Which businesses do ChatGPT and other AI assistants name? Data from 709 test answers
In 709 answers from ChatGPT, Google AI Mode, Gemini and Copilot, the business being checked was named in 36, or 5.1%, and the four assistants named a median of 17 different businesses per question. The assistants rarely agreed, and all 42 pairs of repeated answers contained different lists of names.
In short
- The checked business was named in 36 of 709 answers (5.1%); in 18 of 31 checks no answer named it at all.
- For the same question, the four assistants together named a median of 17 different businesses, and the average overlap between two assistants' lists was about 11%.
- Asking the same question again gave a different list of names in all 42 repeated pairs, but the yes or no verdict for the checked business stayed the same in 38 of them.
- In this sample, 30% of Gemini answers included a source link, compared with 100% of ChatGPT and Copilot answers.
- This is a small sample from two days, mostly small online services and mostly in English, so read it as a snapshot, not a market average.
On this page
Our data In 709 answers from ChatGPT, Google AI Mode, Gemini and Microsoft Copilot, the business being checked was named in 36, or 5.1%. The assistants also disagreed: for the same question, the average overlap between two assistants' lists was about 11%, and all 42 pairs of answers from later repeat checks contained different lists.
This page reports the numbers from 31 checks run on CheckAI on 12 and 13 October 2026, how each one was counted, and what it can and cannot tell you. The figures are a snapshot from 13 October 2026; the "updated" date at the top shows when the text last changed.
Our data The headline figures:
- 5.1%of 709 answers named the business being checked (36 answers)
- 18 of 31checks in which no answer named the business at all
- 17different business names per question across the four assistants (median)
- 11%average overlap between two assistants' lists for the same question
- 0 of 42repeated questions that returned the same list of names
- 30%of Gemini answers carried a source link, against 97 to 100% for the other three
Sample and method
Each check takes one business, drafts customer-style questions from its homepage (the owner can edit them) and sends every question to the consumer versions of ChatGPT, Google AI Mode, Gemini and Microsoft Copilot, logged out. These are the answers the public versions give, not developer API responses.
Our data The 31 checks used 181 questions and requested 724 answers; 709 answers came back. The 15 missing answers (13 from Gemini, 2 from Copilot) are excluded when calculating shares of answers on this page. Demo checks are excluded.
Of the four services, only Google AI Mode offered a country setting in our setup. Our data It was set to the US in 21 checks, Switzerland in 6, Germany in 3 and the UK in 1. We could not set a country for ChatGPT, Gemini or Copilot, so we wrote the place into each question.
What counts as named
A language model reads each answer and lists the businesses it names, in the order they first appear; ChatGPT tags businesses itself and we take those tags as they are. The checked business counts as named when the answer mentions its brand name, another name the owner gave, or its domain. A mention is not necessarily a recommendation. Position is the business's place in that list. The AI visibility glossary explains these terms.
How sources were typed
Each website an answer links to is counted once per answer, by its main domain. We sort domains with fixed lists: directories and review sites (for example local.ch, Yelp and Trustpilot; the list also contains google.com and apple.com), communities (for example Reddit, YouTube, Facebook and Wikipedia), competitors (a site whose name matches a business named in the same answer) and the checked business's own site. Everything else goes into "editorial and other", a catch-all for media, blogs, guides and any site our lists do not cover.
What was left out
We also left out a comparison with Google search positions, because too few checks had that data. Names are compared after spelling is standardised, but variants of the same business name can still count separately. This may make the overlap figures below too low.
How rarely the checked business is named
Our data The business being checked was named in 36 of 709 answers, or 5.1%. Looked at per check, 18 of the 31 checks were named in no answer at all, 13 were named at least once, and in only 1 check, all four assistants named the business. For the median check, none of its answers named the business; the highest share for one check was 59%.
Our data When it was named, the business was usually not alone at the top. In half of those 36 answers, it was among the first two names listed, and it was listed first in 7 of them (19%).
Our data Share of answers naming the checked business, per assistant:
Copilot named the checked business least often. With counts this small, the order of the other three is uncertain: their estimated 90% ranges overlap widely (ChatGPT 4.6 to 11.0%, Google AI Mode 3.3 to 9.0%, Gemini 3.6 to 9.7%). Answers from the same check share the same business and questions, so the real uncertainty is probably larger than these ranges suggest.
How many names an answer contains, and how little the assistants agree
Our data A typical answer named between 4 and 6 businesses (the median per assistant was 4 to 6, the mean 4.5 to 6.1). Few answers named none: between 1.7% and 5.5% depending on the assistant. Taken together, the four assistants produced a median of 17 different business names for a single question (mean 18.3; counted over the 167 questions that all four answered).
To measure agreement we used overlap: of all the businesses either assistant named, the share named by both assistants. If ChatGPT names A, B and C, and Gemini names B, C, D and E, five businesses were named in total and two by both, so the overlap is 2 of 5, or 40%. We computed this for every question and pair of assistants, then averaged it.
Our data The average overlap across all pairs was 10.6%. By pair:
Our data Only 35 of the 167 questions that all four assistants answered (21%) had at least one business that all four named. A business named by one assistant was often missing from the others.
