Should our reputation priorities for AI assistants differ from those for human customers?

Where do AI assistants and customers agree?

On the basics: both put rating and price first, and both value rating well above review count.

Baig and colleagues (opens in a new tab) asked twelve AI models, including GPT-4o-mini, Gemini and Claude models accessed through developer access, to pick one of five invented hotels. Each hotel’s rating, review count, price, certification and other details were randomized. The researchers then measured how much of each model’s choice each detail explained.

Rating took 35.7% of the weight and price 33.8%. Rating outweighed review volume by nearly four to one. That matches decades of research on human guests. One hotel-industry review of studies found rating’s link to hotel performance was 0.888 against 0.055 for review count, more than ten times larger.

So the investments you already make for guests, a strong rating and a defensible price, are the same ones the AI rewards most.

Where does the AI depart from what customers value?

In two places: it over-rewards eco-certification and ignores whether management replies to reviews.

The eco-certification result is the surprise. Human studies find green labels do affect bookings, but modestly, and mostly among environmentally minded travelers. The AI panel placed it third, rivaling review volume.

Chain membership also behaved unexpectedly. Brand is often treated as a way to reassure an unsure guest, yet the models gave chains a small penalty rather than a bonus.

Should you stop replying to reviews?

No: replies still matter to people and to later reviews, but do not expect them to move AI picks directly.

In the audit, a hotel card that said “Management responds to guest reviews” gained nothing. Human research cited by the authors finds that hotels which start replying see their online reputation improve, partly because later reviews change. That indirect route still works through the reviews themselves.

It matters because reviews are where AI assistants build a brand’s reputation. In our study of 79 brands asked “Is this brand legit?” on four AI engines in September 2026, 88.0% of answers cited a review or complaint platform. Claims tied only to review platforms were negative 56.5% of the time. Trustpilot and the BBB alone made up 61.7% of review-platform citations.

So the reply line on a listing is not the lever. What customers end up writing on the main review sites is.

Why does review volume count for more with AI than you might expect?

Because assistants treat a large review count as proof that the rating is real.

In the hotel audit, a high rating moved choices more when it was backed by many reviews. The authors read this as the models using volume to judge how far to trust the rating. They warn that this favors established properties over newcomers with thin review histories.

Our own data on live ChatGPT answers points the same way, for local services rather than hotels (our local picks study). The median business ChatGPT listed had 235 reviews, against 133.5 for those it left out, while both had a median rating of 4.9. Among businesses listed at least once, 69.2% of those with more reviews than the local median were listed in all three repeat runs, against 44.1% of those with fewer. Those are associations, not proof of cause.

For a new location or product, the review count is the gap to close first.

Can you trust what an assistant says about why it chose?

Not fully: the models’ stated reasons left out factors they acted on and stressed ones they ignored.

The hotel audit asked each model to explain its pick. The explanations broadly tracked the real weights, with telling gaps. List position was almost never named, despite a 4.1% share of the weight. Review volume was rarely named, despite a 9.4% share. Brand was named in up to 55% of one model’s reasons, while it carried 2.0% of the weight.

The practical lesson: do not set reputation priorities by asking an assistant what it values. Measure what it does instead.

How do AI assistants describe a brand’s reputation?

As legitimate but flawed: answers nearly always pair a yes with a list of problems.

In our reputation study, every complete answer said the brand was legitimate, and 99.7% also made at least one negative claim. Negative claims reached the first paragraph in 10.9% of answers, but in 31.6% of Perplexity’s. Claims tied only to editorial review sites were negative 19.9% of the time, and those tied only to the brand’s own website 6.4%.

A human shopper can choose to read your site or your reviews. The AI assembles both into one verdict, and the complaints come from third-party review platforms. Google’s summaries may weigh that criticism differently; see how AI Overviews treat negative sources.

What should you do about it?

Keep serving customers first, then add the few signals the AI weighs more heavily.

1. Hold rating and price discipline. Both audiences put them first.
2. Grow review volume steadily. Ask every satisfied customer, and start early for new locations and products.
3. Earn and display real certifications. In hotels, eco-certification counted far more with AI than with people. Claim only what you hold.
4. Keep replying to reviews, for people. Count it as customer service, not AI work.
5. Watch the complaint platforms. Resolve problems where Trustpilot, the BBB and similar sites record them.
6. Measure behavior, not explanations. Track what assistants recommend and say over repeated questions.

For help measuring what assistants say about your brand, see our generative engine optimization service.

What does the research not tell us yet?

The causal evidence comes from one industry and a stylized test, so treat the specific weights as a starting point.

• The only controlled audit is of hotels, with synthetic cards. Whether eco-certification carries similar weight in other industries is untested.
• “Management responds” was a single line on a card. How assistants treat actual reply text on review sites has not been measured.
• The human comparison uses published studies, not the same hotel cards, so the authors compare order of importance, not exact sizes.
• Our reputation and local studies are observational, collected on one or two days, and coded by AI models rather than people.
• Model versions change often, and the authors warn the weights may drift.

Frequently asked questions

Do AI assistants care about responses to reviews?

Not directly, in the evidence so far. A line saying management replies to reviews carried 0.1% of the weight in an audit of twelve AI models, although replies may still improve the reviews that assistants read.

Does eco-certification help with AI recommendations?

In hotels, yes. Eco-certification was the third most important signal in a controlled audit, with 13.1% of the weight, far more than human studies would suggest.

Are AI assistants’ explanations of their recommendations accurate?

Only partly. In a hotel audit, models rarely mentioned list position or review volume even though both affected their picks, and one model cited brand in up to 55% of its reasons despite brand carrying 2.0% of the weight.

Which review sites do AI assistants cite about a brand?

Mostly a few big ones. In our study of “Is this brand legit?” answers, Trustpilot and the BBB made up 61.7% of review-platform citations.

Sources

• Baig, Gillani and Ali (2026), Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection (opens in a new tab), arXiv:2606.16344.
• Underneath (2026), “Is this brand legit?” How AI assistants build a reputation
• Underneath (2026), Which Google Maps businesses does ChatGPT recommend?