- ChatGPT, Gemini and Claude have no privileged access to review platform data. They read the same public pages as anyone else, and they read only part of them.
- The figure needed to judge whether a score is representative reaches none of the three, and each one handles that gap differently.
- The limit is not a matter of model quality. No system can know which customers a business chose not to ask for a review, because that information is published nowhere.
The test
Increasingly, people choosing a supplier ask an AI assistant whether it is worth it, and receive a few lines summarising scores and reviews. We wanted to understand what data that summary rests on. Our starting assumption was that some systems enjoy privileged access to review platforms. The test disproved it.
On 30 September 2026 we put the same questions to the free versions of ChatGPT, Gemini and Claude, in sessions without memory, about two company profiles published on a review platform. This is an exploratory check, with a single run per system, and generative systems do not always answer the same way. The main finding, however, does not depend on any single answer.
What reaches an AI system
The score and the review count arrive first, because they sit in the page metadata. Then a handful of reviews, the most recent ones. On a profile of roughly 300 reviews, one system read about twenty, 7 per cent, without encountering a single negative one. The negative reviews were further down the list, and only the system that paged through ten pages found them.
What does not arrive is the figure that matters most. Some platforms publish the split between invited and organic reviews, with the monthly trend, on pages a user can consult in a browser but which they exclude from indexing. Other data is not public at all and stays visible only to the profile owner. An AI system, in practice, sees less than a person searching on their own would see.
Three different answers to the same gap
Asked how many reviews were invited and how many organic, none of the three systems had the figure. Gemini answered with unsourced estimates, laid out in a table as though they were measurements. Claude stated that the platform does not publish that data, when in fact it is published on a page the system cannot reach. ChatGPT explained where the information sits and why it could not read it, without inventing it.
To a reader, the three answers look equally reliable. All are orderly and confident in tone, and only one says clearly what it does not know.
A limit no model can overcome
The score also measures a company's invitation policy, and public data shows it. In our test, the profile of a large operator fed by invitations stands at 4.3 across roughly 3,600 reviews. The same group's profile that is not fed by invitations stands at 1.7, and a competitor that does not invite customers to review sits at 2.3.
That comparison shows that the way reviews are collected weighs on the final number. It does not show whether the collection is representative. Establishing that would require knowing which customers were invited and which were not, and at what point in the relationship, before or after a complaint. It is written in no public source. A more powerful model will read more pages, but it cannot read information that does not exist.
What an AI system would need
A public register of brands whose collection process has been verified as representative by an independent third party would give AI systems precisely the information that is missing today. To be a source rather than a shop window it must sit on a domain other than the brands' own, it must be readable by automated systems, and it must also carry suspensions and withdrawals, with date and reason. A seal on a company's website speaks to people, an independent register speaks to machines as well.
We have also written about this in relation to how AI amplifies reputational risk.
Frequently asked questions
Do AI systems have privileged access to review platform data?
No, at least not in the free versions tested. They read the public pages available to anyone, and often only a small part of them.
Can an AI assistant say whether a company's score is trustworthy?
It can report the score and the number of reviews. It cannot establish whether they represent customers' overall experience, because it does not know which customers were invited to review.
Do the results of this test always hold?
System answers can change over time and from one request to the next. The structural limit does not, because it concerns information that is published in no source.
© Fametrue 2026. All rights reserved.
Back to the Observatory