“Best price guaranteed”?What hotel sites promise, and what ChatGPT makes of it
TL;DR: Across 5,981 independent hotel websites in California, Florida and New York, 9.1% promise a best-price guarantee, and the “better conditions” on offer for booking direct are most often a discount: 15 hotels mention champagne. Asked whether the best price is on the hotel’s own site, ChatGPT mostly says “not necessarily”. In its own quoted prices, direct loses to the discount site Super.com and splits roughly evenly with Booking.com, Expedia and the other big OTAs.
The oldest line in hotel marketing, checked
Every hotelier has heard the pitch: book direct, it’s cheaper, and you get “better conditions”, something the OTAs can’t give you. A glass of something on arrival. A later checkout. A person who reads your special request. Most hotel websites carry some version of the promise, or so I assumed.
This study asks three questions, in order of how cheap they are to answer. How many hotel sites actually promise their best price is direct, and what do they offer on top? When you ask ChatGPT whether that promise holds, what does it say? And when ChatGPT quotes prices, who comes out cheaper?
About one independent hotel in eleven promises the best price
I took every hotel with a website in California, Florida and New York from this site’s hotel database, 13,042 in all. Those three states are where the database is deepest in the US. I read each independent hotel’s homepage, plus up to two pages it links to about booking direct. 5,981 homepages were readable. Chains were sampled a few pages per brand, because a chain’s guarantee is one corporate policy printed on a thousand property pages.
Three things were counted separately: an explicit guarantee (“Best Rate Guarantee”, “lowest price guaranteed”, “price match”), softer “book direct and save” messaging, and perks promised for booking direct. A perk only counts when a book-direct phrase sits next to it, since “breakfast included” on its own is an amenity.
Homepage or a linked book-direct page. Whiskers are 95% intervals. Chain brands, counted once each: 13 of 39 state a guarantee.
The guarantee is a minority habit: 9.1% make it, and 17.1% make any price promise at all. When it’s there, it isn’t always easy to find. 19% of homepage guarantees sit in the footer, and 38% link to terms explaining how to claim it. Only 5.2% of hotels have a dedicated “why book direct” page at all.
The ten most common engines and tools on independent homepages, plus hotels where none was detected. Hover a bar for n.
“Better conditions” mostly means a discount
Share of all 5,981 independent hotels. Counted only next to a book-direct phrase.
If you were counting on the bottle of champagne, I have bad news: 15 hotels out of 5,981 mention one. The most common written pitch is “it’s cheaper here”, usually as a percentage off. Breakfast, free cancellation, a later checkout and an upgrade each show up on two percent of sites or fewer.
Perks may well exist at the front desk, or for the guest who asks nicely. They just aren’t written anywhere a booker, or a model reading the page, can find them.
ChatGPT hears the promise and doesn’t sign it
Next I took all 547 independents that promise a guarantee, and 547 that make no price promise at all, paired one for one on state and star rating. Each got two questions in ChatGPT:
- “Is the best price for hotel in city on the hotel’s own website?”
- “Where’s the cheapest place to book hotel in city for two nights in November?”
That made 2,188 answers. Each was read in full by a model and labelled against a fixed schema: the verdict, every price quoted and who it came from, the channel called cheapest, and what it told the user to do. The labeller never knew which group a hotel was in.
“Not necessarily” is ChatGPT’s house answer, guarantee or no guarantee. The promise does register: ChatGPT mentions it in 48% of answers about guarantee hotels, against 6% for the rest, and it cites the hotel’s own site more often.
“Tells you to book direct” is the softest label in the study: two independent readers agreed on an answer’s recommendation 81% of the time, against 97% for the guarantee row. Own-site citations aren’t a judgement call: they come straight from the links ChatGPT attached.
Who is cheaper, according to the prices ChatGPT quotes
About two answers in three quote at least one price, and ChatGPT says where it found them. Google Hotels leads (29% of all answers on its own, more in combination), ahead of OTA pages, hotel sites and metasearch. When one answer quotes both the hotel’s price and a third party’s on the same basis (per night with per night, total with total), it has run a tiny price test for us, and the outcome can be counted from its own figures.
The rest of each bar is “direct cheaper” plus a sliver of “equal”. Pooled over both questions. Hover for n.
Stop at the first pair and you have a tabloid headline: hotels that promise the best price are undercut in 84% of ChatGPT’s comparisons. Then look at who does the undercutting.
Super.com, a discount travel site, is the cheaper seller in 53% of those cases, and many of its deals come with conditions: member pricing, app-only rates, non-refundable terms. Take Super.com out and the third party still wins more often than not. Restrict the comparison to the big OTAs and it’s close to a coin flip: direct is cheaper in 47% of comparisons for guarantee hotels and 55% for the rest. When a third party is cheaper for a guarantee hotel, the median gap is 17% of the hotel’s own price.
One result I didn’t expect, and can’t fully explain: guarantee hotels come out worse than no-promise hotels in every version of this comparison. My best guess is that hotels confident enough to print a guarantee are also the ones with the widest OTA distribution, so there are simply more third-party prices around to undercut them. It’s a guess.
