Lawyer, plumber, dentist: ChatGPT names nobody nine times out of ten

On 10 August, OpenAI updated its advertising policies. One line in the changelog, dated April, is worth pausing on: “medical, legal, and financial advice contexts are no longer categorically blocked from ads by default.” In other words, a conversation where someone is looking for a dentist can now carry an ad.

That raises a question that is anything but theoretical for a practice, a firm or a workshop: in that same conversation, what does the assistant answer when nobody is paying?

We have the data to answer, and it needed no new calls. Our three national studies - France on 5 August, United Kingdom on 31 July, Brazil on 6 August - total 4,092 real answers. Keeping only the trades that exist in all three countries gives 3,780 paired answers across 18 trades, 39 cities and 3 assistants. The trade varies; everything else is matched. That is exactly the experiment you want.

One engine refuses. The other two, never.

In every answer we looked for an explicit refusal: “I can't provide specific recommendations”, “je ne peux pas fournir”, “não posso recomendar”. Across 1,260 answers per engine:

This is not a difference of style. It is one engine telling the customer, in nearly one answer out of two, that it will name nobody - while the other two name.

And the split does not follow regulation

We set out to find a line between regulated professions and free trades. There isn't one. Here is ChatGPT's refusal rate, trade by trade, on 78 answers each:

• Solicitor 92%, accountant 92%, electrician 92%, plumber 91%, dentist 88%.
• Insurance broker 63%, locksmith 45%.
• Wedding photographer 31%, garage 28%, estate agent 23%, driving school 19%, marketing agency 14%, hairdresser 10%, beauty salon 8%, recruitment agency 3%.
• Bakery 0%, restaurant 0%, hotel 0%.

The plumber sits at 91%, just under the solicitor. The driving school - a licensed activity if ever there was one - sits at 19%. The dividing line is not the qualification: it is what a bad choice costs you. A flooded kitchen, a dangerous circuit, a lost case, a botched tax return, a tooth. Against: a mediocre croissant.

The more the choice matters, the less the assistant makes it for you. It names bakeries without hesitating and refuses to name a plumber.

The gap also runs between countries, and it is wide: ChatGPT refuses in 31% of cases in France, 46% in Brazil and 53% in the United Kingdom. Same product, same week, three thresholds.

What it does instead

A refusal is not silence. Across ChatGPT's 543 refusals, 60% cite a directory or a professional body - Yellow Pages, the Law Society, Doctolib, Checkatrade - and 20% list selection criteria: check the specialism, ask for a written quote, read the reviews.

Translated for a firm: on the question that should have brought you a client, the assistant answers by sending that client to a directory. The directory wins the conversation. You are not even mentioned in it.

The advertising door is not open either

You might assume advertising offers the way out. OpenAI's text, updated 10 August 2026, says precisely the opposite for most readers of this article. It has to be read closely.

Put the two measurements side by side. A French dentist: refused in the free answer 85% of the time, and ineligible for the ad because they are not American. A Brazilian dentist: 96% refusals, same closed door. A solicitor, anywhere: refused nine times out of ten, and legal-services advertising is permitted nowhere.

Let us say this honestly, because it is what makes the finding useful rather than accusatory: declining to single out a professional whose practice requires a licence is a defensible position. OpenAI's usage policies have framed licensed personalised advice since October 2025. We impute no motive. We record a state of affairs: for these trades, and especially outside the United States, the channel is closed on both sides at once.

The measurement trap - and it is ours too

Here is the part that should worry anyone buying an AI-visibility tool, including ours.

We ran those 543 refusals through our own competitor extractor. It produced 3,099 “competitors”, or 5.7 per answer. Not one refusal returned zero.

The names it keeps? “Consultation initiale” 15 times, “Certificações e Licenças” 13 times, “Especializações” 12 times, “Devis” 12 times, “Qualifications et certifications” 11 times. These are the sub-headings of the advice the assistant gives instead of a name. A report built on that would tell a London solicitor they have five competitors, one of them called “Initial consultation”.

When an assistant answers with advice instead of names, a naive measurement counts the advice as competitors. Including ours - we published that hole yesterday.

And in insurance broking, the few genuine names that survive the sort say something else again: Aon 14 times, Towergate 12, Hiscox 10, Marsh 10. National and global brokers. The high-street broker never appears - even when the assistant does consent to name someone, it names the big ones.

What to do on Monday morning

The limits, and a mistake we fixed along the way

Refusal detection is a list of phrasings, therefore fallible. We checked it both ways on random samples: of eight answers flagged as refusals, eight were. Of eight unflagged, two were refusals we had missed - Brazilian Portuguese phrases the refusal in the subjunctive after “embora” (“embora eu não possa fornecer”), and our pattern only looked for the indicative. Fixed, then re-checked on a fresh draw: ten out of ten correct. The first figure we had, 62% for regulated professions, was therefore a floor; the right one is 75%.

The rest, as always: these answers date from late July and early August 2026, one model per engine, and we have shown elsewhere that answers move from one day to the next. 3,780 answers are not a census. But the gap between 92% and 0% does not close with noise.