AI Didn't Kill the Middleman. It Promoted Him.
Ask ten operators what they think AI agents will do to their business, and eight of them ask the same question back before you can answer. Does this cut out the middleman?
No. It hired one. Gave him a better suit and a faster car, and kept him on the same commission.
That's not a hunch. In late August, Google rolled its AI chat experience into hotel booking for U.S. travelers, letting people find and book a room without leaving the conversation.[6] The launch partners tell you everything you need to know about where this is headed: Booking.com, Expedia, Hotels.com, Priceline, Trip.com, and the big brand chains, Hilton, Marriott, IHG, Wyndham, Choice.[6] Not one small property on that list. Not one guide, not one boutique inn, not one host who answers their own phone.
That's not an oversight. It's the architecture.
Does AI actually cut out the OTA, or just feed it?
A study of AI hotel recommendations, run across 3,600 prompts, found Booking.com showing up in 95.3 percent of hotel queries.[11] Across the board, OTAs pulled in 22.9 percent of all the URLs an AI assistant cited when answering a hotel question. Independent hotel websites got 11.8 percent.[11] Less than half.
So when a guest asks an assistant to "find me somewhere nice near the coast, book it," the assistant isn't touching your site. It's reading a structured feed from an OTA that spent years building exactly the kind of machine-readable data these tools eat for breakfast. Then it books through that OTA, and that OTA takes its usual cut, somewhere between 15 and 25 percent, same as it always has.[11]
One analyst summed it up in a way that stuck with me: AI has added a discovery layer on top of the funnel. It has not removed the commission layer underneath.[10] The assistant is new. The toll booth is the same one that's been there for fifteen years. It just got a faster road leading to it.
This matters because the instinct, when a new technology shows up, is to assume it disrupts the incumbents. Sometimes it does. Here, so far, it's doing the opposite. The OTAs had the cleanest data, the deepest inventory feeds, the APIs already built. When the machines went looking for something to read, the OTAs were the only ones who'd already written the book in a language the machines could parse. Independent operators, most of them running a website built by a nephew in 2019 and a PDF for a menu, weren't in the conversation. Not out of malice. Out of readiness.
The numbers behind why this stings
Here's what's actually at stake, in dollars, not vibes.
Cloudbeds tracked ninety million bookings across independent hotels in 180 countries and found OTA share climbing to 63.4 percent in 2025, up from 61.3 percent the year before.[1][10] In some markets it's pushing toward 80 percent.[1] That's before AI agents were doing any meaningful share of the routing. Early estimates now put AI travel agents at 3 to 5 percent of hotel bookings in major markets, with projections running to 10 to 15 percent by 2028.[9] Small number today. Not a small number in three years, if the current bias holds.
And the bias has a cost you can put a figure on. A direct booking averages $516. The same booking through an OTA averages $312.[1] Direct guests cancel at 10.6 percent. OTA guests cancel at 21.8 percent, twice the churn, twice the empty rooms on a Tuesday you can't sell twice.[1][10] Acquisition cost for a direct guest runs 2 to 5 percent. For an OTA guest, 15 to 30 percent or more.[1][5] One estimate suggests that shifting just ten percent of OTA volume back to direct adds roughly $64,000 a year to net operating income for a typical property.[5] That's a renovated bathroom. That's a full-time hire. That's not nothing.
So every fraction of a percent that an AI assistant routes toward an OTA instead of toward you isn't an abstraction. It's a guest who would have paid you $516 and shown up, paying an OTA $312 worth of margin instead, with a one-in-five chance they don't show up at all.
To be fair to the uncertainty here: nobody has clean public data yet on exactly what share of AI-originated bookings land direct versus through an OTA.[2] The researchers who'd know are honest that the empirical gap is real. What we have instead is circumstantial, but it's consistent circumstantial evidence, stacked five sources deep, all pointing the same direction. When the assistant cites a URL, it's usually Booking.com. When it books, it's usually through the door Booking.com built. You don't need a perfect dataset to smell a pattern. You just need to have worked a dining room long enough to know when the house is quietly taking a cut nobody mentioned on the menu.
What operators are already running into
One host running their properties through Guesty told the site plainly: they wished their system connected to more agentic channels. Not because they wanted another OTA relationship. Because they'd already figured out, on their own, faster than most of the trade press, that the assistants are the new front door, and their software wasn't built to answer that door yet.
That's the honest state of the plumbing right now. The PMS companies are racing to catch up. The OTAs already have the pipes laid. The independent operator is standing in the yard wondering why the water only runs to the big house next door.
Trade advisors are framing this year, and the next two, as the window to fix it. The pitch is that a hotel starting now can plausibly cut its OTA dependency from 65 to 70 percent down to 35 to 40 percent within eighteen months, by rebuilding its data, its content, and its booking flow so an assistant can actually read it.[13][14] I don't love the urgency language these pieces use, it reads like a timeshare pitch half the time, but the underlying math is sound. The assistants default to whoever's easiest to parse. Right now that's the OTA. That's a fixable condition, not a law of physics.
What actually changes the routing
The fix isn't complicated, even if the execution takes real work. An assistant needs a plain, structured, current answer to three questions: what do you offer, what does it cost, and can it book it right now without a phone call. If your rates live in a PDF, if your availability lives in someone's head, if your cancellation policy is a paragraph buried in a confirmation email nobody reads, the assistant skips you. Not out of judgment. Out of budget. It has a task and a deadline, and your site cost it too many seconds to parse.
That's the whole game right now. Not marketing. Not brand voice. Legibility. The operators who get booked by AI in the next two years won't be the ones with the prettiest photography. They'll be the ones whose information a machine can lift cleanly off the page in under a second, same as a line cook needs the ticket to say "no dairy" in plain letters, not buried in a paragraph about the guest's dietary journey.
Direct bookings were always worth protecting, long before any of this. They pay better, they cancel less, they let you actually know who's walking through your door on a Friday night instead of finding out from an OTA's dashboard three days out. AI assistant bookings didn't create that math. They just raised the stakes on whether you show up in the conversation at all.
If you want to know where you currently stand in that conversation, whether an assistant can even find you before it defaults to the usual names, bookaible has a way to check how visible you are to agents right now, before the 2028 numbers stop being a projection and start being your P&L.
The middleman isn't dead. He's just running a faster operation now, with better handwriting. The only real choice is whether you're still one of the ones he has to route around.