In my last post I gave you a test. Ask an AI assistant where to buy what you sell, and see who it names. I told you to run it once.
That advice is good but was incomplete. One run tells you almost nothing, and this week I can show you why.
I took the same test and repeated it. Same question, same country, 11 times per question on ChatGPT, spread over five days. Then I compared it with Perplexity and with Google page 1.
The short version: the list of brands an AI names is not a ranking. It is closer to a draw from a pool, and how big your share of the pool is matters more than whether you showed up in one screenshot.
Below you get the numbers, what I would and would not conclude from them, and a method you can run on your own store in about 10 minutes.
What I tested
This time I tested the way a UK buyer would ask.
Location: London IP address, verified before every session
Three buyer questions, one per niche:
where to buy designer radiators online UK
where to buy PPE and workwear in bulk UK
where to buy commercial catering equipment online UK
ChatGPT: Temporary Chat (no history, no memory), web search on, fastest model setting, every run in a fresh chat. 11 runs per question, 33 in total, between 8 and 12 October 2026
Perplexity: one run per question per session
Google: the unpaid results on page 1, with Google set to the United Kingdom and English, and personalisation switched off so my own search history could not influence what I saw. I captured them next to the AI runs
I counted the brands each answer named, which pages ChatGPT cited as sources, and how much two answers to the same question overlapped.
Finding 1: ask twice and you mostly get different answers
To measure this I used overlap: of all brands named in two answers combined, how many appear in both? 100 percent would mean identical answers. I averaged it over every pair of runs.
Designer radiators: 29 percent overlap. 19 different brands appeared across 11 runs. No brand appeared in every run
PPE and workwear in bulk: 31 percent overlap. 22 different brands. No brand in every run
Commercial catering equipment: 60 percent overlap. 11 different brands. Three brands appeared in all 11 runs
So the answer to “does ChatGPT give a stable answer?” depends on the market. Catering is concentrated around a few large retailers, and those retailers showed up every time. Radiators and PPE are fragmented, with many small suppliers, and the names rotated from run to run.
Here is the uncomfortable part for a small or mid-sized store. If your market looks like radiators or PPE, a single screenshot of ChatGPT naming you is luck as much as visibility. A single screenshot of it not naming you is luck too.
Finding 2: being on Google page 1 did not predict being named
In PPE, not one brand that ChatGPT named matched a Google page 1 result, in any of the 11 runs. Zero. The suppliers ChatGPT recommended and the suppliers Google ranked were two different groups.
In catering the picture was different: between two and four of the named brands matched page 1 in each run. In radiators, none to two.
I would not conclude that Google rankings do not matter. I would conclude that you cannot assume page 1 carries over to AI answers. You have to measure them separately.
One more detail. A supplier that sat on Google page 1 on 8 and 9 October was named by ChatGPT once, after it had already dropped off page 1. Ranking and being named did not move together.
Finding 3: what ChatGPT cites is not what you would guess
I read the titles of the pages ChatGPT used as sources. A lot of them were:
product pages (in radiators)
“About us” pages (in PPE and catering)
news and trade press articles, for example about a retailer opening a new shop (in catering)
In catering, an unrelated recruitment agency page appeared as a source in 4 of 7 runs. I have not read why, and I will not guess.
What this suggests: the page that earns the mention is often not the category page you spent a year optimising. It may be a product page, an About page, or something written about you elsewhere. That is a pattern in the titles I saw, not a measured rule.
Finding 4: the engines disagree with each other and with themselves
Perplexity gave different lists than ChatGPT for the same question. One retailer named in all three Perplexity catering runs on 8 October was missing from both runs on 9 October, and back on 12 October. I only have one Perplexity run per session, so I treat this as a hint that it drifts, not as a measurement.
Time of day showed no visible effect on ChatGPT. Answers were more alike inside one day than between days (for radiators, 54 percent overlap inside 12 October, but 21 to 30 percent against earlier days). Sample is small, so I would test it before repeating it.
