AI-native distribution is the one thing Google won't build

Ask Gemini to find a hotel in Zermatt and it does the job well. It reads the request, compares options, shows rates, hands over a link. Then it steps aside. The booking happens somewhere else — on Google Hotels, on the property's own page — after the conversation has passed you along.

That is the hotel answer most of the industry expected Google to ship. It looks like search because it is search. Google built the tool it knows how to build: a place to discover, compare, and click out to a bid. To an advertising business, a hotel is an ad. The surprise isn't that Google did this. The surprise is what it tells you about the category Google left for someone else.

The booking breaks where the link begins

Every AI hotel path on the market today ends the same way. The traveler gets a recommendation, a price, a link — then leaves the conversation to finish the stay. Gemini passes you to Google Hotels. The new direct-booking aggregators pass you to the property's booking engine. Different doors, same threshold: the transaction completes outside the assistant.

Watch where each one strains. Gemini's hotel results run on Google's own Hotels and Hotel Ads stack, so it can answer a specific date pair with a live rate, then send you out. The newer aggregators have the opposite trouble — they often can't answer the date pair cleanly, and they send you out anyway. The shared weakness is the last step. The assistant recommends. It does not close.

Push gets you found, pull gets you booked

This is the line the whole category turns on, and it is older than AI. Hotel connectivity has always run in two directions. You can push your rates and availability out to a channel as a feed, refreshed on a schedule. Or a system can pull from you — ask your reservation system, live, for these exact nights, right now. The split is written into the industry's own standard: OpenTravel has separate messages for pulling a reservation and for pushing one out.

Most of what gets sold as agentic distribution is push. Take DirectBooker, the most visible of the new entrants, founded by people who built Google's Flights and Hotels products. It gathers direct rates from hundreds of hotel systems, normalizes them into one feed, and presents that feed to the assistant. By its own account it blends caching with real-time updates. Caching is the tell. A cache answers from what it stored, not from your reservation system at the instant of the question. That holds up for the vague top of the trip — somewhere in the Alps in March. It frays the moment a guest names two dates, because a per-night rate is often computed rather than looked up, and last-room availability turns over by the hour.

Push earns you a place in the recommendation. The booking is a separate problem.

Pull is the pattern the GDS already runs

Here is the part the agentic story keeps stepping around. The architecture for booking on a traveler's behalf already exists, and it has run for forty years. It's the GDS.

When a travel agent searches Amadeus for a hotel in Zermatt, the system holds no inventory of its own. It queries the connected reservation systems live, keeps the intent in a record while the agent builds out the rest of the trip, and at the moment of booking goes back to the source to confirm the room is still there at that rate — the whole round-trip, in the normal case, inside a few seconds. That is pull: live query, held record, re-validation at commit.

And the GDS solved the part the feed players skip — the money. There is a settlement layer underneath, with defined commission and a clear answer to who pays whom. The industry even standardized it: HEDNA, HTNG and OpenTravel built a shared payments model for exactly this, and the two largest systems run their own payment engines beneath the search. The connection carries a standard, too. A reservation system that answers too slowly is read not as "please wait" but as "no availability" — the request times out, and the hotel falls out of the results. Speed is a condition of being sold.

Structured data is the easy half, and the GDS has set that standard for decades. The commerce layer is the hard half. AI-native distribution needs both, running live. The GDS is proof the whole thing is buildable, because it is already built.

The agent already left the building

If software shopping a live source on a traveler's behalf sounds like the future, look at the corporate channel. The human agent has mostly been replaced — by the traveler and an API. Self-booking tools now run between roughly half and three-quarters of managed travel in mature markets, and most large North American programs require one. Analysts call the move from GDS-only content to a mixed, API-first model the biggest change in corporate distribution since the GDS itself.

So the precedent for agentic booking already exists. A person states intent, software queries live inventory, the booking completes against a real reservation system. That has been the corporate norm for years. The consumer assistant is walking the same road, a few steps behind.

Why the company everyone's watching won't build it

Which brings it back to Google. It has the hotel rails — Hotel Ads pulls live rates, the knowledge graph holds the content, the reach is unmatched. It has, arguably, more of the pieces than anyone. And it is pointing them at an auction. Hotel Ads is the auction, and it already powers Gemini's hotel answers. Asking an advertising business to build neutral booking infrastructure is asking it to compete with its own model. If your tool is a hammer, the work in front of you looks like a nail.

That is the opening. Real AI-native distribution — pull, held intent, live re-validation, settlement underneath — sits in a spot the largest player has every reason to walk past. The blueprint is public, too. It has been running in the GDS the entire time, waiting for someone to rebuild it for the assistant. Whoever does owns the moment the booking actually closes.