Hotel prices on ChatGPT: what it answers, searches for and cites, across 792 questions

Because on 31 October the co-founder of a travel company wrote to us. He had asked ChatGPT about a Las Vegas hotel, and the answer cited our page on the Venetian’s prices. That was why he got in touch, and it left us wondering how the page had ended up there.

In the screenshot, the answer called our page “an independent published comparison”, then reined itself in: “historical quotes, not current offers”. It went on to list what a fair comparison would need: the same date from Booking.com and from the Venetian itself, the same suite, the cancellation terms, the total to pay, and whether the 55-dollar resort fee was in or out. ChatGPT had just written our to-do list.

Our own logs already hinted at something. When ChatGPT opens a page while answering someone, it does so with an agent called ChatGPT-User, and OpenAI says that agent acts when a user asks a question and is not used to crawl the web automatically. In October, page for page, it opened our price pages far more often than our hotel pages.

So we put to ChatGPT the questions of a traveller trying to find out what a hotel costs: the price a night, the cheapest month, where to book, the total with taxes, one booking site against another, the hotel against the booking site. 792 questions about 149 hotels and 14 destinations, on 31 October 2026, then 136 of them again to see whether the answers held.

One thing to say once, and plainly: we questioned a ChatGPT model through the API, with web search switched on and a US location, not the app a traveller has on their phone. We wrote the questions ourselves, so the rates describe this test, not real traffic.

Yes, and it nearly always gives one: there is a price in 95 of 99 answers to “how much is a night”. What is missing is the night. Only 18 of 99 answers say which date the price is for, and 95 end by asking for yours.

A typical answer goes like this. Asked what a night at the Standard, High Line in New York costs, ChatGPT said roughly 400 to 500 dollars a night including taxes and fees, for some dates, and much more on busy nights; tell me your dates and I’ll look for an exact rate. Its source was Expedia. It is an honest answer and almost no use: which dates, which room, read when?

That finding killed one of our own ideas. In tests built on advice questions, the “where should I stay in…” kind, a price hardly ever appeared, and we said so in our piece on ChatGPT ads for hotels. We had concluded that ChatGPT simply didn’t have prices. Ask it for the price, though, and you get one. It lacks the date, and the date is exactly what a hotel holds.

It looks for the official site and the booking site together, and often adds the year by itself. Across 792 questions the model ran 1,776 searches of its own: 790 named a booking site, 616 contained the word “official”, 508 a year. The commonest shapes ran along the lines of “hotel name, official, Expedia, rates”.

Identical searches barely recur, 1,632 different forms in 1,776, but the pattern does: it wants to set the official source beside the one that sells. Reviews and forums it doesn’t look for: “review” appears twice, “reddit” never.

Then it comes down to who turns up first. Within an answer, the first page the search finds is cited 51% of the time, the second 23%, the fifth 6%. Much of the game is played before a word of the answer is written.

The booking sites and Google, well ahead of anyone else. Expedia is among the sources in 136 of 792 answers, Google Hotels pages in 125, Kayak in 86. Of the 596 questions about a named hotel, 379 cite at least one booking site or Google Hotels.

Each has its own question. Kayak takes the cheapest month, cited in 66 answers of 100; Expedia the price a night, 52 of 99; Google Hotels the question of where to book, 34 of 99, and the one between 2 booking sites, 32 of 99.

Forums and review sites ChatGPT finds and almost never cites: 1 citation in 4,911 pages found. On prices, Reddit stays bedtime reading.

On how the bill works and on “direct or booking site?”, where the hotel’s site is cited more often than any booking site. On the price a night and the right month, the hotel all but disappears.

Many of the official pages that get cited are of one kind: the Broadmoor’s hotel policies, Marina Bay Sands’ booking terms, Resorts World Las Vegas’s offer with no resort fee, La Mamounia’s FAQ. Pages that say how the bill works. In the United States that bill also has a rule: since 12 May 2025 the Federal Trade Commission has required a mandatory resort fee to sit inside the total price shown.

To my mind this is the lesson for anyone running a hotel. The booking sites have won “how much”, because they hold the price with its date. “How does it work”, with taxes, fees, cancellation and the perks of booking direct, goes to whoever writes it clearly on their own site.

Only where we have a price page, and I think because it is the missing page: a price with its date, in a table. Lumina is cited in 12 answers of 792, and all 12 fall among the 136 questions about hotels with a price page in the guide. On guide hotels without one: 0 of 172. The page cited is not always the price page: of the 12 answers, 6 cite the price page and 6 the hotel page, which on hotels without a price page is never cited at all.

Take Giraffe Manor, in Nairobi. Asked what one night costs with taxes and fees, ChatGPT answered with the figures from our page:

For 2026, one night at Giraffe Manor is US$1,413 per adult sharing a room—so US$2,826 total for two adults. The published rate is an all-inclusive rate, so taxes and standard service charges are included; there’s no separate resort fee listed.

ChatGPT’s answer in our test, 31 October 2026, quoted as given. Its only source was our page on Giraffe Manor’s prices.

There is a second reason, and it concerns the booking sites. We split the hotels with a price page into 2 groups: those the booking sites cover thinly, where the model’s search found one of their listings in at most 2 questions of 4, and the rest. Where booking sites are thin, our page is found in 32 questions of 40; where they are dense, in 47 of 96. The first group holds 10 hotels, and nearly all have a family likeness: Amanjiwo, Blackberry Farm, Eden Rock, Aman New York, Giraffe Manor, Ritz Paris, Amangiri, Claridge’s; the exceptions are Beaches Turks & Caicos and Atlantis Paradise Island. Many of them, I think, are booked by writing to the hotel rather than by clicking.

Position matters here too. When our page came 1st or 2nd in the model’s search, counting both times we asked, it was cited in 14 cases of 57; from 6th down, in 2 of 52. Of 40 price pages from other sites that ChatGPT cited, 16 had a table, 15 at least 5 prices, 13 a date in the text. None had a price in the title.

More ideas than survived, and that’s as it should be: each was something we could have sold to a hotel. These 4 don’t stand up.

And there is one thing we can’t explain. In our test the Las Vegas price pages were never cited, 0 times in 32 questions, counting both times we asked, yet one of them is exactly what the real app showed the co-founder who wrote to us. A model through the API and the app on a phone are not the same thing, and that zero proves no absence.

I’d publish myself the answer ChatGPT now fetches from Expedia, and put it where it looks. There are 5 moves, and each comes with the evidence that made me choose it.

A page like this replaces none of the others. For getting the hotel’s name into answers when the question starts from the trip rather than the price, there’s our guide on how to get your hotel on ChatGPT, and our August study of 3,288 answers from ChatGPT, Claude and Gemini shows how hard that is: on questions about the trip, only 189 answers of 1,380 named one of the world’s 50 best hotels. And if you wonder who sells your hotel below your own price, we measured that in is it cheaper to book directly with the hotel.

Inside ChatGPT and Claude, Lumina’s app shows the prices collected for the nights you ask about, with the day we read them. I’d still start with the hotel’s own site: it is the source ChatGPT searches for with the word “official”, and the hotel is the one that sets the night, the room and the bill.