Agentic Search and Your Direct Channel: Preparing for AI Booking Agents

Google started booking hotels inside AI Mode on August 27, and its ten launch partners were OTAs and big chains. Hilton now lists large language models as a risk to its direct channel in its 10-K. The bookings an agent makes go to whoever it can read, price and complete. This is the layer an independent hotel controls: a clean rate feed into Google, structured data on the website, rates an agent can actually see, parity monitoring, and a booking engine vendor with a dated plan for agent connections.

Hilton added one sentence to its annual report in February, and it's the most useful sentence an independent owner will read about distribution this year.

"Consolidation of internet travel intermediaries, or the entry of major technology platforms, such as large language models, into the internet travel bookings business, also could divert bookings away from our direct channels and increase our hotels' cost of sales," the company wrote in its 10-K, as Skift reported. According to Skift, it was the first time Hilton had formally named AI as a risk to how it sells rooms. Marriott's filing carried a similar warning: AI platforms like ChatGPT could steer guests toward the OTAs and push distribution costs up.

Read that as an owner. The two biggest hotel companies in the world, with the largest loyalty programs and the deepest direct booking engines in the business, told their investors the new front door might not be theirs.

I spent years at Starwood as VP of sales strategy and operations across more than 800 hotels, and a good share of that work was the same argument in an older form: who owns the guest at the moment of search, and what does it cost to get them back. Then I owned ten hotels of my own, where that argument stopped being strategy and became a line on the monthly channel report. The OTA row was always the one that moved. When it moved up, margin moved down, and nobody needed a slide to explain why.

Agentic search is that row with a new name on it. The rest of this piece is about the part a hotel can actually control.

The Problem: The Search Box Is Becoming a Booking Agent

On August 27, Google began rolling out hotel booking inside AI Mode in the U.S. A traveler types what they want in plain language, gets a short list with reviews, picks a property, taps "Continue on Google," chooses a room, reads the cancellation terms and pays with Google Pay without leaving the conversation. TechCrunch listed the launch partners: Booking.com, Choice Hotels, Expedia, Hilton, Hotels.com, IHG, Marriott, Priceline, Trip.com and Wyndham.

No independents. No booking engine vendors.

Google's product lead for the feature told PhocusWire that the booking partner, hotel or OTA, stays merchant of record "for now," and that ranking is organic with "no way for partners to influence the ranking from a commercial standpoint." He also said the rollout used a mix of integration paths, and that "as we scale, we will rely on UCP," the Universal Commerce Protocol Google named hotels as its next vertical for at I/O in May.

Google isn't alone. Expedia and Booking.com have run apps inside ChatGPT since October 2025, built on the Model Context Protocol, per Skift. Meta's Muse agent reached 750,000 hotels by connecting to Expedia once, as Max Starkov pointed out in a HospitalityNet panel last week. That's the pattern to notice. An agent builder wants inventory at scale with one integration, and the OTAs already have it.

Travelers are ready faster than most hotel websites. Phocuswright found that 56 percent of U.S. travelers used AI for planning, booking or in-destination help on at least one trip in the past year, up from 33 percent in the first half of 2025. By July, 44 percent said they would book directly inside an AI platform, and 40 percent would let an assistant handle the hotel booking for them.

Here's the risk in one line. An agent books what it can read, price and complete. If your rates are invisible to it, it doesn't conclude you're sold out and move on politely. It books the hotel down the street, or it books you through someone who takes 18 percent.

An agent books what it can read, price and complete. If your rates are invisible to it, it books the hotel down the street, or it books you through someone who takes 18 percent.

The Data: Where the Agents Are and What They Can See

Start with where an agent can complete a hotel booking today, and who holds the transaction when it does.

That last row is the one most owners haven't tested. On October 18, Terence Ronson ran Claude and ChatGPT agents against the websites of ten large hotel groups and published what happened. Agents "reached the checkout page at five or six of each tool's ten groups." His conclusion is the one to tape to the revenue manager's monitor: "Today, access is often decided by bot-management settings rather than by policy."

One line from that test is worth more than the table. An agent "cannot tell a sold-out hotel from a silent refusal." A firewall rule written in 2023 to stop rate scrapers now tells a guest's assistant that you have no rooms.

The problem underneath is plumbing. In the same HospitalityNet discussion, Alessio Re of Klartika put the independent's position plainly: "An independent hotel cannot expose live rates to an agent until its booking engine or channel manager vendor ships the integration, the chokepoint that stalled Book on Google." He also made the hopeful point. "Hotels running metasearch already send Google Hotel Center the ARI feed that live prices in an assistant depend on." ARI is availability, rates and inventory. Most independents already pay to produce it.

Now the money. I modeled a 100-room independent at 72 percent occupancy and a $200 ADR, about 26,280 room nights a year, and asked what happens as a share of those nights starts arriving through AI channels. The only question that matters is where the agent sends them.

