What makes ChatGPT or Gemini recommend one hotel over another?

Which hotel details move an AI assistant’s pick the most?

Guest rating and price: together they outweigh every other detail an assistant was shown.

Baig and colleagues (opens in a new tab) ran a pre-registered audit of twelve AI models. Each was asked to recommend one of five invented hotels. Every hotel card listed a rating, review count, date of the latest review, whether management replies to reviews, chain or independent, nightly price, and whether it held a Green Key eco-certification. Each detail was assigned at random, so its effect can be read directly.

The panel included OpenAI’s GPT-4o-mini, three Google Gemini models, four Anthropic Claude models and four smaller open models. The study made 61,459 model calls in total. Here is how each detail changed the chance that a hotel was the one recommended:

All twelve models preferred higher-rated, cheaper hotels. They disagreed sharply on eco-certification: its effect ranged from almost nothing to 29.9 points depending on the model. The models’ own stated reasons are a separate matter, covered in whether AI explanations can be trusted.

What is each detail worth in dollars per night?

Because price was randomized too, each detail can be priced: a full rating step was worth about $126 a night.

The researchers converted every effect into the nightly price change that would cancel it out. Moving from 3.9 to 4.7 stars was worth $126.4 per night. Eco-certification was worth $46.4, a large review count $33.2 and fresh reviews $6.2. Chain membership cost $7.2. A visible management response was worth $0.5 per night, which is indistinguishable from nothing.

These are the assistant’s trade-offs, not a guest’s. They tell a revenue manager where the AI channel puts its weight.

Does the order of the list matter?

Yes: on average a hotel listed first was picked more often, and one model leaned on order heavily.

Across the panel, a hotel in fifth place was recommended 3.7 points less often than the same hotel in first place. Being listed first was worth $11.7 per night on average, and as much as $18.8 for the business traveler. That gain comes from placement alone, with nothing about the property changed.

The average hides one outlier. Gemini 2.0 Flash showed a first-position advantage of about 26 points, roughly ten times the panel average. Order is usually set by the booking site, search tool or interface feeding the assistant, which a hotel does not control. Our guide on list order in AI recommendations covers tests beyond hotels.

Does the traveler’s request change the weights?

Yes: the same details counted differently for a budget family, a business traveler and an eco-minded couple.

The audit used three traveler personas. Eco-certification was worth $65.4 per night to the eco-conscious couple against $36.8 for the budget family. A top rating was worth $98 per night for the family against $153 for the business traveler, because the family was more price-sensitive. The chain penalty was steepest for the eco-couple and vanished for the business traveler.

So a hotel that courts one kind of guest should make the details that guest cares about plain on every listing.

Where does the assistant get its hotel information in the first place?

Mostly from booking sites, at least in one Gemini audit of Tokyo hotels.

The audit above handed each model a fixed list. Real assistants search first. Zhu and Chang (opens in a new tab), of Blossom AI, a company, put 156 Tokyo hotel questions to Gemini 2.5 Flash with Google Search in March 2026. Online travel agencies such as Booking.com and Expedia supplied 55.3% of the 1,357 cited sources. Hotels’ own websites supplied 8.2% of citations in English and 11.0% in Japanese.

In a small check of 14 hotel websites, the hotels Gemini cited directly had deeper content, such as long FAQ pages and neighborhood guides. That check shows association, not cause.

Do real assistants weigh reviews the same way?

The closest real-world evidence says yes for review volume, though it covers local services, not hotels.

In the audit, the API-served models answered almost identically when the same set was repeated. Real consumer apps add web search, and their answers vary more. We checked live ChatGPT answers to 120 local questions, such as the best dentist in a city, over two days in September 2026 (our local picks study).

ChatGPT listed 67.7% of businesses ranked 1 to 3 in Google Maps. At the same Maps rank, more reviews than the local median went with a 19.5-point higher chance of being listed, and a rating of 4.8 or more added 9.9 points. Our study shows association, not cause, but it points the same way as the controlled audit: rating and review volume travel with being recommended.

What should you do about it?

Treat the AI channel as a buyer who reads your rating, price and certifications, but not your replies.

1. Protect the rating and grow review volume. These are the largest levers in the audit and in our live data.
2. Price with the AI trade-offs in mind. A higher rate costs recommendations unless the rating and other signals justify it.
3. State real certifications on every listing. Eco-certification counted heavily with several models, especially for eco-minded travelers. Only claim what you hold.
4. Keep facts identical across booking sites and your own site. Assistants read listings you do not own.
5. Build pages that answer guest questions. Deep FAQ and area guides went with direct citation in the Tokyo audit.
6. Check the answers regularly. Ask the main assistants the questions your guests ask, several times, and note who is named and why.

For help setting up that monitoring, see our generative engine optimization service.

What does the research not tell us yet?

The research shows how assistants choose among hotels they are given, not how a hotel gets into the list.

• The hotel cards were synthetic, with no photos, review text or amenities, and every hotel was a 4-star near the city center.
• The audit tested models through developer access in English with US-dollar prices, not the consumer ChatGPT or Gemini apps.
• It covered one turn of conversation; follow-up questions could change the weights.
• The Tokyo citation audit covered one engine, one city and one month, and measured citations, not bookings.
• Model versions change often, and the authors warn the weights may drift.

Frequently asked questions

Does replying to reviews help a hotel get recommended by AI?

Not directly in the evidence so far. A line saying management replies to reviews had no detectable effect across twelve models, worth about $0.5 per night.

Do AI assistants prefer chain hotels?

No. In the audit, chain membership slightly lowered the chance of being picked, by 1.8 points, and the penalty was largest for eco-minded travelers.

Does eco-certification matter for AI hotel recommendations?

Yes, more than many managers expect. Green Key certification raised the chance of being picked by 11.6 points across the panel, though some models gave it almost no weight.

Can a hotel get cited by ChatGPT or Gemini instead of Booking.com?

Sometimes. In Gemini’s Tokyo answers, hotels’ own sites were 8.2% of English citations, and the cited hotels tended to have deep, question-answering pages.

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.
• Zhu and Chang (2026), The End of Rented Discovery: How AI Search Redistributes Power Between Hotels and Intermediaries (opens in a new tab), arXiv:2603.20062.
• Underneath (2026), Which Google Maps businesses does ChatGPT recommend?