Every hotel conference this year has the slide with the collapsing funnel. Every second trade piece says AI has already changed hotel distribution. Go looking for the number underneath, and the ones you find say something more useful than the slide does.
The company with the most to lose puts it under one percent
Booking Holdings is the largest distribution partner most hotels have. On its second-quarter earnings call on 4 August 2026, CFO Ewout Steenbergen put referral traffic from large language models, paid and unpaid together, at “less than 1% of total room nights”, and added that it “hasn’t moved so much recently, so no material change over the last few months or quarters”.
In the same call, CEO Glenn Fogel described how Booking’s own brands fare inside those systems: they are “showing up an awful lot”. He called it “an interesting difference between the two”.
That difference is what this article is about. Appearing in an answer and receiving a booking are two measurements on two different axes, and between them sit two conversions that nobody has published: from being named to being clicked, and from being clicked to being booked.
A hotel-side panel run by Lighthouse recorded AI traffic rising from 0.51% to 0.90% of total visits across more than 17,000 properties, with ChatGPT accounting for 98.8% of it. A separate measurement on the same provider’s 80,000-hotel network puts it at “just under 1% of total hotel website visits”. Those two are one company in two cuts, not two independent confirmations. The independent part is that Booking counts room nights and Lighthouse counts website visits, two firms with two different denominators, and both land under one percent. Adobe measured 194% year-on-year growth in AI traffic to US travel sites in May 2026, and in the same report that this traffic converts 28% worse than other sources while spending 70% longer per visit.
The research stage tells a different story from the booking stage. Phocuswright’s tracking has US travellers using AI for planning, booking or in-destination help rise from 33% in the first half of 2025 to 43% in the second and 56% by March 2026. People are using these systems to decide. The bookings have not moved with them, at least not yet, and not in any number a public company has reported.
What we measured, and what it is not
We read 39,833 answers about hotels across four brand sets, drawn evenly across the period, and extracted every source those answers carry. Roughly nine in ten of those answers name their sources: between 90% and 94% across the four sets. Mention rates cover May to 13 July 2026, citations run to mid-September.
This measures whether a brand appears in an answer and what that answer was built from. It does not measure clicks, and it does not measure bookings. Where those two things are known, they come from the companies quoted above, not from us.
Finding 1: there is no dominant source
| Share of citations | Domain |
|---|---|
| 4.1% | google.com |
| 2.3% | booking.com |
| 1.0% | all.accor.com |
| 0.9% | h10hotels.com |
| 0.9% | facebook.com |
| 0.8% | barcelo.com |
| 0.7% | melia.com |
| 0.6% | tripadvisor.com |
| 0.3% | reddit.com |
| 0.3% | expedia.com |
The most-cited domain in hotel answers reaches 4.1%. In the largest of the four sets, citations spread across 10,368 different domains. There is no list of ten places to be, and there is no single site that decides whether a property appears.
Two details in that table are worth more than the ranking. Booking.com is the only booking platform that registers at all: Expedia sits at 0.3%, hotels.com at 0.2%. And review sites and social networks barely register: TripAdvisor 0.6%, Facebook 0.9%, Reddit 0.3%, Instagram 0.3%. That runs against the advice most hotels are given, and against what we expected before the measurement. The first version of this article recommended maintaining review and forum presence for exactly this purpose. The data does not support it.
Finding 2: the gap between two hotels is wider than the gap between two assistants
Share of answers in which the brand appears, whole percentages, same questions per set:
| Set A (452 questions) | Set B (28) | Set C (40) | Set D (25) | |
|---|---|---|---|---|
| Gemini 3.8 Flash | 74 | 68 | 70 | 1 |
| GPT-5.6 | 68 | 67 | 64 | 4 |
| Claude 5 | 67 | 51 | 38 | 1 |
| Grok 4.5 | 57 | 43 | 32 | 1 |
| Perplexity | 50 | 28 | 63 | 0 |
Question counts are those active in the shared window, May to 13 July 2026, so that every column covers the same period.
