand lives inside a single engine.
Read it this way: of every 100 brands an engine names in a market, 74 are named by that engine and by no other. Track one engine and you are looking at one fifth of the market, with almost no overlap to borrow from. Counted per market, so a brand present in both markets counts once in each.
The findings
What changes between two identical questions.
The same question, asked three times, does not return the same brands
For each question and each engine we ran three passes. The list of brands came back identical 2 times out of 205. On average only 13% to 23% of the brands seen for a question survived all three runs. One screenshot of one answer tells you nothing about your position.
The engines barely agree with each other
Of the 1,226 brands named, 74% were named by a single engine out of five. Between 2% and 4% were named by all five. Being visible in ChatGPT says almost nothing about being visible in Perplexity, and a tool that tracks one engine is tracking one fifth of the picture.
A buyer sees 6 brands. The engines know 45
The average answer names 6.2 brands, median 6. The same question, across 5 engines and 3 runs, surfaces about 45 distinct brands. So a buyer who asks once sees 1 brand in 7 of those the engines are willing to name. Which 7th they get is close to a lottery.
Most brands are named once and never again
57% of the brands named appeared in exactly one answer out of the 300 or so in their market. The top 5 brands took 14% and 20% of all namings in their market. The head is small and it repeats, and that is the part worth fighting for.
Being first is not a position you hold
The same brand was named first in all 3 runs 50 times out of 205. The other 155 times, first place changed between 2 identical requests sent a minute apart.
How to read any number
How to judge an AI visibility number someone shows you.
A single test proves nothing
If an agency shows you one screenshot of ChatGPT naming you, that is one of the six brands the engine happened to name that second. Ask what happens on the second and third pass.
One engine is not the market
74% of brands live in exactly one engine's answers. Pick the engines your buyers actually use, and measure all of them. Published research agrees: 86% of the top 50 most-mentioned sources are not shared across ChatGPT, Perplexity and Google AI Overviews (Ahrefs, June 2025).
Ask for the denominator
Any number about AI visibility is meaningless without the count of answers behind it and the date it was taken. "You appear in ChatGPT" is not a measurement. "Named in 41 of 300 answers, measured on 15 September" is.
Aim at the head, not the tail
Appearing once in 300 answers is what happens by default. The work is to move into the handful of brands the engines repeat, which means being present in the sources those engines read.
The limits
What 615 answers cannot tell you.
It does not say which brands the engines never name, because we only measured brands that appeared at least once. It does not generalise to B2C, to languages beyond French and English, or to markets shaped by one dominant platform: in one of the two markets a single well-known platform took 52% of all the answers on its own, which compresses everything below it. Two markets is two markets. We publish the method so anyone can repeat it, and we will publish the third and fourth markets as they come.
Our own line, for the record: we run the same measurement on ourselves every week, 20 buyer expressions across five engines (27 questions until September 20). In the week of September 14, 2026 we were named in 0 of 135 answers, the same as the day we started on August 31. The weekly reading is on /share-of-answer, and what a visibility score can and cannot measure is at /ai-visibility-score.
Questions
Fair questions about this study. Straight answers.
Why measure the same question three times?
Because once tells you nothing. Across the 615 answers in this study, the same engine asked the same question three times returned the identical list of brands 2 times out of 205. Only 13% to 23% of the brands seen for a question survived all 3 passes. A single run is one draw from something that keeps moving.
Does this mean AI visibility cannot be measured?
It can. It cannot be measured once. A share built on 300 timestamped answers is stable enough to compare month to month, which is why our audits run 20 questions across 5 engines 3 times each. What cannot be measured is a screenshot.
Are the engines really that different from each other?
Yes. 74% of the 1,226 brands named in this study were named by a single engine out of five, and between 2% and 4% by all five. Published research points the same way: 86% of the top 50 most-mentioned sources are not shared across ChatGPT, Perplexity and Google AI Overviews (Ahrefs, June 2025), and only 12% of the URLs cited by ChatGPT, Gemini and Copilot rank in Google's top 10 for the same prompt, against 76% for Google AI Overviews (Ahrefs, August 2025).
Which brands do the engines never name?
This study cannot say, and we would rather say so than guess. We counted the brands that appeared at least once; a brand that appeared zero times in 615 answers is indistinguishable from a brand that does not exist in that market. What we can say is that appearing once is the normal state: 57% of the brands named showed up in exactly one answer.
Can I see the raw answers?
Yes. Every answer is timestamped and stored with the sources the engine cited, and the two audits behind this study were locked on 19 September 2026 and have not changed since. Ask and we will show you the per-question counts. The method is public at /methodology.
Free, 2 minutes
Before you buy any audit: run one free check.
5 buyer questions, 2 AI engines, and you see who gets named instead of you. The verdict shows on the page, and a copy lands in your inbox.
For journalists and analysts
Take the numbers. Here is how to credit them.
Copy this line
NextChannel Studio, "Do AI answers change every time you ask?", 615 measured AI answers across two B2B markets, September 2026. https://nextchannelstudio.com/do-ai-answers-change-every-time
Every figure on this page comes from 2 audits delivered on 18 September 2026 and locked the day after, stored as raw answers with the sources each engine cited. Ask and we will send the per-question counts. If you find an error, tell us and we will correct the page with a dated note.
Where to go next
Run the same measurement on your own market.
The $1,500 audit is this study run on your buyers' questions: 10 questions per territory (5 to 7 territories), five engines, three passes, 750 to 1,050 timestamped answers, delivered in 48 hours without a call. The method is at /methodology, what the rest of the market charges is at /geo-price-comparison, and the sourced market numbers are on /ai-search-statistics. Journalists and analysts: the per-question counts are available on request.