The Travel Search Moved Before the Booking Did
Harry Hayman reads two 2026 travel data points together and finds the disruption hoteliers are bracing for has already happened.
Every hotel operating meeting in 2026 has the same ghost sitting at the table. Somebody eventually says it out loud: what happens when the machines start booking the rooms. Budgets get held back, integrations get postponed, and a great deal of nervous energy goes into defending the reservation system.
The reservation system was never the target. Harry Hayman spent a week reading two sets of numbers that were published four months apart, and together they describe a change that already finished happening one step earlier than almost anyone is defending.
The two numbers that belong side by side
The first number comes from Phocuswright research reported in March. Fifty six percent of United States travelers used artificial intelligence for planning, booking or in destination help on at least one trip in the past twelve months. In the second half of 2025 that figure was 43 percent. In the first half of 2025 it was 33 percent. PhocusWire summarised the study, titled The AI Surge, with a line that does the work of a paragraph: the fastest shift in travel behavior just became the default.
Every generation posted double digit gains between the second half of 2025 and the first half of 2026. Among travelers who use these tools at all, 94 percent have used them for travel specifically. More than half expect to use them more next year.
The second number comes from an earnings call. On August 4, Booking Holdings chief financial officer Ewout Steenbergen told analysts that inbound referral traffic from large language models, paid and unpaid combined, accounts for less than one percent of total room nights. He was direct about the trend as well. That has not moved much recently, he said, with no material change over the last few months or quarters.
Read alone, either number produces the wrong conclusion. Read together they describe the actual situation. More than half of American travelers now research trips with a machine, and almost none of the resulting bookings arrive with a machine’s fingerprints on them.
What Booking Holdings actually said in August
The most useful sentence on that call was not the one about referrals. It came from chief executive Glenn Fogel, who noted that third party data shows the company’s brands turning up an awful lot inside the language models. He called the difference between the two observations interesting, which is executive language for something that does not yet have a name.
High visibility. Almost no referred bookings. PhocusWire’s account of the call records both facts in the same paragraph.
What that describes is a research layer that has quietly become invisible to attribution. A traveler asks a model for hotel options, reads the answer, then opens a browser and types the property name directly. Nothing in that sequence produces a referral. The company that appeared in the answer gets the booking and no analytics dashboard on earth records why.
Booking Holdings also disclosed that its direct mix held in the mid sixty percent range in the second quarter, roughly flat year over year, even under search engine pressure from Google’s display changes and AI overviews. Its published financial detail sits with investor relations. Flat direct traffic during a year of that much upheaval is not a reprieve. It is what stability looks like when a company is already visible inside the new front door.
The research layer is where the loss happens
Picture the moment as a guest actually experiences it. Somebody is going to a conference in an unfamiliar city. They ask a model which hotels near the convention center have a proper desk in the room, a bar open past eleven and a gym worth the name. They get four suggestions and a short reason for each.
If a property is not one of those four, nothing else in the funnel ever gets tested. The rate strategy is irrelevant. The booking engine is irrelevant. The photography, the loyalty programme and the abandoned cart email are all irrelevant, because the guest never arrived at the point where any of them operate. There was no cart to abandon.
That is the disruption. It happened at the shortlist, not at the checkout, and it happened to properties whose published information cannot answer a specific question in plain language.
The Phocuswright work adds one detail that should reassure any operator who reads it correctly. Half of travelers who encounter answers inside a search engine still click through to the source websites afterwards. The front door of travel remains open. What has changed is who gets to be one of the doors the traveler is shown, and that decision now happens before a single click is recorded anywhere.
It is worth naming the emotional shape of this, because it explains the misdirected budgets. An agent that books a room is a threat a hotelier can picture, and picturing it feels like preparing for it. A shortlist assembled without the property, in a conversation nobody logged, on a Tuesday afternoon, is not a picture at all. It is an absence, and absences do not show up in weekly reporting until the quarter closes light and everyone blames the market.
Models assemble those answers from whatever is written about a property across the open web: the hotel’s own site, its listing data, review text, structured markup such as the Schema.org Hotel vocabulary, and the descriptions distributed to every channel that resells the rooms. When those sources disagree, the model either picks one or leaves the property out of the answer entirely. Neither outcome is visible in a revenue report.
What an operator does about it this quarter
The work is unglamorous, which is probably why it keeps losing budget arguments to projects with better slides. It is also cheap by hotel standards.
Start with the property’s own website, because it is the one source nobody else controls. Room types described the way a guest would describe them rather than the way a property management system exports them. Amenity detail written out instead of implied by an icon. Rate rules stated in sentences. Neighbourhood language that matches what a visitor would actually type, including the landmark, the district name and the transit line, not only the postal address.
Then reconcile everything downstream. The same facts should appear on every channel that lists the property, in the same terms, with no version dated three years ago still circulating. Distribution hygiene has always been on the industry checklist, alongside the operational guidance published by bodies such as the American Hotel and Lodging Association. What has changed is the stake. It used to be about tidiness and rate parity. It is now about whether a property appears in the answer at all.
The test is simple enough to run before the next management meeting. Ask three different models the question a real guest would ask about the property, and write down what they get wrong. That list is the work order. Harry Hayman has watched hospitality operators run this exercise and come away quieter than they went in, because the errors are rarely exotic. They are old room counts, retired restaurant names and a closing time that changed in 2024.
Why this matters in a convention city
Philadelphia sells itself on the things that do not fit into a booking engine field. Walkability. Food. Music on a Tuesday. Visit Philadelphia built a whole marketing practice on that texture, and independent operators here have always competed on it rather than on price.
Texture is exactly what a language model is good at repeating and exactly what most hotel data files cannot express. A property with real character and a thin digital description will lose the shortlist to a chain box with better structured information, which is an outcome nobody in this city wants and everybody in this city can prevent.
That is the consulting argument Gemini Strategic Consultants has been making to operators through 2026, and it comes with an expiry date attached. Right now, accurate and specific property information is an advantage, because most properties do not have it. In eighteen months it will be the baseline, and the advantage will belong to whoever did the boring work first.