Note: this is a dated study of one specific sector. We published it in July 2026 and have left it as it was, with its data and its date, because the method works for any local business — hair salon, garage, gestoría (accountancy and admin firm) — even though the sector we measured was dental. The results are from that day and have not been run again.

On 17 July 2026 we ran a simple experiment to see which are the best dental clinics Barcelona AI assistants recommend. We took the four questions a real patient asks when looking for a dentist in Barcelona and put them to ChatGPT and Perplexity, with no tricks and without asking them to name anyone:

What are the best dental clinics in Barcelona?
I need a dentist in Gràcia, which one do you recommend?
Best dental implant clinic in Barcelona.
Emergency dentist Barcelona Sunday.

(We asked in Spanish; these are our translations.)

Eight answers. Between them, the two AIs named more than thirty different dental clinics in Barcelona.

Do you know how many appeared in the answers of both? Two. And both of them in the same question.

If you are looking for the best dental clinics in Barcelona according to AI, the awkward conclusion is that there is no single answer. There is one per engine, and they look nothing alike.

How we did it (and what it does not prove)

We start here, because this sector sees far too many "studies" that nobody can check.

We called the OpenAI API directly (with web search switched on and an approximate location in Barcelona) and the Perplexity API (the sonar model, location Spain/Barcelona). No system prompt: just the question, as it was. No session memory. We saved all eight answers word for word, and the 45 sources they cited.

And now the caveats, which matter as much as the results:

It is a single run, on a single day. These answers vary between runs and change over time.
It is not a ranking of clinical quality. An AI naming a clinic does not mean it is better. Not naming it does not mean it is worse. That is exactly what we are trying to explain.

That said, the patterns were too clear to be chance.

Finding 1: zero overlap in three of the four questions

For "best dental clinics in Barcelona", ChatGPT built a table with Nart, Den, Aparicio, Oris, Pina and Puyuelo, among others. Perplexity answered with Muro, Molins, AE Clínica Dental, Pifarré, Aribau, QMax…

Not a single clinic in common. None.

For "dentist in Gràcia":

ChatGPT recommended starting with Poldent, and mentioned Falcó, Odontquality, Sanitas Milenium and Anmella.
Perplexity suggested Friedländer, Diego Caballero and Miravé. Zero overlap again.

For implants:

ChatGPT gave a shortlist of four (Nart, Birbe, Barcelona Dental Institute, Clínica Dental Barcelona).
Perplexity found two, and neither was one of those.

Think about it from the patient's chair. Two people with the same toothache, in the same barrio (neighbourhood), on the same day. One uses ChatGPT and the other Perplexity. They get two lists of dentists without a single name in common. It is not that one is right and the other wrong: they are parallel universes.

That is why asking ChatGPT once from your own phone tells you nothing. You have answered one question of an exam with four subjects. It is also why more and more customers ask ChatGPT instead of Google: the board has multiplied.

Finding 2: the only question where they agreed

The only two clinics named by both engines — Clínica Dental Barcelona and Clínica Dental Raslan — both appeared in the same question: Sunday emergencies.

That is no coincidence, and it is the most revealing result of the experiment.

"Emergency Sunday" is a question with an objective, checkable answer: either you open on Sunday or you do not. There are opening hours, a phone number, an address. When there is a verifiable fact, the two engines converge on it, because both found the same explicit, unambiguous information.

"Best dental clinic in Barcelona" has no objective answer. That piece of data does not exist anywhere. So each engine makes it up from what it has read — and since each has read different things, each makes up something different.

There is the whole mechanism, in one sentence: the vaguer the question, the more the source decides. And which sources each engine trusts is not the same, as we saw next.

Finding 3: they do not read the same sites (not even close)

Between them, the two AIs consulted 45 different domains. Only 6 appeared in the citations of both: Doctoralia, Top Doctors, Sanitas, Adeslas and the websites of two clinics. A 13% overlap.

And the profile of what each one reads is very different:

ChatGPT cited 17 domains and leaned on medical directories and institutional sources: Top Doctors, Doctoralia, the COEC (the Catalan College of Dentists), the Consejo General de Dentistas (Spain's General Council of Dentists), CatSalut (the Catalan public health service). It even opened its answer with a reminder that it is worth checking the dentist's professional registration.
Perplexity cited 34 domains, many of them listicles and aggregators of the "best dental clinics in Barcelona" kind: dentalq.es, topdentistas.com.es, ascensium.es, dentistasplus.com, doctorplus.es, aesinergy.es. An Instagram profile, too.

