ChatGPT skips travel apps it's already connected to
Three travel apps, all connected and authenticated, and ChatGPT used none of them without a fight. Being available to the assistant is not the same as being called by it.
A traveler asks ChatGPT for a hotel near Times Square under $400 a night. The Booking.com app is installed, connected, signed in to a real account. ChatGPT ignores it and returns generic web results scraped off the open site. Ask again, by name, and it still balks. That is what Skift found when it ran the test in early July, and the detail that matters to any hotel commercial team is not that the app failed — it is how the machine explained itself.
The refusal. Told to use Expedia, ChatGPT said the app was "not available in this environment as an executable tool," though it was connected and authenticated to a live account. Booking.com got a longer brush-off: a confident paragraph claiming no "stable API-style search result export" for live inventory. Both explanations were false. The app worked fine once it ran. What broke was the model's willingness to call it, and it papered over that gap with plausible engineering language a hotelier has no way to check. Only a blunt "you are wrong" jolted it into trying. Cards, rates, deep-linked booking buttons — all there, once the argument was won.
Nobody owns the miss. Reached for comment, Expedia confirmed the behavior is a known bug and put it on the ChatGPT platform rather than its own app. Fine for Expedia's engineers. Useless for the hotel. Inside a company, a failed tool call becomes a ticket someone is paid to close. When it fails for a traveler mid-search, nothing happens: no ticket, no error even, just generic results and a guest who taps back to Google. The brand never learns it was passed over.
The tell is Claude. Same connectors, same accounts, different disposition. Claude offered the relevant apps up front with a connect button, asked whether to search the active Booking.com or Expedia link, and volunteered others — Wyndham, Tripadvisor — without being asked. It pulled the same card carousel with no arguing. So this is not a capability ceiling in agentic booking. The pipes work. Whether they get used comes down to which assistant the guest happened to open.
One caveat. This was one tester and a snapshot in time, and Skift notes chatbot behavior shifts session to session with model updates and prompt phrasing. The pattern still held across three apps and both platforms.
The stake. For two decades the fight was for position — in search results, app stores, metasearch. The AI app was pitched as the way past all that: connect once, and the assistant walks the traveler straight to you. Skift's test says connection is the starting line, not the finish. The call is discretionary, it varies by platform, and the brand cannot see when it is being passed over. A referral channel you can't audit and can't count on is not yet a channel.
Enjoying this analysis? Hospitality.today delivers daily insights on hotel distribution, AI trends, and travel commerce — straight to your inbox. Subscribe for free at Hospitality.today →