The same dentist, three different answers

On 1 September 2026 Mador put the same 5 questions about a dentist in Turin to three assistants, and looked at where one dental practice in the city came up. Grok named it 5 times out of 5. Perplexity 3 times out of 5. ChatGPT never did.

They are the questions of someone with a sore tooth and no name in mind, asked in Italian: can you recommend a good, honest dentist in Turin, is there a recommended dentist in the centre. All 15 answers came back within a minute, so the day and the hour have nothing to do with it. The only thing that changes is who is answering.

The practice's own site: read 4 times, once, never

Every answer carries the list of sites the assistant opened to write it. The practice's site is on 5 of those lists: 4 of Perplexity's, 1 of Grok's, none of ChatGPT's.

It would be natural to think that whoever reads more names more. That is not how it went. Perplexity read the site 4 times and gave the name 3 times: on the question from someone who is scared of the dentist it opened the site, then recommended other people. Grok read it just 1 time and named the practice in all 5 answers. Where the name came from the other 4 times, this measurement cannot say. It says the site was not among the things read, and that MioDottore, the portal where people book appointments, was there all 5 times.

ChatGPT listed the sites it read in 3 answers out of 5, 20 sites in all. The practice's site is not one of the 20. Neither is MioDottore, which the other two opened in every one of their 10 answers. And of those 20 sites, 16 were read by ChatGPT alone.

Three assistants, 63 sites, 1 in common

Between them, the three read 63 different sites. 54 were opened by one assistant only. 8 were opened by two. That leaves 1 that all three read, and it is not the practice's.

The same question, asked in the same minute, sends three assistants off to read three different parts of the web. The Turin practice is in two of those parts and missing from the third.

Which of the three numbers is the right one

All three. Someone who had only tried Grok would say the practice always comes up. Someone who had only tried ChatGPT would say it does not exist. Both would be right, and neither could decide anything on the strength of that. The number to work with is the sum: 8 answers out of 15, a little over one in two, and inside it there is one assistant in three that never says the name.

That is why opening ChatGPT on a phone and asking is not a measurement. It is one answer in three, and which one depends on the assistant that happens to be on the phone.

What this does not say

This is 5 questions, 3 assistants and a single day. ChatGPT's 0 out of 5 is the figure for these 5 questions, and other questions may change it. There are more than three assistants in use, and about the others nothing is written here. And a site that gets read is not a site that convinced anyone: Perplexity opened it 4 times and gave the name 3.