And here is the detail that made us raise an eyebrow most. Perplexity presented Clínica Dental Muro as "considered by many the best clinic in Barcelona" (our translation), citing as its source a page on muroclinicadental.com titled "best dental clinic in Barcelona". It named Clínica Dental Molins citing an article hosted on clinicadentalmolins.es called "28 best dental clinics in Barcelona". And it named AE Clínica Dental citing a post on the aeclinicadental.com blog about the best dental clinics in Barcelona.

Let us be clear: those clinics have done nothing wrong. Publishing content about your sector is legitimate, and it is working for them. What the data shows is how the machine works: when there is no official truth to consult, the engine trusts what it finds written down, and sometimes what it finds written down was written by the very business being recommended.

Finding 4: the AI believes what it reads

Two examples from the same batch, without leaving our four questions.

Perplexity described a Barcelona clinic as "the only orthodontics clinic in Barcelona" (our translation). That is obviously false: there are dozens. It is marketing copy from a website that the model repeated without filtering.

And when answering about Gràcia, it slipped in among its sources a Doctoralia page for… Gràcia in Badalona. A different town.

We are not telling you this to laugh at Perplexity. We tell it because it explains why this cannot be fixed just by "writing well": the model does not judge, it cross-checks. And what it cross-checks depends entirely on where you appear, how you appear and how much ambiguity there is. If your information is scattered all over the web in a mess, the machine gets muddled along with you.

What this means if you run a clinic in Barcelona

Three concrete things.

Winning in one engine gets you nothing in the other. Nart came up twice in ChatGPT and zero times in Perplexity. Muro was "the best in Barcelona" for Perplexity and non-existent for ChatGPT. They are separate competitions, and we go into them in ChatGPT, Perplexity and Gemini do not recommend you in the same way.

Your real reputation does not travel on its own. There are clinics here with decades behind them that did not appear in any of the eight answers. Not because they are bad: because they were not legible to the engine we asked. That is exactly what we explain in why your business is invisible to AI even if you show up on Google.

And this changes by itself. Our test is a snapshot of 17 July. Repeated in September it will give something else, because the models get updated and the sources move. It is not an exam you pass once: that is why what we do is regular measurement, not a report and goodbye.

What we are not going to do

We are not going to finish this with a list of five tricks for getting into ChatGPT. For two reasons.

The first is that it would be a lie useful only to us: looking at this data, any generic recipe — "get into this directory", "use these words" — works for one engine and not the other, and goes stale within a quarter. What matters is not the list of actions. It is which of them apply to your clinic, in your barrio, in what order, and against which engine. That cannot be written in a blog post: it has to be measured case by case.

The second is that we do not even know whether you have a problem. Your clinic may show up perfectly in three engines, and this article is not about you. Recommending work before looking at your data would be exactly what we are criticising.

FAQ

Can I run this test myself?

You can, and you should, even if only once. But be careful with the conclusion: asking from your phone, logged in and with your history, gives you one answer from one engine at one moment. Our experiment shows that with that snapshot you can believe you are visible while being invisible across 75% of the map.

My clinic did not come up in your test. Is that serious?

Do not draw conclusions from this. It is four questions out of the hundreds a patient asks, on one particular day. What it does indicate is that you are not among the names the AI gives by default for the most generic and most valuable queries, and that does deserve a proper measurement.

I was on the list. Is that it, then?

Congratulations, and watch out: you came up in one engine, for one question, on a Friday in July. Nobody's position in this test is guaranteed. Staying there takes more than appearing once.

Why do we publish the sources?

Because a study that cannot be checked is not a study, it is advertising. We saved the eight answers word for word and the 45 domains cited. If you are a customer and want to see them, we will send them to you.

Run the test yourself, then decide

You can repeat the method above this afternoon with your own business, and it costs nothing: four questions written the way a customer who does not know you would write them, in two assistants, without being logged in. Note down who comes up. If it is not you, you know where to start.

What that home-made test does not give you is continuity, and that is exactly what is needed: our snapshot is from 17 July and in September it would give something else, because the models update themselves. Asking once is an anecdote; asking the same thing every month is data.

That is what we do on Grow (€9/month) — every month, ChatGPT and Perplexity, always the same questions — and on Pro (€24/month), every week and across all four assistants, with up to eight competitors tracked. Before any of that, creating and publishing your page is free, with no card needed: start here.

What we will not promise you, after showing you this data, is that an assistant will name you.

The best dental clinics Barcelona AI recommends: we asked ChatGPT, and this is what it said