ChatGPT app development for hotels: what to ask before you sign

ChatGPT app development for a hotel involves three jobs and some paperwork: a server that answers the assistants, the hotel’s data in a form the server can read, the cards a guest sees in the conversation and, last, publication.

The server is a web address on the publisher’s domain, and one server answers both ChatGPT and Claude. It offers tools, one for each kind of question it can answer: which rooms are free on given dates, how to get there. The guest types, ChatGPT picks a tool, and the server sends back the data and, where it helps, a card. Our app launched on 3 October 2026 with 7 tools and had 21 by the 8th, plus 4 used only by the cards.

An app of this kind is for the guest who is already weighing you up: on 4 October 2026 ChatGPT opened ours in 9 questions out of 9 when we called it by name and in 0 out of 3 when we did not. What gets a hotel into the answers in the first place is measured in our study of 3,288 answers from ChatGPT, Claude and Gemini.

Six parts of the project drive the cost and the timetable of a hotel’s ChatGPT app, and the heaviest is the connection to the system that holds prices and availability. Lumina publishes no price list for this work: the figure for your hotel can only come from someone who has looked at your booking engine, and it is worth asking for it line by line.

In early October 2026 the machines cost us little: collecting one resort’s pages cost, in one test, $0.53 and 3.5 minutes; having a model read them, $2 on average. The bill comes from people, who decide which of two sources is right and rebuild a card that misbehaves inside the conversation.

Our own timetable ran from 2 to 7 October 2026. The server was online after the first night and the first test inside ChatGPT came at midday on the 3rd. On the 5th our checker found 0 errors in the package, the file holding the name, logo, description and instructions that is uploaded to OpenAI; on the 7th we closed version 1.7.2, by which time the cards had gone from version 3 to version 45. This was an app that reads and does not book, needs no sign-in and has no booking engine behind it, built by several AI agent sessions working in parallel. Much of the calendar went on things the documentation does not cover: whether the assistant decides to call the app, cards that shift about, duplicate copies of the app on one account, wrong data.

Your project will have two waits that ours did not: access to the booking engine, which your provider decides and which you should ask for first, and review, on which OpenAI writes that timelines may vary and that it cannot take requests to speed things up. We have yet to go through it: on 7 October 2026 our package was ready and unsubmitted, because the company’s identity verification, which OpenAI requires before any submission, was still outstanding. Start yours on day one.

Prices and availability come from the system that holds them: the booking engine, the channel manager or the property management system (PMS). Your provider grants the access, and a developer can only ask for it.

Our app has no such connection: it shows prices read on a stated date and says so. On 4 October 2026 we put 12 questions to it, each with a budget, and in none of them were two prices comparable: one source gave the price for the right party but with taxes extra, another included taxes but not the number of guests. A traveller who asked for a villa under $1,500 a night was offered one at $2,645, and for families the app had only the price for two adults.

So the first call is to your booking-engine provider. The Hotels Network, SiteMinder and Mirai already describe, on pages we read on 8 October 2026, an app or a connection that carries rates and availability into the assistants: ask what is live today under your contract. How it works, and why payment still happens on your own site, is in our guide to hotel MCP and direct booking.

The data to prepare is the set of answers a guest wants before booking, written down and dated: how to get there, the rooms, children by age, meals, the price with taxes and transfer, building works and closures. Where a fact is missing, the app can lose to ChatGPT on its own.

It happened to us on 6 October 2026 with a family’s question, one of 5 we put to ChatGPT with our app and without it, with the judging criteria written beforehand. ChatGPT alone had found on the web the age from which one resort offers babysitting; we held 60 kids’ clubs with no ages recorded, and the answer with the app had taken 170 seconds against 92. Overnight we loaded 201 facts about families and, asked again, the app won, in 42 seconds: the full account is in our guide to ChatGPT apps for hotels.

Two things then misled us. The new facts seemed to make no difference, because the app was answering from a copy computed the night before: ask how long a corrected fact takes to reach the answers. And on 6 October 1,072 facts about 31 hotels turned out to be stored under names the system no longer read: as far as the app knew, one hotel had nothing to do in the rain, while its spa, gym and lounge sat in our records.

The rest of the work is choosing between sources that disagree. On 3 October 2026 our first method, reading everything and then choosing, got 49% of 77 checkable answers wrong; the fourth, which starts from the official site and accepts “they do not say” as an answer, left 1 serious error in 171. The commonest cause was old documents left online undated: one from 2020 gave a villa as 446 m², the 2025 one as 860. Take them off your site, or date every one.

Seven mistakes cost us days: six showed only inside a real conversation, in ChatGPT or in Claude, and the seventh was on the server. Each is now an acceptance test, to copy into your contract and rerun after every change.

OpenAI also lists what to check before packaging, on its page about connecting and testing a server: among other things, that the cards render without errors.

OpenAI asks for a verified identity, a package whose fields have exact limits, eight test cases, a video and four public pages; Anthropic asks for less. A test case is a question written down with the result you expect. The rows come from OpenAI’s submission procedure, review requirements and guidelines, read on 8 October 2026.

Once the app is approved, you choose when to publish. Featured placement and suggestions during a conversation cannot be requested, OpenAI writes, and a press release for the launch has to be agreed with its press office first.

We put our own five positive cases to the live server on 5 October 2026, using a script that checked 24 things in the answers. That same day two older cases that had never been rerun no longer held: one returned 235,000 characters, the other showed three photos of a spa, a villa and a kayak beside two videos of the coral reef. Rerun all eight before every submission.

Anthropic asks you to submit the server from Claude’s developer portal, from an account on a paid plan, as a connector identified by its URL alone: this is on its page about publishing. After an automated scan it is listed, by default, as a “Community” connector, and every connector in the directory can be suggested in a chat when relevant. Submit from the hotel’s account, because the listing belongs to the organisation that submits it; we have submitted nothing to Claude yet.

The brief for a developer has seven parts, and it is the document to send to whoever is quoting.

Three kinds of provider develop ChatGPT apps for hotels: your booking-engine provider, an agency or developer who knows MCP, and a hotel group’s own technical team. Whichever you choose, see that the domain the server runs on, the account that submits the app and the data are yours. If you are not yet sure an app is the right route, start with our comparison of a ChatGPT app, a chatbot on your site and a custom GPT.

Whoever runs the server maintains the app, and the contract should say how quickly they respond. Data, prices and cards change on the server; the name, logo, description and instructions live in the package, which has to be uploaded again after every change and, once the app is public, reviewed again, as OpenAI’s submission procedure sets out. OpenAI scans the server daily: a new tool waits until it passes the checks, a changed one carries on under its old definition, and an app that is unstable or neglected can be removed from the directory. On 4 October 2026 a database change of ours broke the answers twice, for about 2 and about 5 minutes: ask how they roll back.

Whoever you approach, put these six questions to them.

We built our own, for ChatGPT and Claude, and everything above comes from it: you can try it in ChatGPT and in Claude in a minute.