Gemini tells customers who you are for. ChatGPT almost never does

Short answer: when Gemini recommends a business, it says who that business is for in 48.1% of its passages: "best for emergency call-outs", "ideal if you are a student", "perfect for a first home". ChatGPT does it in 4.6% of its passages, ten times less often. Across our four countries, that is 2,849 passages out of 5,917 for Gemini and 470 out of 10,105 for ChatGPT. The label goes to some trades much more than others: half of the hotels recommended in Brazil get one, three plumbers in a hundred.

One engine sorts businesses by customer. The other lists them.

What was measured

Our four studies, re-read on October 2, 2026: the same buyer questions ("the best hotel in Rio", "a good accountant in Lyon") put to ChatGPT (gpt-4o-mini) and Gemini through their APIs, without web search, in the United States (collected September 13, 2026), the United Kingdom (July 31), France (July 30) and Brazil (August 6). 3,600 answers, cut into passages (a numbered item, a heading and what follows, a paragraph); we kept the 16,022 passages that name a business, with the same cleaning as our previous edition (platforms, directories and advice items removed). In each, we looked for a statement of who the business suits, in English, French and Portuguese: "best for", "ideal for", "perfect if", "if you are looking for", "idéal pour", "si vous cherchez", "ideal para quem", "se você procura". The detector was tested on 10 phrasings it must catch and 5 it must not ("the best plumbers in Lyon", "ideal location"), then read on samples from all four countries. We also tried to measure neighbourhoods; that detector picked up "District Attorney" and advice about neighbourhoods, so it is not published.

What the label looks like

Gemini often builds its whole answer around it: a heading "Best for emergency and 24/7 service", then a business; "Best for student and shared housing", then a letting agent; "If you want a trendy, edgy look, go to one salon; for a classic experience, go to another". The label is a use case or a customer type, not a quality: emergencies, boilers, high-net-worth divorce, a first home, men's grooming. ChatGPT, in the same questions, mostly gives a numbered list with a line of description, and leaves the reader to work out which name fits them. The gap holds in every country: in the United States, 55.8% of Gemini's passages against 2.8% of ChatGPT's; in France 54.2% against 9.0%. Brazil is the lowest for Gemini, 31.6%.

Hotels get a customer, plumbers do not

The label follows the trade. 53.5% of the passages about Brazilian hotels say who the hotel suits, 52.3% for French accountants, 38.0% for French hotels; hairdressers and beauty salons come close to a third. At the bottom, Brazilian lawyers (3.1%), plumbers (3.3%) and electricians (3.6%), and British gas engineers (5.8%). These are the trades where the assistants name the fewest businesses at all and send people to platforms instead. Age followed a similar line in our previous edition (AI cites how long you have existed in 6% of its recommendations): the assistant describes businesses whose offer differs from one to the next, and says little about trades it treats as interchangeable.

What it means if you are the business

When Gemini names you, there is about an even chance it also tells the customer who you are for. That sentence decides whether the reader keeps you on their list: a family looking for a quiet hotel skips the one labelled "best for nightlife". The label is written from something the engine has read. So say it yourself, in plain words and in the same terms everywhere: who you serve best, for which job, and who you are not for. "Boiler repairs and emergency call-outs in North Leeds" gives an engine a label; "quality service at fair prices" gives it nothing to sort with, as we saw with prices (AI praises a clear price more than a low one). None of this guarantees that you are named; it gives the answer a reason to send you the right customer when you are.

What this does not prove

• These are the models' answers through their APIs, without web search: the way our audit asks. Consumer apps with search switched on may write differently.
• Four studies collected between July 30 and September 13, 2026, one pass per question. The models have moved since, and nothing here measures that drift.
• A passage can name several businesses; a label is then counted once for the passage. A heading such as "Best for emergencies" counts for the passage it opens.
• We do not know what the engine reads to write the label. That it describes a use case rather than a quality is what we observed; that a business's own wording feeds it is a reasonable reading, not a measurement.
• Trade figures rest on 100 to 300 passages each: clear at both ends of the table, not precise to the point.