What happens when you ask the same question again
Our data Two of the businesses were checked a second time with the same questions, in a later check on the same day (about 40 minutes and about four and a half hours after the first). That gives 42 pairs of answers: same business, same question, same assistant, two different checks. In none of the 42 pairs was the list of named businesses identical, and the average overlap between the two lists was 24%.
Our data Whether the checked business was named changed less often. Whether the checked business was named or not stayed the same in 38 of the 42 pairs (90%), mostly because it was not named either time, and changed in 4. This is a very small repeat sample from two businesses, so read it as an illustration. Our guide on why AI answers change explains how to measure despite this.
What the answers cite
Our data The answers contained 3,044 website links, counted once per domain per answer, across 1,264 different domains. (The per-assistant table further down counts every link, including several to the same site, so its averages are higher.) 60% of those domains were cited only once, although a smaller number of sites may account for many citations.
| Type of site (our lists) | Citations | Share |
|---|---|---|
| Editorial and other (catch-all) | 2,201 | 72.3% |
| Competitors named in the same answer | 523 | 17.2% |
| Directories and review sites | 124 | 4.1% |
| Communities (Reddit, YouTube and similar) | 182 | 6.0% |
| The checked business's own site | 14 | 0.5% |
A note on Google AI Mode: it links to most sources through Google redirect addresses. We resolve those to the destination website, including for earlier answers since 13 October 2026, so the figures here count the real sites. The remaining google.com links are Google searches.
Our data In Google AI Mode, youtube.com appeared in 80 of 181 answers (44%), reddit.com in 46 (25%), and a link to a Google search in 38 (21%). The other assistants rarely did: ChatGPT linked YouTube in 5 of its 181 answers and Reddit in 4; Gemini and Copilot each did so in 1 or 2 answers at most.
Our data Gemini included links less often: only 29.8% of its answers carried a source link, with 1.1 links per answer on average. Every ChatGPT and Copilot answer in the sample had at least one link, and 96.7% of Google AI Mode answers did.
Our data Answers that named the checked business linked to its own site in 36% of cases, against 0.1% of answers that did not name it. This comparison is partly circular: an answer that links to a business's site may also show its domain in the text, and a domain counts as a mention. So read it as two things that tend to appear together, not as proof that one causes the other.
ChatGPT, Gemini, Copilot and AI Mode side by side
Our data Per assistant, over the answers that came back:
| Assistant | Answered of asked | Median words | Businesses per answer (mean) | Answers with a source link | Source links per answer (mean) | Median seconds |
|---|---|---|---|---|---|---|
| ChatGPT | 181 of 181 | 347 | 4.5 | 100% | 9.7 | 15.0 |
| Google AI Mode | 181 of 181 | 390 | 5.9 | 96.7% | 8.1 | 9.9 |
| Gemini | 168 of 181 | 346.5 | 6.1 | 29.8% | 1.1 | 15.7 |
| Copilot | 179 of 181 | 440 | 5.6 | 100% | 3.1 | 26.8 |
Seconds are measured from our request to the finished answer. They include our own technical overhead and do not measure the speed a person sees in the app. Copilot wrote the longest answers, ChatGPT named the fewest businesses per answer but linked the most, and Gemini named the most businesses while linking the least. The guide ChatGPT, Gemini, Copilot and Google AI Mode compared goes through each assistant.
What this means for a business
Advice These points follow from the numbers above. Treat them as things to check, not promises.
- Not being named is common. In our sample most checks got no mention at all, so a "no" is a starting point, not a sign that something is broken.
- Check more than one assistant. With an average overlap of about 11% between two assistants, a result from ChatGPT alone says little about Copilot or Gemini.
- Count across many answers. Lists changed on every repeat in our data, so one answer is weak evidence.
- Read who is named instead, and which sites are cited. For a bike shop in Bern or a tax adviser in Munich, the cited pages show which articles, directories and communities the assistants linked to. Whether your business is described accurately there is a reasonable thing to check.
- Expect fewer clues from Gemini. Fewer than a third of its answers included a source link.
If you want your own numbers, the free check on CheckAI runs 6 of your customer questions through ChatGPT and Google AI Mode; the Single Check and the plans ask all four assistants. The method page describes the counting in more detail.
Limits of this data
- Small. 31 checks of 27 businesses and 709 answers; about 170 to 180 per assistant.
- Not representative. Most businesses are small online services whose owners chose to run a check. Local shops, restaurants or trades may look quite different; see the separate tests for restaurants and plumbers.
- Mostly English. 25 of the 31 checks were in English, 6 in German.
- Two days. All answers were collected on 12 and 13 October 2026.
- Logged out, consumer versions. People who are logged in, with history or memory turned on, may see different answers.
- No country setting for three assistants. We could not set a country for ChatGPT, Gemini or Copilot, even for Swiss and German businesses; the place was written into the question instead.
- Counting by a language model. Names are extracted by a model, which can miss or misread a name, and two spellings of one business may count as two.
- Related answers. The 709 answers are not independent: each check contributes answers about the same business, so the uncertainty is larger than a simple count suggests.