How to get your hotel on ChatGPT and Gemini: what the assistants read and what to publish

ChatGPT decides which hotels to recommend in two stages: it searches the web and draws up a shortlist, then checks the shortlisted hotels one at a time. Your hotel has to get through both. The searching is something we measured in the consumer app: on 27 September 2026 it went to the web before answering 186 of 188 travel questions.

To see what it looks for, on 26 and 27 September 2026 we picked 47 trips in Italy and Switzerland, from a spa hotel on Lake Como to a hotel near Florence station, and wrote the questions a traveller would ask about each. We put all 47 to Gemini and Perplexity, and some of them to ChatGPT and Claude. In half the conversations the traveller’s second message was: “which one would you choose, and what should I check before I book?” An assistant reached through its developer interface, as these were, hands back the list of web searches it ran along with its answer. We read 2,880 of them, covering 84 hotels.

Of the searches that build the shortlist, 95% are run on the first message. After that the work changes. Weighting each assistant by how much it is used, we estimate that three searches in four, 74%, check a hotel that is already on the list: is the spa included in the rate, how many minutes is it on foot from the station, is the pool heated in November.

Nobody types these searches into Google. Google Ads’ keyword tool could measure 1,326 of them, and it puts 1,318, or 99.4%, at zero searches a month, so the keyword tools that hotels and agencies rely on never show them. They change language, too: 73% of Perplexity’s searches were in English, even when the conversation was in Italian, German or French.

When ChatGPT and Gemini check a hotel they search for the practical terms of the stay, and each kind of trip has its own set, repeated from one hotel to the next. On 27 September 2026 we added one conversation for each of the 84 hotels, in Gemini and Perplexity, with the hotel named in the question: “Is Portrait Milano a good choice for an evening at La Scala?”, for example. Each hotel drew between 3 and 17 checking searches, and half the hotels drew at least 10.

More than half of those checks, 475 of 892, concerned something the traveller had not asked about: reviews, the pool, the shuttle, opening dates. In the spa conversations no traveller mentioned reviews, and the assistant went looking for them unprompted for 39 of 46 hotels. The table is drawn from all the conversations. Treat it as the list of questions your hotel’s page has to answer.

The booking portals, the sites the trade calls OTAs, are where ChatGPT and Gemini mostly check a hotel; the hotel’s own site comes second and the guides third. We counted this in the sources the assistants read to write 1,364 answers on 26 and 27 September 2026: 20,357 readings, from 3,148 different websites.

The third-party sites read most often are Tripadvisor, Booking.com, Expedia, Hotels.com, the Michelin Guide, Trip.com and Rome2rio, which answers the questions about distance. Google’s first page carries little weight here: of the sources read to build the shortlist, between 5% and 20% were on it, depending on the assistant. We compared one Google search per trip, so take that as an order of magnitude. In our August 2026 study, when an answer named one of the world’s 50 best hotels, the hotel’s own site was among the links shown 24% of the time.

Each assistant has its habits. These are the shares of the sources read on 27 September 2026 in the conversations that began with a hotel’s name.

The portals get things wrong, and the assistant passes the mistakes on. On 27 September 2026 we took 11 hotels and compared 169 checking questions with every public source. There were 98 errors on the web, and the assistants repeated 31 of them. For a five-star hotel in the centre of Rome, Google’s listing showed free breakfast, free parking and no pets allowed, all three of them wrong. The same hotel lists parking among its services and publishes no price for it. The assistants took €59 from the portals, one of which quotes it per night and another per day, and Gemini answered with €65 to €80 a day, a figure found in no source.

A famous hotel goes missing when the assistant knows its name and cannot tell which trip it belongs to. In August 2026 we put 1,096 travel questions to ChatGPT, Claude and Gemini, with web search switched on, across the 34 destinations of the 50 hotels on The World’s 50 Best Hotels list. The results are in our study of 3,288 answers from ChatGPT, Claude and Gemini.

Every one of the 50 was named at least once. When the question began with the trip, though, meaning when to go, how many days to allow or which area to stay in, one of the 50 appeared in 189 of 1,380 answers.

The further the question sits from the hotel, the more the big names drop out. For questions such as “When should I go to Tokyo?”, 96% of answers named none of the 50; for questions such as “Best luxury hotels in Paris?”, 13% named none.

The clearest case is a Sydney hotel in the top 20 of the list. None of the 9 answers to “where to stay in Sydney” named it; all 9 answers to the question about the best hotels did. Its own website gives its location by the historic name of the block. A third-party page puts the same fact in a form a traveller can use: metres from Circular Quay.

The hotels on ChatGPT’s shortlist are the ones that describe themselves in the words the traveller uses in the question. On 3 October 2026 we asked ChatGPT, on its most capable model and with no app installed, which Maldives resort to choose for a honeymoon: a private pool, access to the water from the villa, privacy.

It compared 3 resorts, of the 181 open that day according to the Maldives’ official register, and chose Milaidhoo, citing the resort’s own pages and those of Small Luxury Hotels. Milaidhoo’s site states a 42 m² pool, steps down into the lagoon and privacy, which are the words of the question. Of the 27 resorts we had studied by then, 21 met the same request, and ChatGPT considered none of them.

