Who does ChatGPT recommend for a dentist in Toronto?

We put a single, specific consumer question to gpt-5.4-mini: “I need a good dentist in Toronto for a dental implant. Which practices would you recommend and why?” The prompt was sent three separate times on 2026-09-23, each run started fresh, with the model's default settings and no web browsing enabled. Locale was set to en-CA.

Running the same prompt more than once is the whole point. A single answer tells you what the model said that one time. Three answers, sent minutes apart with nothing else changed, tell you whether the model is consistent, and how much of its answer is stable versus how much shifts run to run.

In none of the three runs did gpt-5.4-mini name a specific Toronto dental practice. Every response opened with some version of the same caveat, that it should not “rank” or recommend a clinic without current, verifiable information, and then pivoted to describing the types of practice worth considering.

The categories repeated across runs, worded differently each time: a prosthodontist or periodontist-led implant clinic for complex cases, a general dentist who places a high volume of implants for straightforward single-tooth cases, and a university-affiliated or teaching clinic, mentioned in all three runs as a lower-cost, slower option. One run used the phrase “Toronto Implant Institute / implant-focused practices” as a heading, but the text under it described a type of practice, not a named business, so we are not treating it as a real recommendation.

The shape of the answer was consistent: a caution about not ranking clinics blind, a short list of practice categories with a “best for” note on each, a set of vetting questions, and a closing offer to narrow things down given more detail. That structure showed up in all three runs almost like a template.

What drifted was the wording and the exact categories offered. Run one listed three categories; run two listed three with different headings and added a “red flags to avoid” section that run one did not have; run three folded the categories differently again and used the phrase “specialist-led implant centers.” None of the three runs used identical section headings, and the response length varied from 2,716 to 3,752 characters for what was, functionally, the same answer.

A dental practice owner reading this cannot conclude that ChatGPT never names dentists, or that Toronto implant dentistry specifically triggers a no-names policy from every AI assistant. We tested one model, on one date, with browsing switched off, using one exact phrasing of the question, three times. That is a narrow window.

A browsing-enabled session might pull in current directory listings or review sites and name real clinics. A different model, Gemini, Claude, or Perplexity, might answer with more or fewer names. A rephrased question, such as one that names a neighborhood or asks for “top-rated” clinics instead of a general recommendation, might behave differently too. None of that was tested here, so none of it can be claimed here.

Even with zero practices named, the model handed the user a specific checklist, unprompted, in every run: implant volume per year, whether the same provider places and restores the implant, whether CBCT 3D imaging is used, the stated success rate, whether grafting is needed, and what happens if the implant fails.

That checklist is not a set of ranking factors, we did not test whether answering it changes what any AI tool says. It is simply the set of questions this model told a real user to ask before booking. A practice whose own website already answers those questions in plain language is answering the exact questions this model is coaching patients to ask, regardless of whether that changes anything about AI visibility.