Across the first three sets the assistant matters: a brand that appears in seven of ten Gemini answers may appear in three of ten from Grok.
Set D is a single independent property with no chain affiliation, and it is the sharpest number here. The figure that follows comes from a different window than the table above, its own full measurement period of 9 March to 22 July 2026, and covers nine models rather than five: the property appeared in 79 of 5,477 answers. That is 1.4%, against comparable questions that put other properties in two of three answers.
It is not evenly invisible either. Of those 79 mentions, 40 came from one model, GPT-5.6. Perplexity named it nine times, Google’s AI Overview eight, Gemini and Claude six each, Grok 4.5 and the classic Google top ten five each, and two of the nine models never did. The property effectively exists on one assistant.
One property is a data point, not a rule about independent hotels: one house, one market, one language, four and a half months. It still shows that the surface decides less than the property does.
Two Google surfaces sit beside that table and measure something else. Google’s AI Overview named the brand in 20% of cases in Set A, the classic top ten in 12%. Both answer “does the brand appear”, neither on the same surface as an assistant’s reply, and the Overview box does not appear for every query. They belong next to the table rather than in a ratio with it.
Finding 3: the referral measurement sees less than it reports
An attribution study published on 2 September 2026 ran 62 days on one Paris hotel. Of 35 visitors who told an on-site survey that an AI assistant had sent them, 11 carried a traceable tag. The other 24 carried nothing: twelve arrived bare, twelve arrived from google.com untagged. The author’s line: “The AI you can measure is the AI that tags itself.”
One property and 68 survey answers are a weak base for a claim about volume, which the author says himself. They are a reasonable base for a claim about detection. So both halves hold at once: the referral share is small, and the usual tools measure it smaller than it is.
The counter-argument, which is stronger than the slide
The strongest case against the collapse story is not scepticism, it is data. SiteMinder, across more than 130 million bookings in 20 markets, found direct booking revenue share “staying within 1.5 percentage points of the prior year in 95% of markets”. Average booking values in that data: 516 USD through a hotel website against 312 USD through an OTA. A consultancy analysis from 8 July 2026 adds a second measurement, Expedia Group’s own Path to Purchase research, which metered over 70,000 travellers passively rather than surveying them: they viewed 141 pages of travel content before booking, rising to 277 in the US, across roughly 45 days. If AI were collapsing the funnel into one decisive exchange, that number should be falling.
Its conclusion fits our own data better than either the burning funnel or the dismissal: “The funnel didn’t collapse. It reordered itself.”
Marriott and Hilton put the same thing in more careful language. Both had carried AI in their risk factors a year earlier, as an internal technology and security matter: their own use of it, and attackers using it. What is new in the February 2026 filings is the link to distribution. Marriott writes that AI capabilities at travel intermediaries “may change the way guests plan, book, and pay for travel”, with the consequence of “potentially eroding brand loyalty, increasing distribution costs”. Hilton names intermediaries attracting customers “including with the aid of AI, which may adversely affect our ability to sell rooms directly”. Neither attaches a number. A 10-K is a liability document: both groups considered the risk material enough to disclose and neither could quantify it.
What to do with this, in three steps you can check yourself
1. Put the facts where a machine can read them. Room types, distances, breakfast terms, family suitability, accessibility, as text on your own pages rather than inside images or a PDF. Adobe’s content-visibility measure puts hotel product pages at 73% machine-readable, which reads better than it is: roughly a quarter of what you published is not being read.
2. Stop optimising for a source list that does not exist. No domain in our measurement reaches 5%, and the review and social sites that usually top the recommendation lists sit below 1%. The practical consequence is that your own pages and your own structured facts carry more weight than any single third-party profile, because they are what an assistant reaches for when it needs a specific detail.
