Three AIs, one question, the same minute: ChatGPT named no business at all

This whole series changed one thing at a time: web search, the phrasing of the question, the pages read. The engine stayed constant. This morning we did the opposite: the same question, at the same minute, sent to all three assistants in parallel. Six cities, six trades. Nothing changes but the AI you are talking to.

The result we did not plan for

ChatGPT named no business. Not one, in any of the six cases. Asked for the best plumbers in Toulouse, it answers “Search Google Maps and the yellow pages”, “Check the legal notices”, “Insurance and company number”. On lawyers in Strasbourg: “Selection criteria”, “Languages spoken”, “Who will work on my case”. On dentists in Nice, it asks a question back and stops there.

This is not one polite refusal in one odd case: it is the same answer six times over, across six trades and six cities. The most-used assistant in the world, on the most ordinary question a customer can ask, recommends nobody.

Six cities, six trades, one question: ChatGPT cited no local business. It explains how to look for one.

And the other two barely overlap

That leaves Gemini and Perplexity, which do name businesses. Across the six cases, 42 different businesses were cited. None is named by all three engines - necessarily, since the third names none - and 33 out of 42, or 79%, are named by only one.

The overlap between the two engines that do answer varies enormously by trade. On restaurants in Lille they share four names out of twelve; on hairdressers in Bordeaux, three out of eleven. But on dentists in Nice, Perplexity cites seven practices and Gemini none - it answers with booking platforms and selection criteria. On lawyers in Strasbourg, Gemini names none either.

Translated for a business owner: “I am visible on AI” means nothing until you say which one. You can be cited by Perplexity and entirely absent from Gemini on the same day, at the same hour, on the same question - with nothing having changed on your side.

A figure we threw away

Our automated count first claimed 109 businesses, with 93% cited by a single engine. We did not publish it. Re-reading the lists, most of the “businesses” were not businesses at all: they were instruction sentences that the extractor takes for proper nouns - “Check the legal notices”, “Selection criteria”, “Response time and availability”.

That is the third time this week the same trap has closed, and it teaches something: when an assistant answers with advice rather than names, a naive measurement counts the advice as competitors. The 42 businesses above were confirmed one by one, by hand. The spectacular figure was false; the one that remains is verified.

What the market measures, from another angle

Our finding meets, by an entirely different route, what SOCi's 2026 local visibility index reports: across more than 350,000 locations in 2,751 brands, ChatGPT reportedly recommends only 1.2% of them. In other words, the overwhelming majority of local businesses does not exist in its answers - which is exactly what six questions on an August morning made visible.

The same body of analysis puts the use of AI for a local search at 45% in 2026, against 6% a year earlier. A channel multiplying sevenfold in a year, in which one major assistant names almost nobody: that is where the game is being played.

What to do with this

• Stop speaking of “AI visibility” in the singular. It is three measurements, and they do not converge: your presence on one says nothing about the other.
• Conclude nothing from a single ChatGPT test. If it does not cite you, that may simply mean it cites nobody in your trade - testing it alone teaches you nothing about yourself.
• Look first at the engine that cites your trade. Across the six cases, Perplexity named businesses in all six; Gemini in four. Start where there is a list to get into.
• Beware of “competitors” that are not. If a report shows you “Selection criteria” as a rival, it is counting advice, not businesses.

The limits, as always: six questions, one country, one morning, one model per engine. This is not a census. But the protocol - same question, same minute, three engines - removes the easy explanation: it is not the day, not the phrasing, not luck. It is the engine.