It says it cannot recommend, then it names eight businesses

Short answer: in more than half of its answers, ChatGPT begins by saying it cannot recommend anyone, and then recommends eight businesses. Across the 872 answers of our United States study, 249 of ChatGPT’s 436 answers open with a first-person refusal: no real-time data, no specific names. All 249 name businesses anyway - 1,824 names in total, a median of eight per answer, exactly the median of the answers that carry no warning. Gemini wrote a sentence like that once in 436.

A warning that changes nothing about the list is not caution. It is a sentence.

What was measured

Our United States study of 13 September 2026: 13 cities, 18 trades, two assistants (ChatGPT and Gemini), two phrasings of the same question, 872 answers in all, 436 per assistant. We counted the answers containing a first-person refusal - a sentence where the assistant says of ITSELF that it cannot give real-time, current, specific, individual or personal recommendations. The pattern is deliberately narrow and reads the subject of the sentence: advice such as “never hire someone who cannot provide a current license number” is about the contractor, not about the assistant, and is not counted. One Gemini answer was removed from the count for exactly that reason, after reading it. What the pattern misses, it misses in one direction: 249 is a floor.

The warning follows the trade, not the assistant

Inside ChatGPT the warning is not spread evenly. It is close to automatic where a wrong choice costs money or safety, and absent where it costs an evening.

This is not the ranking of licensed occupations. Hairdressing requires a state license across the United States and gets the warning in 4 answers out of 26; wedding photography requires one nowhere and gets it in 13. What the order tracks looks closer to the cost of being wrong: a bad attorney, a bad furnace installation or a bad broker is expensive and slow to undo. A bad croissant is neither.

And it changes nothing about the list

The warning comes first: at the 6th character of the answer in the median case, and within the first 200 characters in 248 of 249 answers. Then the list follows, unchanged. We also tested whether it made the list safer, that is whether it pushed the assistant towards the best-known names: it does not. The share of names that travel to five of the thirteen cities or more is 16.2% in the answers that carry the warning against 17.0% in those that do not. Same median of eight names, same profile of names. The assistant warns you, then does exactly what it would have done without warning you.

What it means if you are the business

Being named is not worth the same everywhere. In a restaurant answer the name IS the recommendation: the customer reads it and books. In an answer about attorneys, brokers, dentists or heating - the trades where a single customer is worth the most - the name arrives with an instruction to go and check. That check happens away from the assistant, on whatever the customer can find: a license register, an insurance certificate, a professional body’s page, reviews, an address that exists. If none of that is findable under your name, you were named and you still lose, at a step you never see. The practical consequence is narrow enough to act on: in those trades, write down the evidence the second check is looking for - license number and issuing authority, insurance, service area, guarantee - in text, on pages you own, not inside an image and not only on a profile that belongs to someone else.

What this measurement does not prove

• It counts a PHRASING, not an intention. The pattern reads a first-person refusal; a warning written some other way escapes it, so 249 is a floor and so is 57%.
• One date and one pass: 13 September 2026, one model per assistant. A house style is a product decision and can change with a version.
• No customer was observed. That a warning changes what someone does next is plausible, and is not measured here.
• It does not change who is named: the same median of eight names, and 16.2% against 17.0% of names travelling to five cities or more. The hypothesis of a safer list was tested and set aside.
• Restaurants and bakeries carry only 10 answers each, not 26: their zero rests on less material than the trades at the top.
• United States only, questions asked in English. Nothing here says what an assistant does in French or in Portuguese.