On 27 August 2026, Google started rolling out hotel booking inside AI Mode in the United States. A traveller describes the stay they want, compares options in the conversation and books without leaving Google. The launch partners are the companies you would expect: Booking.com, Expedia, Hotels.com, Priceline, Trip.com, Hilton, IHG, Marriott, Choice and Wyndham. Google’s own post spells out who carries the transaction: “the hotel or booking platform will act as the merchant of record and handle any customer service” (Google, August 2026⁠ (opens in a new tab)).

It is US-only and English-only for now. Flight price tracking from the same announcement already works in more than 180 countries and territories. AI Mode itself has been available in Australia since 8 October 2025 (Google Australia⁠ (opens in a new tab)), and Alphabet says it passed one billion monthly users this year (Alphabet, July 2026⁠ (opens in a new tab)). Google hasn’t said when, or whether, hotel booking will come to Australia. If it does, the question for a destination marketer is what the website is for once an assistant can take the booking.

Our view: much the same as now, with less room to be vague about it.

Where the booking happens

Most DMO and regional tourism websites don’t take bookings. They inspire, inform and send Visitors on to Operators, whose own booking pages or channel partners close the sale. Agentic booking moves the transaction from those Operator pages into the assistant, through a platform the assistant has a deal with. That is a real shift for Operators and their distribution costs. For the destination site, it changes less than the headlines suggest, because the destination site was never where the money changed hands.

What it does change is the value of the step before the booking. An assistant that books a hotel still has to decide which town, which weekend and which hotel, and it decides from what it can read.

What assistants read when they plan a trip

Travel is the category where AI assistants lean hardest on the open web. Similarweb found that 22.6% of ChatGPT answers about travel included web citations in May 2026, against a 6.8% average across all conversations, the highest rate of any industry it measured (Similarweb, September 2026⁠ (opens in a new tab)). Travel questions are local, seasonal and time-bound. The model’s training data goes stale in exactly the places a trip depends on: what’s open, what’s on, how long the drive takes after the bridge works.

That is the territory a destination website owns. The booking platforms hold inventory and prices. They do not hold the council’s road closure, the shoulder-season festival, the cellar door that opens only on long weekends, the morning ferry that stops running after Easter, or the fact that the walking track to the falls reopened in September, muddy but passable. Nor do they know which week after the school holidays the coast goes quiet. A destination site that states those things plainly, keeps them current and lets assistants read them is the source an agent needs before it books anything.

Who arrives from an assistant

When assistants do send people to websites, they send a particular kind of visitor. Adobe’s data on US travel sites shows traffic from AI sources up 194% year on year in May 2026. Those visitors were 21% more engaged and stayed 70% longer per visit than visitors from other sources, and they converted 28% less (Adobe, May 2026 data⁠ (opens in a new tab)).

Take the conversion figure at face value: AI-referred visitors bought less on the spot. For a booking site that is a problem. For a destination site, which rarely takes bookings, the more useful signal is the longer, more engaged visit. Those visitors are reading, and reading is what a destination site is for, which is why referrals to Operators, not bookings, are the right measure of whether it works.

Keep the Australian numbers in proportion too. The Europ Assistance Holiday Barometer found 12% of Australians had used AI tools to plan or book a holiday, against 22% of Americans (reported by The Canberra Times, April 2026⁠ (opens in a new tab)). Adoption here is early, which argues for preparing calmly this year.

Four things to do while it is still early

1. Make the facts unmissable. Opening days, seasonal closures, drive times, event dates, accessibility details and booking links belong in plain sentences on the page, not only inside a widget, an image or a PDF. An assistant reading your site at speed takes what is stated, and an AI-referred visitor checks the same page before they commit.

2. Keep Operator Products current and linked. Every Product page should send the Visitor somewhere they can act: the Operator’s own booking page, a phone number, an enquiry form. When an assistant cites your Product page, that link is how the Operator gets the business. Stale profiles are the fastest way to lose it (see ATDW in 2026).

3. Let the right agents in. Check that your robots.txt allows the search and on-demand fetchers of the assistants you want to be found by. A blanket block set by a security tool can quietly remove you from the answers. Our llms.txt explainer lists the crawlers by company and what each one does.

4. Measure the visitors assistants send. Set up an AI referral segment in GA4 this month, so you have a baseline before summer: Measuring AI referrals in GA4 walks through it.

On Roam

On every Roam site, AI Discovery is live at no extra cost: an AI agent that visits receives clean, structured markdown instead of raw HTML, so the facts above are easy to read. Concierge AI, an optional add-on, lets a destination answer Visitors directly from its own Products and events. See how AI Discovery works.

For the longer view of how AI answers are changing the path to a destination, read How AI search is changing the way travellers find destinations.