ChatGPT apps for hotels: what they are, who has one and when you need one

A ChatGPT app for a hotel is a service the hotel runs, which ChatGPT calls while the guest is typing. The guest asks “do you have a pool villa from 12 to 16 January?”. ChatGPT passes the question to the hotel’s service, gets rooms, prices and conditions back, and shows them in the conversation, sometimes as a card with photos and buttons.

Without an app, ChatGPT answers from what it finds on the web: the hotel’s site, the booking portals, the reviews. It knows a great deal, but it cannot tell whether that villa is free tomorrow or what it costs tonight. The app is the direct line to whoever can.

The service is a program on the internet, which engineers call a server. It talks to the assistants in a shared language, the Model Context Protocol, or MCP. The same server answers both ChatGPT and Claude; our guide to MCP and direct bookings explains how it connects to a booking engine.

Almost every large hotel group already has an app in ChatGPT, and so do the first individual hotels, from Madrid to Palmanova. On 6 October 2026 the Travel category of ChatGPT’s public plugin directory held 388 listings; on 2 October it held 377, of which 79 had the word “hotel” in their name or description. Nearly all promise the same thing: search and book.

Wyndham, Radisson, Shangri-La, Barceló and Motel One are in the directory too, along with other individual hotels such as Hotel Moderno and Hotel Liabeny in Madrid. We read the listings on 2 October 2026. Accor’s announcement is in its press release, Mirai’s on its blog.

An app does not make ChatGPT recommend your hotel: in our tests ChatGPT opened one only when the user called it by name.

On 4 October 2026 we installed our own app in ChatGPT and asked 12 questions a traveller might ask, such as “which Maldives resort for our honeymoon?”. In 3 of them we left the app unnamed: ChatGPT answered from the web and the app stayed closed all 3 times. In the other 9 we typed “@Lumina” before the question, and it opened the app all 9 times. This was ChatGPT’s default model.

A traveller who has not installed the app would first need ChatGPT to suggest it. To see whether it does, we used Codex, OpenAI’s tool for developers, which draws on the same plugin directory. On 3 October 2026 we put 6 travel questions to it: it answered from the web and looked for a plugin in none of the 6. When it does look, it chooses among about 150 of the 3,915 plugins we counted in the directory, and for travel it finds three: Booking.com, Tripadvisor and Skyscanner. How ChatGPT decides what to suggest, OpenAI does not publish.

What OpenAI does say, on its page about review and distribution, is that a published plugin is found through a direct link or by searching for its name. A featured place or a suggestion during a conversation goes to plugins that show real-world utility and satisfied users; few receive one at publication, and developers cannot ask for it. In Claude every connector in the directory can be suggested, and the order depends on use.

A guest who has never heard of you finds you in the pages the assistant reads on the web. Given a travel question, ChatGPT runs searches of its own, much as a person would on Google, and answers from what it finds: on 27 September 2026 it did so in 186 conversations out of 188.

On 26 and 27 September 2026 we asked ChatGPT, Gemini, Claude and Perplexity about 84 hotels and read the 2,880 searches they ran. By our estimate three in four, 74%, were checking a hotel already on the shortlist: the breakfast, the parking, the walk to the centre. For those checks the assistants mostly read the booking portals.

To get onto the shortlist, your pages have to say plainly who the hotel is for and what it offers. In our study of 3,288 answers, from August 2026, when the question started from the trip and not from the hotel, only 189 of 1,380 answers named one of the world’s 50 best hotels. The guide on how to get your hotel on ChatGPT says where to begin.

A hotel’s own app is for the guest who already has you on the shortlist and has to decide which room, at what price, on what conditions. That is where the assistant stops today.

On 3 October 2026 we asked ChatGPT 6 questions about hotels already chosen. In 5 the answer ended with “ask the hotel” or “get it in writing”. When the assistant does answer, it often repeats what it finds: on 27 September 2026, across 11 hotels and 169 practical questions, we found 98 errors on web pages, and in 31 cases the assistant repeated the error.

With an app, the hotel answers, in the conversation and from its own data. There are two kinds of app. The portals’ apps, such as those from Booking.com and Tripadvisor, compare many hotels and help a traveller choose, and so does ours. A hotel’s or a group’s own app comes in afterwards: the guest has chosen, asks about availability and conditions, and goes on to book.

Today a hotel booking is completed on the hotel’s own site. OpenAI’s guidelines ask plugins to complete the purchase on their own site, and the groups’ apps do exactly that: ALL Accor searches in the chat and closes on ALL, Marriott hands over to Marriott, and Aman’s listing says so.

What the app can do is take the guest to the booking engine, the system that sells rooms on your site, with dates, guests and room already chosen. To show a true price and availability in the chat, the app needs a connection to that system. Without one it shows prices read on a given date, and has to say so. Ours does: it does not book, and every price carries the day we read it.

For the app to work, the hotel needs the answers a guest asks for before booking, in writing and with a date. This is the part we underestimated.

On 6 October 2026 we ran a single test: the same question, a family with children aged 2 and 9, put to ChatGPT with our app and without it. The answer without the app was the more useful one: it gave a resort’s minimum age for the kids’ club and its babysitting price, and our records had neither. We loaded 201 facts about families, each with the sentence from the page it comes from. On the next run the fuller answer was the one with the app.

Reading the sites of 30 resorts on 3 October 2026, we counted 288 answers that nobody publishes: the rate for a solo traveller, the distance from the beach to the reef, the price of a babysitter. We also counted 72 cases where sources contradict each other, often the official site contradicting itself: a 2018 press release says 20 minutes by speedboat, today’s FAQ says 30, and six portals still repeat the first figure.

Building our own app taught us that the work lies in the data and the details. The first server was answering inside ChatGPT the day after we started, on 3 October 2026; the rest of the week went on fixing three mistakes.

The first mistake: the card changed height at every tap and pushed the conversation up and down. Every card now has a fixed height. The second: on 4 October the card showed one price and ChatGPT’s text another, $2,101 and $1,101 a night for the same resort. Both were real, a pool villa and the cheapest room, on different dates, but the guest sees a contradiction. The third: in a single reply we opened a map, a table and the prices. We now open one card per reply, the one the question asks for.

One thing we did not expect: the server never receives the guest’s message. It receives the request ChatGPT composes on the guest’s behalf, with the language and the area. In a test on 3 October ChatGPT added a budget of $1,000 a night that the user had mentioned in another conversation.

To publish an app in ChatGPT you prepare a package, send it to OpenAI and wait for review; for Claude the path is similar and shorter. OpenAI’s procedure has five steps.

Three rules in the guidelines weigh on a hotel. A plugin cannot serve advertisements. The listing describes what the app does, with no comparisons and no claims that cannot be verified: we had to change “knows every resort in the Maldives” to “more than 180 resorts”. And payment happens outside ChatGPT, on the seller’s own site.

Our own app is added by hand today, in a minute, from the page that explains how to use it. It is not in ChatGPT’s public directory yet: the package has been ready since 7 October 2026 and is waiting for the company to be verified.

For an independent hotel it makes sense to get there in stages: the data and the pages first, then the app, often through the booking engine provider. Building your own is worth it for a group with several properties and one central reservation system.

The four routes, your own app, your provider’s app, a chatbot on the site and a custom GPT, are compared in the guide ChatGPT app or hotel chatbot. The guide to ChatGPT app development for hotels sets out what to ask whoever builds yours.