In a town of 10,201 people, two AIs out of three name your bakery
48 answers, 8 communes, from Paris to a sub-prefecture of 10,201 people. Gemini and Perplexity name real bakeries everywhere - with the address. ChatGPT answers without a question only in Paris.
In this article
A study published this month puts a number on the demand: 23% of people used an AI tool during their most recent local-business search, and 31% use one at least monthly for a recommendation. Another, in May, argued that smaller geographies are easier to break into, the competitive field being thinner.
Neither answers the question a small-town business actually asks: is my town too small for the AI to know anyone in it?
We could not answer it either. Our three national studies cover 39 cities, and the smallest is Grenoble. So last night we measured where our customers actually live.
The protocol
Eight communes on a gradient from the capital down to Saint-Amand-Montrond, 10,201 inhabitants, by way of Nantes (335,634), Chartres (38,209), Vierzon (24,694) and Guéret (12,839). Two trades - bakery and car garage - so that no result hangs on bread alone. Three assistants. The question is identical everywhere: only the commune changes.
48 answers, plus 32 control draws on ChatGPT: without re-asking, you cannot tell an effect from a coin toss. That control overturned our first reading, and it is worth saying what it corrected.
The good news, and it is clear-cut
Gemini and Perplexity name real local businesses everywhere, down to 10,201 people. Zero answers without a name out of twelve. And not generalities: actual shops, with the address.
Saint-Amand-Montrond
: Aux 3 Choux, Maison Aveline, Aux Épis d'Or, Au Merle Blanc, La Chaumière, Pâtisserie-Chocolaterie Franck Cosyns.
Guéret
: J&F Boulangerie Pâtisserie, avenue de la Sénatorerie - named by both engines, with the same address, Perplexity stating it got it from the Petit Futé guide.
A shop in a sub-prefecture of ten thousand people is therefore perfectly nameable. The fear of being “too small for AI” does not survive measurement - provided you look at the engines that go and search the web.
ChatGPT is another story
Across 32 draws spread over 8 communes, exactly one never gets a question back: Paris. Everywhere else - Lyon included - ChatGPT answers one or two times in four with a question instead of a list:
“Would you prefer bakeries in the centre of Guéret, or shall I include the chains? Do you also want addresses and opening hours?”
This is not a refusal - it never says it cannot. It hands the question back. For the customer in a hurry who wanted a bakery, the outcome is the same: they have no name.
And when ChatGPT does agree to name bakeries, the count collapses with size: 8 and 8 in Paris, 6 in Lyon, 2 in Nantes, 2 in Guéret, 1 in Saint-Amand-Montrond.
The trade matters more than the size
The control is what showed this, against our first reading. Across the 32 draws: the garage gets names 16 times out of 16. The bakery, 6 times out of 16. Same engine, same night, same communes.
In other words the dominant factor is not the population: it is the trade. A bakery in Grenoble - 156,000 people - gets the question handed back as often as a bakery in Vierzon.
What we did NOT find
We were looking for a fallback to national chains: unable to know the local baker, the assistant would cite the big brands instead. It is a seductive hypothesis, and on our measurements it is false. Share of cited names that are national chains, metros versus small communes: ChatGPT 17% then 0%, Gemini 21% then 32%, Perplexity 27% then 28%. No clean effect.
Likewise, May's claim - that smaller geographies are easier - only half appears. Agreement between the two searching engines rises from 7% in Paris to 27% in Guéret on garages, but stays flat on bakeries. One trade out of two: we report it as an indication, not a result.
What to do on Monday morning
If you are in a small commune, your channel is the engine that searches.
Gemini and Perplexity already name you. Do not judge your visibility on ChatGPT, which will ask your customer a question instead of citing you.
Your address is what makes you nameable.
Both engines that cited J&F Boulangerie gave the street. That is what separates a shop from a word.
Regional directories weigh more where you are than in Paris.
Perplexity cites the Petit Futé for Guéret. In a commune of 12,000, a handful of sources makes the whole answer - so being in them changes everything.
Do not conclude you are invisible because one test failed.
Our measurement shows the same shop cited by two engines out of three. The third is not saying you do not exist: it is asking which neighbourhood.
The limits, and a method error of our own
Eight communes, two trades, one night, one model per engine: this is a sounding. We did not verify that every named business exists - we read the answers, we did not audit the bakeries. And the comparison with our national studies is dated on another day, which forbids treating it as a before-and-after.
Finally, an error found by reading and not by the machine: our refusal detector missed “je n'ai pas d'accès” when the apostrophe is typographic rather than straight. It is the same family of gap as yesterday's missed Portuguese subjunctive. Fixed, and reported here because a figure published without its known defect is worth less than a smaller, honest one.