Act 1. I have seen this movie before
I was working in hotels when Booking and Expedia arrived, so I did not read about this shift, I lived it. At first it looked like a pure threat. An intermediary, the online travel agency (the OTA), was inserting itself between the property and the guest, and the commission it wanted, fifteen to twenty-five percent, felt like a tax on our own rooms, with the direct relationship and the pricing power that came with it slipping to the platform. Plenty of operators reacted to the threat and nothing else, kept it at arm’s length, and lost either way.
But it was never only a threat. For the properties without a big brand’s marketing budget, those same platforms were suddenly the widest distribution any independent had ever had, a shelf we could never have built ourselves. The threat and the opportunity were the same event, and which one it became for you depended entirely on whether you adapted. The properties that moved early, that learned how the channel actually chose what to show and made themselves fit it, got the better outcome. The ones that treated it as optional got the tax without the reach.
That is the whole reason I trust the pattern I am about to describe, because it is the same one. Whoever owns the layer between the guest and the property owns the margin, and a new intermediary is now taking that layer. The difference is the stakes. Last time the platform still handed the choice back to a human who read the list and clicked. This time the intermediary reads the list, makes the choice, and completes the booking itself. The threat is larger, and so is the opening for whoever adapts first.
Act 2. What we measured: the named page is not the ranked page, and the same names keep repeating
Before the forecast, the evidence. Two of these findings are our own research, and both are ChatGPT-centred, which matters for how far the conclusions travel.
Ranking first no longer means getting named. Across five anonymised hotel clients, our Search Console data showed that pages sitting in positions one to three captured only 0.35 percent of AI impressions, while the majority landed on pages ranking four to ten. The pages that got surfaced were the ones tied to what each property is genuinely known for, not the highest-ranking commercial pages the marketing team had chosen to push. One client, an airport hotel, ranked well on its meeting-room pages and lost its AI visibility entirely to its access and layover pages, which broke the commercial positioning the team had built on purpose. The full study is The page that ranks first is not the page the AI names.
The mechanism is query fan-out. When an AI answers a question it quietly breaks that question into a set of smaller ones and retrieves against each, so the page that answers the sub-question set broadly is the one that gets named, not the page ranking first for the head term. In our study of 2,867 ChatGPT fan-out sequences, to be published, the first query behaved like classic SEO, with a steep position-sensitive drop from 40.2 percent of citations at position one to 12.7 percent at position ten. By the middle of the fan-out that gradient collapsed, from 13.9 percent at position one to 4.7 percent at position ten, a 67 percent fall in how much Google position mattered. Domain overlap with Google fell from 34.4 percent on the first query to 17.6 percent by the middle. Ranking first on the fifth query, at a 24.3 percent citation rate, actually underperformed ranking third on the first query, at 26.0 percent. You cannot use Google rank as a proxy for whether an AI will name you.
And the answers concentrate. This is the part that decides the fate of the long tail. As the number of slots in an answer shrinks, the same established names fill them. When ChatGPT began naming fewer brands per answer, it did not distribute the fewer slots evenly, it handed them to the entities the system already recognised and trusted. In our runs, citations concentrate on a small set of validation sources, and visibility compounds, because a brand cited often becomes a familiar entity that gets surfaced again. The rich get richer, and the small hotel, the independent rental, and the single listing get squeezed out of a shelf that keeps getting shorter.
Put those together and you have the world as it stands: a generic-query world, where “best hotels in Zurich” returns the same handful of names, and the long tail loses not because its rooms are worse but because it is invisible to the machinery doing the naming.
Act 3. The shift: the agent reads the shelf, and it reads it for one person
What the Agents API actually changed
On 10 September 2026, OpenAI released the Agents API in public beta. It runs on the same harness that powers Codex, managed by OpenAI, and it lets a developer build cloud agents that run for hours, execute code, work with files, coordinate sub-agents, and use tools including web search. Billing is by token with no extra fee. It supports MCP, the protocol that lets an agent connect to an outside system, so an agent can reach a booking system a business chooses to expose. In plain terms, the thing that used to answer a question can now run a task from end to end, and it can transact against a system you own if you offer one.
This is the moment the intermediary stops handing the choice back to a human. The guest delegates the trip, and an agent searches, compares, checks real availability, and books, without a person reading a single line along the way.