What I would not conclude
Be careful with anyone, including me, who turns this into a rulebook.
This is a pilot: 3 niches, 1 country, 4 ChatGPT sessions over 5 days. It is not a rate for your market
Brand names in an answer are not the same as cited pages. I read the page titles, not every page
Nothing here explains why a brand gets named. I checked page rendering, robots.txt, structured data and FAQ text on a dozen pages. Those checks did not separate cited from uncited pages. The only thing that lined up with an absence was a robots.txt block on one site, and two other fully open pages were also absent
I did not test Bing, and I will not claim anything about how Bing relates to ChatGPT from this data
If someone sells you “the 7 things AI prefers”, ask for the repeated runs behind it.
The 10 minute test for your store
This is the method I would use on your store. It gives you one number you can track.
Step 1. Pick 3 buying questions. Write them the way a buyer would, not the way you title your pages. Examples: “where to buy [your product] online [your country]”, “best [your product] for [use case]”, “who sells [your product] in bulk”.
Step 2. Set up a clean ChatGPT. Open a Temporary Chat, turn web search on. If you sell to another country, use a VPN in that country, because the answers are location sensitive. Start a new chat for every run.
Step 3. Run each question 5 times. Space them across the day if you can. After each run write down three things:
Were you named? (yes or no)
In which position?
Which page was cited for you, if any?
Step 4. Calculate your AI Visibility Rate. Appearances divided by runs. Named in 3 of 5 runs means 60 percent.
Step 5. Repeat once on Perplexity and write down Google page 1 for the same questions. Now you have three columns: ChatGPT, Perplexity, Google.
Step 6. Read the result. This is my rule of thumb, not a standard:
Named in 0 of 5: you are not in the pool for that question
Named in 1 or 2 of 5: you are on the edge of the pool. You are visible sometimes, which is also why one screenshot can mislead you
Named in 3 or 4 of 5: you are a regular candidate
Named in 5 of 5: you are an anchor name, like the three catering retailers in my test
What to do with each result
If you are named in 0 or 1 of 5:
Check that AI crawlers can reach you. Open yourdomain.com/robots.txt and make sure you are not blocking GPTBot, OAI-SearchBot or PerplexityBot by accident
Look at which pages the AI cited for your competitors. Open them. Ask what they say that your equivalent page does not
Make sure the pages a buyer would land on from an answer (product pages, About page) say plainly who you are, what you sell, where you deliver, and what makes you different, in text, not only in images or scripts
If you are named in 2 to 4 of 5:
Find the runs where you were missing and see who was named instead. That is your real competitor set in AI answers, and it can differ from the one you see on Google
Run the same test again in two weeks, so you know whether you are moving or only seeing noise
If you are named in 5 of 5:
Protect it. Keep the cited pages live, fast and accurate, and do not redirect or remove them without checking
For everyone:
Do not judge a change from one run. Compare a rate from 5 runs before and after
Keep the log. A simple spreadsheet is enough
One more source worth checking
Microsoft expanded the AI Performance report in Bing Webmaster Tools on 1 July 2026, with four views: Intents, Topics, Citation Share and Compare. If your site is verified there, open it and look at Citation Share and Intents. It is the one place where a search engine reports citation data about your own pages directly, instead of you sampling it by hand.
The takeaway
One AI answer is a sample of size one. My 33 runs say that for fragmented markets the names rotate, Google page 1 does not carry over automatically, and the pages that get cited are not always the ones you would bet on.
So stop treating AI visibility as a screenshot and start treating it as a rate. Run the 5 run test this week, write down your number, and run it again in two weeks.
If you would rather have it done for your store, I will run it for you. I check your three most important buying questions on ChatGPT and Perplexity, compare them with Google, and send you what I find.
If you decide to run the test yourself, send me your results at neven@lobit.agency. For each of your three questions, how many times out of 5 were you named? I will look at them and help you work out how to appear more often in AI answers.
I reply to every email.
Kind regards
Neven Lovreković