Nobody knows which row a given hotel lands in by 2028. I don't either. But I've read enough owner statements to know that $88,000 a year flows almost straight to GOP at a property this size, and it's decided by technical choices most owners have never been asked to make.

The Mechanism: An Agent-Ready Distribution Layer

An AI agent reads your hotel the way software reads anything: through feeds, markup and pages that load without a human clicking. So the fix is a set of systems, owned by someone, that make the direct channel the easiest thing for an agent to read and book. Six parts.

1. A clean rate feed into Google. Google's free booking links already put a hotel's own booking site next to the OTAs on Search and Maps "at no cost," ranked partly on "historical accuracy of the prices provided to Google." The feed behind them comes from Hotel Center, usually through the booking engine or a connectivity partner. It's the nearest thing an independent has to an on-ramp for AI Mode, because it's the same rates-and-availability data an assistant needs. Check that your official site link is live, that the feed covers every room type, and that the price accuracy report in Hotel Center is green.

2. Structured data on every page that matters. Agents and the models behind them parse schema.org markup far more reliably than they parse a designer's layout. A Hotel block with address, check-in and check-out times, pet rules, amenities and star rating, plus HotelRoom entries for each room type with bed configuration and occupancy. Google documents the markup it reads in its structured data guide. Fees belong in there too. A resort fee an agent can't find is a resort fee the guest discovers at checkout, and the review that follows gets read by the next agent.

3. Rates an agent can actually see. Load your booking engine with a script-free fetch and see what comes back. Many engines render the rate calendar entirely in JavaScript, behind a cookie banner, inside an iframe on another domain. A person gets through. An agent may get an empty page. Then check the bot rules. The Hotel-Online test showed access is often set by bot-management defaults, and Imperva's figures in the same piece put malicious bots at 44.5 percent of online travel traffic, so the answer isn't to open everything. Block scrapers. Allow the verified agents of the major AI platforms through to the rate page, and log every visit by user agent so you can see who's trying.

4. Parity monitoring that watches what the agent sees. Agents compare. Phocuswright found price comparison drives more action on AI recommendations than anything else, at 44 percent. If the OTA shows $189 and your site shows $199 because a wholesale rate leaked, the agent books the $189 and it's right to. A rate shopper running daily against your own site, the main OTAs and Google's displayed price catches the leak before an agent does.

5. One set of facts, published once. Pet policy, parking cost, early check-in, cancellation terms, what's included. Written in one place, pushed to the website markup, the Google Business Profile, the OTA extranets and the guest messaging assistant from the same source. When the website says pets are $50 a night and Expedia says $75, an agent either picks one at random or flags the property as unreliable. Neither helps.

6. A dated agent roadmap from your booking engine vendor. UCP, MCP and Google's agent payments work are vendor-level integrations. A hotel doesn't build them. The vendor does, or it doesn't. Ask in writing: when will you support hotel booking in AI Mode, do you plan a UCP integration, do you expose availability over MCP, and who handles modifications and cancellations made by an agent. Lighthouse's Juanjo Rodriguez said on the HospitalityNet panel that "agentic bookings are still a small share of volume, but growing extremely quickly." Your contract renewal date is probably closer than the vendor's roadmap.

A firewall rule written in 2023 to stop rate scrapers now tells a guest's assistant that you have no rooms.

The Framework: An Agentic Booking Readiness Check

This is the check I'd run before spending a dollar on anything new. Most of it is configuration of systems the hotel already pays for.

Two more items don't fit a table. Ask one AI assistant a week to find and price a two-night stay at your hotel for real dates, and write down where it sends you. And tag AI referrals separately in GA4. Visits from chatgpt.com, perplexity.ai and gemini.google.com show up as referrers, and they deserve their own channel grouping so you can watch the trend before it shows up as an OTA bill.

Our earlier pieces on direct booking economics and metasearch bidding cover the cost side in more depth. This check is about whether the agent can find the door at all.

Implementation: Eight Weeks to an Agent-Readable Direct Channel

None of this needs a new platform. It needs one owner (usually the revenue or e-commerce lead), the booking engine vendor on the phone, and someone who can edit the website.

Two cautions.

Don't open the firewall wide to fix week five. The bot traffic hitting hotel sites is mostly hostile, and the goal is a narrow door for known agents into the rate page, with logging, not an open house.

And don't wait for the vendor to call you. At Collective Retreats, which I founded and ran across ten locations until it was acquired in 2025, the integrations that shipped on time were the ones we asked about in writing with a date attached. The ones we assumed were coming mostly weren't. Independent hotels are a small line on a booking engine vendor's roadmap until enough of them ask the same question.

If you want an outside view of how your rates, content and channel mix look to an AI agent today, our AI revenue optimization and forecasting work starts with exactly this readiness check and the channel cost model above. The channel manager piece covers the distribution stack underneath it.

Hilton put the risk in writing in February. An independent can watch the same traffic arrive, but only if someone tests the door this month.

Frequently Asked Questions