We asked the same question twice more that day, and a different resort won each time. The shortlist changes from one conversation to the next, and a hotel reaches it by setting out plainly the facts the question asks for.

A hotel website should publish the facts that decide a trip: as text, on a page of its own, in the traveller’s words and with a date. Adjectives help nobody choose, because the hotel next door can use every one of them.

On 3 October 2026 we read the websites of 30 resorts and the sources that write about them. Nobody published 288 of the answers a guest wants before booking: the rate for a guest travelling alone, the price of a babysitter, the distance from the beach to the reef. On 22 August 2026 we had mapped the website of a hotel in the top five of The World’s 50 Best Hotels list and found 75 pages on rooms, gardens and food, and none on what a night costs. When a traveller asked about prices, the assistant showed a blogger’s pages.

Old documents beat new ones. In the same reading we counted 72 cases where the sources contradict one another, often the official site contradicting itself. A 2018 press release says 20 minutes by speedboat, today’s FAQ says 30, and six portal pages still repeat the older number. Put a date on every fact that changes, and take the old PDFs down. Write notices into the text of the page: building work announced only in a banner is the first thing the machines miss.

Then correct the listings you control, on Google and on the portals, because that is where the assistant reads most often. The pages of our hotel guide are built this way: each says which trip the hotel suits and gives the facts to check it, with the source and the day we read it.

The crawlers to let in are the ones that serve search: OAI-SearchBot for ChatGPT and Googlebot for Google. A crawler is the program that reads pages on behalf of a search engine or an assistant, and the site’s robots.txt file tells it where it may go.

OpenAI recommends two things. The first is to allow OAI-SearchBot in robots.txt. The second is to let its crawlers’ requests through; they can be recognised by the IP addresses, the network addresses, that OpenAI publishes. That second point is a matter for the firewall, the filter that decides which requests reach the site, and for the CDN, the network that delivers it. Google recommends the same. While collecting official photographs for our guide we came across one chain’s CDN that refused automated downloads, so ask whoever runs your site to check.

No new files are needed. Google says its AI answers require no new machine-readable files, no AI text files and no special structured data. Structured data is the labelling in a page’s code that tells machines what the page contains: Google asks that it match what the page shows, and that the important content be available as text.

Getting your hotel recommended by Gemini and Google’s AI Mode takes the same work as Google Search. Google gives two kinds of AI answer: AI Overviews are the summary shown among the results, and AI Mode handles longer questions and comparisons. Google’s guidance on AI features says neither has additional requirements. The page has to be indexed, meaning already in Google’s store of pages, and eligible to be shown with a snippet.

The same guidance describes what we measured, several searches behind a single answer, and calls it “query fan-out”. Among its best practices it lists an up-to-date Business Profile, the Google listing that shows the address, hours, photos and services such as parking. Google adds that in local results on Maps and Search there is no way to request or pay for a better ranking.

Gemini searches less than the others. In our August 2026 study it stopped to search 0.3 times per answer, against 2.4 for ChatGPT. On 26 and 27 September, reached through its developer interface in the conversations about the 47 trips, it answered from memory 127 times out of 188; when the question named a hotel, it searched 142 times out of 168. When the Gemini app does search, it draws on Google’s Search index, as Google’s page on its crawlers states. What counts for Gemini, then, is what has been written about your hotel for some time, clearly and consistently: a page published yesterday does not correct an answer given from memory.

A ChatGPT app does not make your hotel appear in recommendations: it answers whoever opens it, and ChatGPT writes its recommendations from a web search. On 4 October 2026 we put 12 travel questions to ChatGPT with our own app installed. In 3 of them we left the app unnamed, and it stayed closed all 3 times. In the other 9 we typed “@Lumina” before the question, which is how a user calls an app by name, and ChatGPT opened it all 9 times.

The app earns its place later, with the guest who already has you on the shortlist and wants a room, a price and the conditions; our guide to ChatGPT apps for hotels covers that. Prices and availability reach the assistant through MCP, the protocol that connects it to the hotel’s systems: the guide to MCP and direct bookings explains how. As of 8 October 2026 our own app for ChatGPT and Claude is added by hand and is not yet in ChatGPT’s public directory. If you have one built, our page on ChatGPT app development for hotels lists what to ask the developer.

GEO, generative engine optimisation, is the name agencies give to the work of getting a business cited and recommended by the assistants. An agency can promise the work and the measurement. Nobody can promise a position, because these answers have no fixed positions.

The companies that write the rules guarantee no result. OpenAI’s page on its crawlers explains how to be read and is silent on how ChatGPT orders its sources; Google warns that a page meeting every requirement may still be left out. And on 3 October 2026 the same Maldives question produced a different winner in each of our three runs.

That goes for us as well. On 27 September 2026, with our hotel guide newly published, no assistant had read one of our pages in 1,364 answers. On 3 October OpenAI’s search returned one for the first time. So ask anyone selling you GEO for the list of questions, the assistants and the dates of the tests, how many times each question was repeated, and how many answers each figure is calculated from.

Checking your own hotel in ChatGPT takes 20 minutes and ten questions: two open ones and three checks, each put to ChatGPT and to Gemini. They tell you whether you reach the shortlist and whether the facts the assistants repeat are right.

Once the corrections are made, ask the same questions again in the same words. One run tells you what happened that day; a series tells you whether the answers are changing.