Ask for your dates, and ChatGPT gets careful
The prices in Part 3 are ChatGPT’s. To check some against reality, I read hotels’ cheapest public rate for Thursday 12 to Saturday 14 November, two adults, straight from their own booking engine (SynXis, Sabre’s system), and at the same moment asked ChatGPT for the cheapest price for those exact dates and where to book it.
It didn’t go far. Of 67 SynXis hotels in the sample, the engine answered for 15; the rest sit behind bot protection that shows an automated browser a blank page, and I left it at that. After removing duplicates and one bad read, 12 hotels remain, all in New York State and Florida. Treat this as a spot check.
| Hotel | Engine, 2 nights before tax | Engine, with taxes & fees | ChatGPT's dated answer |
|---|---|---|---|
| Hotel Belleclaire | $1,197 | $1,459 | ~$1,279 (before tax); book or check directly |
| Tarrytown House Estate | $304 | $405 | ~$322 (with taxes); Vio; also check the direct offer |
| The Evelyn | $1,135 | $1,397 | ~$1,207 (before tax); check the official site first, then OTAs |
| Walker Hotel Greenwich Village | $1,420 | $1,729 | ~$1,100–1,200 (unclear); direct (hotel says direct is its lowest), Google Hotels |
| 6 Columbus - Central Park Hotel | $523 | $676 | No price given |
| Henderson Park Inn | $840 | $1,098 | No price given |
| Hotel Beacon | $989 | $1,146 | No price given |
| The Inn at Fox Hollow | $788 | $880 | No price given |
| The Rockaway Hotel | $495 | $676 | No price given |
| The Roundtree Amagansett, a Small Luxury Hotel | $1,490 | $1,775 | No price given |
| The Sherry-Netherland | $1,938 | $2,231 | No price given |
| The Wallace | $1,434 | $1,653 | No price given |
The clearest result isn’t a price. Given exact dates, ChatGPT declined to name one for 8 of the 12 hotels, saying it couldn’t see a live rate for those nights and wouldn’t invent one. Where it did estimate, it landed a few percent above the hotel’s own rate twice (The Evelyn, Hotel Belleclaire), below it once (Walker Hotel, whose cheapest direct rate that night was non-refundable), and pointed to a cheaper third party, Vio, once (Tarrytown House).
My read: the undated prices in Part 3 are whatever Google Hotels happened to show ChatGPT, for dates nobody chose.
If you run a hotel, or book one
For hotels
- A guarantee is visible to ChatGPT and changes the conversation. It brings your site into the answer and gets the guarantee repeated. Whether it then calls you cheapest depends on the numbers it finds, (that is Part 3).
- The promise is only as strong as your rate on Google Hotels. That’s where most of ChatGPT’s numbers come from. If a discount site sits below your official rate there, ChatGPT will find it and say so.
- Write the perks down. If breakfast, a later checkout or a bottle on arrival are real direct-booking benefits at your hotel, they belong on the page, where a guest and a model can read them. Right now almost nobody does this.
For travellers
ChatGPT will mostly tell you “it depends, give me your dates”, which is the honest answer. Its quoted prices are for dates you didn’t pick. A guarantee is worth claiming if you find a lower price elsewhere, but the comparing is still your job.
Methodology
- Hotels: 13,042 hotels in California, Florida and New York with a website, from a Google Places–derived hotel database; websites that were really OTA or social pages were excluded. These three states are the database’s best-covered US ground, about nine in ten of the US hotels it holds, though the coverage still leans toward the big tourist cities.
- Fetching: homepages read on 2026-10-03; 6,262 of 7,096 pages were readable after retrying bot-walled and JavaScript-only pages in a real browser. Independents were all fetched; chains three pages per brand site, analysed at brand level.
- Claim detection: visible text only, English and Spanish phrasing, with negative controls (“we guarantee a great night’s sleep” must not count). A hand check of 25 random guarantee matches found all 25 genuine, and rendering 398 random pages in a browser found no guarantee that appears only after JavaScript runs.
- ChatGPT: answers scraped from the US, logged out, on 2026-10-03. 547 guarantee hotels and 547 no-promise hotels matched on state and star rating, two questions each.
- Reading the answers: each answer was read by a model against a fixed schema, blind to the hotel’s group, with a validator rejecting incomplete labels. A second, independent labelling of 120 random answers agreed on who was cheaper in 98% of cases (Cohen’s κ 0.94).
- Comparisons: only prices quoted in the same answer, on the same basis. Intervals are Wilson 95%.
Limits
- ChatGPT’s prices are ChatGPT’s, mostly Google Hotels snapshots for other dates. The engine check covers 12 hotels.
- A homepage isn’t a hotel’s whole offer: member rates, emails and front-desk perks are invisible here.
- The two groups are matched on state and stars only; guarantee hotels may differ in size, distribution and price level.
- One day, logged-out ChatGPT, one country, English.
FAQ
Summarize with AI
More on hotels, prices and ChatGPT
Disclosure: I work at Lighthouse, which owns The Hotels Network, one of the booking tools in the engine chart. It was counted with the same code as everyone else.