The assistants recommend your hotel. Then they send the guest to an OTA.
GapWatch measures what AI assistants tell travellers looking for a hotel — and, for every answer, where it sends the booking. For an owner-operator, that is a margin measurement rather than a marketing one.
The finding
Across 120 answers to hotel questions in two North American cities, measured across four assistants:
| Where the answer sends the guest | Share of answers |
|---|---|
| An online travel agency or metasearch | 77.5% |
| A brand site — Marriott, Hilton, IHG | 29.2% |
| The property's own booking page | 0% |
Zero. On any of the four systems.
The part that makes it fixable
In 30.8% of answers, the assistant already tells the guest to book directly with the hotel.
It is not steering people away from you. But when it needs somewhere to send them, it reaches for TripAdvisor or Booking.com — because the property's own domain is not in the material it reads.
The recommendation you want is already being made. It simply arrives without a direct channel attached.
The arithmetic
At 15–25% commission, a $250 room night gives up $37.50 to $62.50 on every OTA booking. That is the cost of a guest who had already chosen you: the assistant did the persuading, the platform collected the toll.
Unlike a distribution contract, this one is not fixed. It is a function of what a property publishes, and where.
What clients ask us
Does this apply to a hotel inside a major flag?
It changes the picture rather than removing it. In our measurement the brand site collected direct bookings in 29.2% of answers — better economics than an OTA, but still not the property's own page, and not the property's own guest data.
What about a property that just opened?
That is the most interesting case and the shortest window. A property weeks old has almost no accumulated review or citation history, and how quickly the answer layer picks it up is measurable — but only while it is still new.
Do you do this in French?
Yes, and separately rather than translated. Answers in different languages draw on substantially different sources.
What about apartments and mixed-use?
Rental discovery behaves differently from hotel discovery. In one national apartment measurement, complaint and review content outweighed ordinary listing content by several times — a very different balance from the properties we measured alongside it.
What an engagement looks like
A frozen baseline. Monthly measurement of where answers send the booking, by property and by market. The ranked list of sources those answers are built from. Then the work: the property pages, the structured data, the review and directory layer, and the content that makes a direct channel available for an assistant to point at.