3. Check more than one assistant, and more than one language. An independent travel study in April 2026 found that all five engines it tested picked a different number one brand, and that Spanish-language answers favoured global booking platforms (Booking.com +34%, Expedia +31%) while US-centric premium brands lost ground. A single check in English describes one system in one language.
What none of this shows
- Nobody has published an independently audited visibility measurement for hotels. Every index we found comes from a company that sells measurement or consulting. This one included. Read all of them as vendor data, ours no differently.
- No hotel group has published an AI share of bookings. The risk is disclosed, the number is not.
- Neither OpenAI nor Google publishes travel query volume. They announce features, not demand. Google’s hotel booking inside AI Mode launched on 27 August 2026 in the US in English, and its own footnotes exclude the European Economic Area, so a hotelier in Spain is outside the booking feature and inside the visibility question.
- The widely repeated claim that more than a third of travellers now start with an AI engine has no source in the release that popularised it. We looked.
- Our figures are Mediterranean, cover four brand sets, and count appearances in answers. The two conversions after that, from appearance to click and from click to booking, are not measured here by anyone.
The question worth asking instead
Not whether AI has already changed hotel distribution. By the only figures the companies themselves have published, the booking split has not moved yet. The answerable question is narrower: when a guest describes their trip in their own words, does your property appear in the answer, and what was that answer built from?
Both are measurable per destination and per language, which is exactly the part a single national average hides.
→ How Truffle tracks this per destination and language
Sources
All external pages retrieved on 17 September 2026.
- Booking Holdings Q2 2026 earnings call, 4 August 2026. Quotes taken from PhocusWire’s write-up of the call; we did not obtain a full transcript.
- Lighthouse / The Hotels Network panel, more than 17,000 hotels, analysed by Nicolas Sitter, “The ChatGPT Direct-Traffic Explosion for Hotels (2026)”, 21 May 2026, updated 3 June 2026.
- Lighthouse 80,000-hotel network, cited in White Sky Hospitality, 8 July 2026.
- Adobe Analytics, “AI travel traffic surges as engagement hits new highs”, 17 June 2026. The 86% satisfaction figure in that report comes from a survey of more than 5,000 US consumers rather than from traffic data, so we do not use it.
- Phocuswright Research, published via PhocusWire, March 2026: US travellers using AI for planning, booking or in-destination help, 33% (H1 2025), 43% (H2 2025), 56% (March 2026). Measures usage, not bookings.
- Marriott International Form 10-K FY2025, filed 10 February 2026, and Hilton Worldwide Form 10-K FY2025, filed 11 February 2026, both read in the SEC filing rather than in press coverage. The prior-year filings from February 2025 carried AI as an internal technology and security risk, without the link to distribution.
- SiteMinder, more than 130 million bookings across 20 markets: direct booking revenue share “staying within 1.5 percentage points of the prior year in 95% of markets”, average booking value 516 USD through a hotel website against 312 USD through an OTA. The 36% figure in the same release comes from a separate survey of 700 hoteliers and is not used here.
- Expedia Group Path to Purchase research, metered behaviour of more than 70,000 travellers, cited in White Sky Hospitality, 8 July 2026. We did not read the Expedia original.
- Nicolas Sitter, “AI Traffic Is Undercounted (2026): Four Measurements of One Hotel”, 2 September 2026.
- AIVO Research, Travel & Hospitality AI Visibility Report, April 2026: 1,500 queries, 5 engines, 18 brands, English and Spanish.
- Google, “How to book travel with AI Mode in Search”, 27 August 2026. Its own footnotes exclude the European Economic Area.
- Our own measurement: 39,833 answers across four hotel brand sets, sampled evenly; mention rates May to 13 July 2026, citations to mid-September 2026. Mediterranean market focus.
A note on sources 2, 3, 9, 10 and 12: every one of them, ours included, is published by a company that sells measurement, consulting or software in this market. None has been independently audited. That is the state of the evidence, not a disclaimer.