Who ChatGPT and Gemini name: 83 small businesses, 498 answers
Measured on 15 and 16 November 2026
We took 83 small businesses and, for each one, put three questions a buyer might ask to ChatGPT and to Gemini. The questions did not mention the business. Then we counted how often it came up in the answer.
19 of the 83 were never named. Gemini named them more often than ChatGPT did. And for four businesses in ten, neither assistant used a single page of the business’s own site.
How often a business was named
Of the 498 answers, 187 named the business. That is a little more than one answer in three. Here is how the 83 businesses divide, by the number of their six answers that named them.
- 0 of 6 answers19 businesses
- 1 of 6 answers12 businesses
- 2 of 6 answers14 businesses
- 3 of 6 answers15 businesses
- 4 of 6 answers14 businesses
- 5 of 6 answers8 businesses
- 6 of 6 answers1 businesses
Each mark is one business. The filled ones were named at least once.
45 of the 83, more than half, were named in two answers or fewer. One was named in all six.
Is that a lot or a little? The same answers give one yardstick. In each set of six answers we noted the other business that was named most often. That name came up in 298 of the 498 answers, against 187 for the business we were checking, and in 49 of the 83 checks it came up in more answers than the business itself. For the 19 that were never named, it came up in 75 of their 114 answers. The assistants did answer these questions with names. They were other names.
Gemini named them more often than ChatGPT
| Assistant | Answers that named the business | Businesses it never named |
|---|---|---|
| ChatGPT | 81 of 249 answers (33%) | 35 of the 83 |
| Gemini | 106 of 249 answers (43%) | 24 of the 83 |
16 businesses were named by Gemini and never by ChatGPT. 5 were named by ChatGPT and never by Gemini. So for one business in four, one assistant named it and the other never did.
Taken one question at a time, both assistants named the business in 60 of the 249 questions, only Gemini in 46, only ChatGPT in 21, and neither in 122. That last number is almost half.
Shops and software companies
| Kind of business | Named, both assistants | ChatGPT | Gemini | Never named |
|---|---|---|---|---|
| 56 online shops | 131 of 336 answers (39%) | 60 of 168 (36%) | 71 of 168 (42%) | 10 shops |
| 27 software companies | 56 of 162 answers (35%) | 21 of 81 (26%) | 35 of 81 (43%) | 9 companies |
The two assistants were furthest apart on software. ChatGPT did not name 16 of the 27 software companies in any of its answers.
The answers about software also kept to a leading name. Check by check, the other company named most often came up in 140 of the 162 software answers. For the shops, the other name given most often came up in 158 of 336.
It was not the robots.txt
The check also reads each site’s robots.txt and its home page as the server sends it. None of the 83 had a rule that turns away the crawlers the assistants use to fetch pages for an answer. None sent a home page that stays empty until a script runs. A site with no robots.txt counts here as having no rule.
Those are the two faults people look for first. They explain none of the 19.
Where the answers came from
For each answer we keep the pages the assistant read or linked to. For 34 of the 83 businesses, the business’s own site was not among them in any of the six answers. The questions had been written from those businesses’ own pages, and the assistants answered them from other sites.
Those 34 were named in 25 of their 204 answers, about one in eight, and 19 of them never. The other 49 had their own site among the sources at least once. They were named in 162 of their 294 answers, and every one of them was named at least once.
Part of that gap comes from how we count. An answer that links to a page of the business counts as naming it, so a business whose site is used is named almost by definition. What the count does show is how often the site was not used at all. It also shows that a name can arrive without it: 15 of the 34 were named from what other sites say about them.
What did the assistants read instead? For each of the 34 we noted the site they read most. It was Reddit for 6 of them and Tom’s Guide for 5. For the rest it was a different site nearly every time: a magazine, a larger retailer, the website of a competitor. In all but one of the 83 checks the six answers drew on at least 13 other sites, and in 75 they named at least twelve other businesses.
What to take from it
Three things follow for a small business. Ask more than one assistant: an owner who asks one and finds nothing has not learned what the other one says. Ask what a buyer asks, with your own name left out, because a question that contains your name tells you little about a buyer who has not heard of you. And look at the sources under the answer. If your own site is never among them, the first job is on the pages the assistant does read: the list, the review, the thread.
We have written up the causes one at a time in why a website does not appear in ChatGPT. Our free check asks three such questions about your business on ChatGPT and Gemini and shows you the answers. There is no sign-up and it does not ask for an email address.
How we counted
- The 83 are 56 online shops and 27 software companies in 21 countries. 61 of them are in the United Kingdom, the United States, Australia, Ireland or Canada. We do not name them.
- We picked them by hand, before any check was run, as businesses to write to about this check. Each is small and independent, sells from its own site and publishes a contact address. Large chains were left out. So this is not a random sample, and it leans one way: 34 of the 56 shops sell food or drink.
- We started with 100. 16 checks did not finish and one came back with four answers instead of six. Those 17 are left out.
- For each business the check read its pages and wrote three questions a buyer might ask, without the name of the business in them. Each question went once to each assistant through its official API, with web search switched on and no extra instructions. An answer from the API can differ from what the same assistant shows inside its own app.
- An answer counts as naming a business when the name is in the text of the answer, or when the answer links to one of the business’s own pages.
- The names of other businesses are read off each answer by a helper model, up to eight an answer. The count for the business itself does not rest on that model. The other business named most often is the highest of many, so it sets a high bar.
- Three questions asked once are a small sample for any one business, and answers change from one day to the next. Read the figures as a picture of these 83 businesses on those two days.
- We make this check and sell a longer version of it, so we have an interest in what it finds. The method is set out here so that you can weigh the figures yourself.