Does ChatGPT Recommend Your Clinic? Check in Twenty Minutes

Ask a clinic owner whether their site ranks on Google and they'll usually know, roughly. Ask whether ChatGPT recommends them when someone in their city describes a sore shoulder, and the honest answer is nearly always that nobody has checked.

We're running that check this month for ten physiotherapy clinics in Mississauga, cold, with no ask attached. This is the method, written down so you can run it on your own business in an afternoon. It works for any local service, and it's built around one distinction most people miss.

The one thing to get right

An assistant can fail to name you for two completely different reasons, and the fix for each is different.

The first is a retrieval problem. The model went and searched, read the top handful of results, and you weren't in what it read. That's a ranking and profile problem, and it can move in a quarter.

The second is a memory problem. The model answered from what it already knew, without searching, and it has never encountered you. Nothing on your own website fixes that, because the model didn't look at it. What fixes it is other people writing about you, which is slow and can't be bought.

So every prompt gets run twice on ChatGPT: once with web browsing on, once with it off. If you're present with browsing and absent without it, you've got a brand problem, and someone selling you schema markup is selling you the wrong thing. If you're absent both ways, start with the retrieval side, because it's faster.

Very few people bother with the second pass, and it's the one that tells you what kind of problem you've got.

Write the prompts the way people talk

Real people don't type "best physiotherapy clinic Mississauga" into an assistant. They describe a situation. So the prompt set has to include the situation.

We use ten prompts for a first look, and they deliberately don't all sound the same. A couple are the obvious ones, "who are the best physiotherapy clinics in Mississauga", because everyone tests those and you should too. A couple add a real constraint the way a real person would: physio near Square One that does direct billing and has evening appointments. Then the ones that matter most, where someone describes what happened to them rather than what they're looking for. "I hurt my shoulder lifting boxes on the weekend, still can't raise my arm, who should I see in Mississauga?" That's how most people ask, and it's where most businesses discover they don't exist.

Round it out with a comparison ("is Clinic A or Clinic B better for a running injury"), a question about you by name ("is [your clinic] any good", which is frequently the ugliest result and the most useful one), a vetting question ("how do I choose a physiotherapist", to see whether the model names anyone while it's giving advice), and a cost question, because a business that gets named inside a pricing answer is trusted in a way no ad can buy.

Swap in your own trade and city. The mix is the point, not the wording.

Run them on five surfaces, not one

ChatGPT with browsing. ChatGPT without browsing. Google's AI Overview. Perplexity. Microsoft Copilot. Five, because they draw on different indexes, and a business that Perplexity cites and Copilot has never heard of is telling you something specific: Copilot runs on Bing, and you're probably not on Bing Places.

Use a logged-out or temporary chat every time. Personalisation and chat memory will quietly hand you a flattering result, and a measurement you can't reproduce next month is worthless.

What to write down

One row per prompt per surface, so fifty rows for a first look. For each one:

Were you mentioned at all. If so, what position among the businesses named. Was your domain actually linked or footnoted. Was what it said about you correct. Every other business it named, in order. And the domains it cited.

That last column is the report. Without it you've got a feeling that you're invisible. With it you've got the three directory pages and the local top-ten list the answers were built from, a note that you're on none of them, and something to do about it on Monday.

What the results almost always say

The shape of the results is fairly predictable, and it's the opposite of what most owners expect going in.

For a local, commercial question, the assistant's answer is assembled mostly from the Google and Bing local packs, from third-party lists and directories, and only then from the business's own site, usually to confirm a fact rather than to discover the business. Which means the highest-leverage work for AI visibility is often not on your website at all. It's your Business Profile, your review count, and whether you appear on the pages the model reads.

This explains a thing clinic owners tell us constantly: they paid for SEO, the agency did real work, and the assistant still doesn't know them. The agency optimised the one source the model weights least.

It also explains why this is ongoing rather than a fix. The pages the model reads change. New listicles get written. A competitor gets forty reviews over a summer. You measure it monthly or you don't know.

What to do with the sheet

If you're absent with browsing on, look at the sources column. Those are the pages to be on. Then look at your Business Profile with a stranger's eyes: category, hours, services, photos, review count against the clinics that did get named. Then, and only then, the website: is the service you want to be found for stated in plain text on a page of its own, or buried in a menu?

If you're present with browsing and absent without it, the work is getting written about, by someone other than you. Local press, an association directory, a guest piece, a case study a supplier publishes. It's slow, and it's the only thing that moves that column.

If the model named you and got a fact wrong, fix the fact at its source, which is almost always your profile or a directory, then check again in a month.

Twenty minutes, give or take

Ten prompts on two ChatGPT settings is twenty observations and about twenty minutes. The full five-surface run is closer to an hour. Either one will tell you more about how your next patient is choosing a clinic than any ranking report you've been sent.

If you'd rather we ran it, say so. We do the first one free for clinics and firms in the GTA, with no pitch attached to the email, because the gap on the sheet makes the argument better than we could. The AI SEO work that follows, if it follows, starts from that sheet rather than from a template.

The reasoning behind the two failure modes, and why local profiles outweigh websites in these answers, is in How AI Decides Which Local Business to Recommend. Clinics specifically have one more constraint on what they can put on a site, covered in Your Clinic Website Can't Use Testimonials.