Back to all posts

What AI can already book, and what it still can't

By Wendy Harris·17 September 2026·5 minute read

An editorial illustration for the article What AI can already book, and what it still can't.

An operator asks a reasonable question this autumn: is there some new AI booking agent I need to plug my property into, and will I fall behind if I do not?

The honest answer is not yet, and quite possibly not soon. But something did change this year, and it is worth understanding what actually happened rather than what the coverage around it implies.

What Google actually launched

On 27 August 2026, Google added a checkout step inside AI Mode in Search. A traveller in the United States can describe the trip they want, compare hotel options inside the conversation, pick a room, review the cancellation terms and pay through Google Pay, all without leaving the chat window (ppc.land, 27 August 2026).

Ten hotel partners are live at launch: Booking.com, Choice Hotels, Expedia, Hilton, Hotels.com, IHG, Marriott, Priceline, Trip.com and Wyndham. It is US only and English only for now, with more partners said to be activating over the following weeks.

Look at that list again. Every name on it is a chain or a global OTA. There is no route in it for an independent holiday cottage, a family-run inn or a small attraction to be booked directly through AI Mode. If a guest completes a booking through this feature at all, they are completing it on Booking.com's or Expedia's inventory, the same inventory those sites have always sold. For a small visitor economy business, this launch does not open a new booking channel. It automates one that already existed, for the businesses that were already on it.

What ChatGPT tried, and then stopped

OpenAI went the other way. It launched Instant Checkout, a button that completed a purchase inside the chat, for a small set of US sellers in late 2025. By March 2026 it had pulled the feature back out of ChatGPT's main interface and moved checkout to third-party apps instead of handling it directly (Skift, 5 March 2026). The reasons reported were specific to travel: prices that change by the minute, cancellation terms that vary by rate, payment across currencies, and the servicing that has to happen when a booking goes wrong after the fact. Several travel companies, Booking.com, Expedia, Skyscanner, Accor and Lighthouse among them, built their own apps inside ChatGPT instead, so a transaction that happens there still runs through those existing platforms.

So the two largest AI surfaces have arrived at roughly the same place from different directions. Booking is starting to happen inside AI conversations, but it is happening through the businesses that already had the scale and the payment plumbing to do it. Neither system has built a door for a small operator's own website to transact directly, and on the evidence so far, neither looks close to it.

Why this matches what travellers are actually doing

This lines up with the shape of AI travel use already visible in survey data. In IMG's 2026 travel outlook survey, of the people likely to use AI tools to help plan travel this year, 75% expected to use it for recommendations and 13% for booking (IMG travel outlook survey, 2026). Most of what is actually happening inside these tools today is a conversation about where to go and what to do, not a transaction. The checkout features are real, but they sit on top of a much larger layer of recommendation, and that layer is the one every visitor economy business is already inside whether it meant to be or not.

That is also the layer with no membership list. Nobody has to sign a partnership agreement to be recommended. An AI system answering "somewhere quiet near the coast that takes dogs" is working from whatever it can find and trust about your business, wherever that information lives: your own site, your Google Business Profile, an OTA listing if you have one, a review platform. Being accurate and specific there costs nothing and needs no integration.

What this means for your website this autumn

Do not buy an "AI booking agent" integration if one is pitched to you. That door is not open yet for an independent operator, on either Google's or OpenAI's current approach, and paying to be ready for a channel that does not exist is money spent on the wrong problem.

Do keep your own booking widget working the way a crawler expects. A related scan of 39 visitor economy sites this summer found every one with a booking widget had it in a form a crawler could reach, which is the right baseline to hold, not a reason to stop checking it.

Do keep your basic facts, opening hours, location, what kind of place you are, consistent everywhere an AI system might pull them from. That is the same conclusion the structured data findings pointed to, arrived at from a different direction. Whether a guest ends up booking direct, through an OTA, or eventually through a chat window, the recommendation has to happen first, and the recommendation runs on the same plain facts either way.

The direction, not the deadline

Nothing here needs action by a date. There is no equivalent of a Cloudflare cutoff to watch for. What is worth watching is the direction: agentic checkout is being built by companies with the payment relationships and scale to absorb the complexity, and it is being built around the parts of the visitor economy that already had a seat at that table. The discovery layer stays open to everyone, and for now, it is where the actual weight sits. That is the part worth getting right this year, whatever happens to checkout next.

Read more

What does AI find when it reads your website? Free scan at glassmark.ai, no developer required.

Run your free scan