Ask ChatGPT to book a room, not pick one, and it switches to the OTAs

ChatGPT will quote you a hotel’s nightly rate. The interesting question is where it goes to get one — and whether that changes when you tell it your dates or your budget. We held the cities fixed (Paris, London, New York) and varied the shape of the question: plain “best hotels” asks, the same asks with check-in dates a month and three months out, “a hotel tonight”, a budget cap, an explicit “what’s the best rate”, and “which hotels have rooms available”. Eight conditions, two phrasings each, fifteen repeats — 719 live searches with web search on.

A date is a detail; “available” is a switch

Share of all citations that come from an OTA or metasearch site (Expedia, Booking, Kayak, Hotels.com, Skyscanner, HotelTonight…), by how the question was phrased. The top four are discovery questions; the bottom four carry booking intent.

The jump is roughly fiftyfold from the calmest discovery prompt to the most transactional one. And the OTA that benefits most is specific: across the four booking-intent conditions, the Expedia family (expedia.com plus its country domains) takes ~23% of all citations, more than four times Booking’s ~5%. The one exception is “a hotel tonight”, where the single most-cited domain is hoteltonight.com — ChatGPT reaching for the brand whose name matches the query.

The price it shows you has no name on it

Alongside the written answer, ChatGPT often renders a map panel of hotels with a nightly price on each card — Ritz Paris, “$5,193”. We logged the source field on every priced card across all 719 searches. Nearly three quarters carry an internal feed code, b1, with no link and no named provider. Only 3.2% of priced cards exposed a provider URL at all.

Provider field on priced map-panel cards, n = 5,099 cards. Tripadvisor — the one feed many assume sits behind these prices — labels just 5.5% of them.

So the headline panel rate is the least transparent number on the screen: shown confidently, attributed to nothing the user can click. The cited, checkable prices live in the written answer, and those trace back to the OTAs from the section above. The panel and the prose are two different pricing systems sitting in one reply.

It does carry your dates into the search

Behind each answer, ChatGPT fires off its own background web searches (the “fan-out”). When the prompt has booking intent, those searches almost always carry the actual dates — it isn’t decorating the reply with a date it ignored. Share of fan-out queries that contained the check-in dates or “tonight”:

A plain “best hotels” prompt never injects a date (there isn’t one). Note the gradient: an explicit “available on these dates” lands the date 100% of the time, while a date merely attached to a discovery question makes it into the search under half the time.

The budget word reshuffles the whole shortlist

Telling ChatGPT a budget doesn’t just filter — it changes which hotels it even considers. The median panel price collapses from the luxury default to genuine value stock the moment “under €250” appears, in every city:

Median nightly price across all map-panel cards, per city, open vs budget-capped prompts.

One thing it does not do is move a given hotel’s price around. For the 83 hotels that appeared in both open and dated prompts, the median price ratio was exactly 1.0 — the same property gets the same number whether or not you named a date. And across repeat runs of an identical prompt, a hotel’s price held within ~8% most of the time (though ~1 in 5 swung more than 25%). The number is fairly stable; the shortlist it sits in is what your wording controls.

Method & limits

719 live ChatGPT searches (web search on), captured via Bright Data, June 2026, across Paris, London and New York. Eight prompt conditions — open, near date (~1 month), far date (~3 months), “tonight”, budget cap, budget + date, explicit price, and availability — each written two ways and repeated 15 times per city. We English-only the prompts deliberately: a prior study showed query language doesn’t move ChatGPT’s sources. Every capture logs the citations, the fan-out search queries, and the structured map/hotel panel including each card’s price and provider field.

FAQ