- Divergent outcomes. Asked the same question, ChatGPT, Gemini, Claude and Perplexity picked three different cruise lines.
- Source dependency. The choice follows whichever platform each assistant consults.
- Blind to method. In none of the 12 initial answers did the agent consider how the reviews had been collected.
- Silent exclusion. A brand can lose customers without knowing, discarded by an agent before any comparison reaches the user.
Same question, three different answers
"I need to book a cruise and I am considering Costa Cruises, MSC Cruises and Carnival Cruise Line. You choose the one with the best reviews and tell me in one line why."
We put this to four AI assistants, three times each, opening a new chat every time. This is the agentic commerce scenario, where the AI does not merely advise but decides on the user's behalf.
In an earlier test we checked what AI systems read from a single brand's profile. This time we asked them to choose, and the answers were not the same.
| System | Choice, 3 runs | Prevailing source |
|---|---|---|
| ChatGPT | Costa, Costa, Costa | Trustpilot |
| Gemini | Carnival, Carnival, MSC | Cruise Critic and trade blogs |
| Claude | Carnival, Carnival, Carnival | Cruise Critic, via a 2024 article |
| Perplexity | MSC, MSC, MSC | Crociere.com, a booking portal |
Each system tends to repeat its own choice, but the systems do not agree with one another. Whoever reads Trustpilot picks Costa, whoever reads Cruise Critic, a portal with a largely North American audience, picks Carnival. Perplexity picks MSC on the basis of a 4.6 out of 5 average published by Crociere.com, a portal that sells cruises from several lines. Pressed on where that figure came from, Perplexity withdraws it as unsupported by the source and falls back on Trustpilot. Four sources of different kinds, three different winners.
Why Costa wins on Trustpilot
On Trustpilot the Italian profile of Costa Cruises records 4.3 across 3,621 reviews, that of MSC Cruises 2.3 across 889, and Carnival Cruise Line 1.5 across 527. The difference, however, lies not only in the numbers but in the collection method. Costa invites its customers to leave a review, so its score is shaped by that invitation. MSC does not invite, and Carnival has not even claimed its profile. Their reviews are unprompted.
Inviting customers to review is a lawful practice, but it makes scores incomparable, because they measure groups of customers assembled in different ways. Costa wins when the agent consults the platform where the company invites its customers, and loses when the agent looks where nobody does. The test does not establish which cruise line is better. It shows that the agent's answer depends on where it chooses to look.
Nobody looks at how reviews come about
When we asked them to choose, all four assistants decided on the score, without asking whether the reviews had been solicited by the company or written spontaneously by customers. Only two systems considered the collection method, and only when we asked them explicitly. Even then, on more than one occasion they read the figure wrongly.
Who answers for the agent's choice?
An agent now sits between the platform and the consumer, and its choice of source determines the outcome. The score is taken as it stands. The Italian Consumer Code, at article 22(5-bis), treats as material the information on whether and how a trader verifies that reviews come from people who actually bought or used the product, and omitting it may amount to a misleading practice. That safeguard, however, concerns the provenance of the individual review, not the representativeness of the whole, and in the test no agent looked for either.
The test does not claim that Costa's score, or any other line's, is distorted. It shows instead that if a score were distorted or unrepresentative, the distortion would travel intact all the way to the purchase. Two questions follow.
- If the agent buys on the basis of a score that does not represent customers' experience, where does responsibility lie? With the company that built the score, the platform that publishes it, or whoever configured the agent?
- If an agent favours one company over another on the basis of a figure nobody has verified, can competition between businesses suffer as well?
What changes for brands
With AI agents a brand can be ruled out before the customer ever sees it. Carnival, which has not claimed its Trustpilot profile, is discarded every time the agent reads that platform, and the customer will never know it was an option. Costa, which invites customers to review there, wins with the agent that reads that source and loses with the others. Reputation is no longer played out only on what people perceive, but also on the source the agent chooses to read.
The answer, though, is not to invite more in order to climb on a single platform. A higher score still does not say who was invited and who was not. What is missing today, and what none of the assistants tested managed to find, is verifiable information on how the reviews were collected.
Teaching AI systems to read how reviews were collected is not enough either. Even the system that spotted the invitations in our test concluded that invited reviews were more representative, when an invitation says nothing about the customers who were never invited. That information can only come from an independent verification of the collection process, published away from the brand's own site and in a form an agent can read too.
Frequently asked questions
Do AI assistants all pick the same company?
No. Asked the same question, ChatGPT picked Costa Cruises in all three runs, Claude picked Carnival Cruise Line, Gemini picked Carnival in two runs out of three and MSC Cruises in the third, Perplexity picked MSC. The difference follows the source each system chooses to consult.
Can a brand be penalised by AI assistants without knowing?
Yes. Each assistant picks its own source, and a brand with a weak or unmanaged profile on that source can be discarded before the customer sees the alternatives.
How can an AI agent tell whether a score is representative?
On its own it cannot, because no public page says which customers were invited to review and which were not. It takes an independent verification of the collection process, published on a source other than the brand's own site and readable by machines.
Method note. Test run on 8 October 2026 on the free versions of ChatGPT, Gemini, Claude and Perplexity, with three runs per system in fresh chats and identical questions. This is an exploratory check, not a statistical study, and system answers can change over time. ChatGPT's answers contained references to the profile of the person asking, although memory was switched off. Trustpilot scores verified on the date of the test.
© Fametrue 2026. All rights reserved.
Back to the Observatory