How an agent actually picks, and why keyword thinking is finished
The old mental model was string matching, where a word in the query met a word on the page and rank sorted the winners. That is not what happens. The query is turned into meaning before anything is retrieved, so what gets matched is intent, entity, and place, not words. Four things follow, and they are why long-tail keyword optimisation is the wrong job now.
Intent, because the agent decomposes one request into a set of sub-intents, your fan-out, and retrieves against each, which is exactly why the ranked page and the named page come apart.
Entity, because a hotel surfaces when the system already associates it with the thing the traveller asked about, not because the matching word appears in a title tag. You are competing to be a recognised entity, not to hold a keyword (the plain version of entity salience).
Local, because a place in the query resolves to a place entity and its known associations, so “near the old town” pulls what the model connects to the old town, not the literal string on your page.
Language, because the query language changes the pool of sources retrieved. In our own runs, the same city queried in English and in the local language surfaces different sources, and that divergence is a finding, not noise.
Now add the two things the Agents API puts on top: action, because the agent does not stop at a recommendation, it completes the booking, and fit, because it is doing this for one identified person whose preferences, history, budget, and constraints it already holds.
The forecast, and it is a forecast
Here is where the evidence ends and the reasoning begins, and I want to be honest about the line. Our data is ChatGPT-centred and it measures naming, not booking. What follows is built on the mechanics above, not measured across agents, and I will say so plainly rather than dress a prediction as a study.
The agent will not ask for the best hotel. It will ask what is right for this specific person, and it will know that person better than any search box ever did. The generic list, the one the big names own through sheer repetition, stops being the only battlefield, because the agent is not running the generic query on the traveller’s behalf, it is running a personalised one. Multiply a single personalised match by the millions of people looking for a hotel, an apartment, or a room every day, and the property that wins is the one that understands exactly who it is for and has made that fit legible to a machine.
THE GENERIC QUESTION
“best hotels in Zurich”
- Big chain 1
- Big chain 2
- Big chain 3
The same big names, every time. The small property is invisible here.
THE PERSONAL QUESTION
“right for this traveller”
- Big chain 1
- Your property quiet, family rooms, near the airport
- Big chain 2
Knowing exactly who you are for is the new advantage.
That is the escape route the concentration finding seemed to close. Concentration is what the generic query produces. Personalisation is how a small property that genuinely fits a specific traveller can be the right answer for that person even when it would never rank on the head term. Deep understanding of your real audience, the guests who actually come for you, stops being a marketing nicety and becomes the ranking signal.
The catch is that understanding your guest in your own head is worthless to an agent. The fit has to be encoded, structured, and verifiable, or the agent cannot use it and moves on to the property that made itself readable.
Act 4. How to adapt: stop selling to humans, start being legible to the agent that serves them
The rule is the same for everyone, and it is one sentence. Stop optimising for a human skimming a page, and start being an entity a machine can verify, match to a person, and act on. What changes is how each type of operator applies it.
What an agent needs before it can book you.
- Recognised Known as a real place the AI already trusts, not just a website.
- Readable Real availability, price and rules as clean data, not buried in prose.
- Bookable A connection the agent can book against directly.
Hotels. Being named is now the floor. Exist as a clean, consistent entity across the validation sources these systems trust, so you pass the recognition threshold at all (the audit is the checklist for that). Then go past recognition to fit: make explicit and structured who you are genuinely right for, the solo business traveller with an early flight, the family that wants connecting rooms, the couple that came for the quiet. That specificity is what lets a personalising agent match you to its person. Stop pushing the generic commercial page and let the pages that describe what you are actually known for do the surfacing, because our data says those are the ones that get named.
Property managers and listings. Your job is machine-readability at scale. Real availability, real price, real policies, structured so an agent can parse them without guessing, across every unit. An agent that cannot read your availability does not wait, it books the listing it can read. The manager who structures the portfolio first wins the default across all of it.
Rentals and serviced apartments. This is the clearest personalisation play, because your inventory is already differentiated by exactly the attributes an agent matches on, location, layout, length of stay, who it suits. Encode those attributes as data rather than prose and you become the precise answer for a specific person, which is the one place the head-query concentration does not reach you.
Real estate and capital. AI visibility becomes a demand and value signal for the asset. A property that agents can find, match, and book has a different forward booking profile from one they cannot see, and that gap belongs in due diligence, acquisition, and repositioning decisions. Invisible to agents is a discount you can measure.
There is also an offensive move hiding in the Agents API, and it is the long tail’s first real counter to the OTA era. Because the agent can transact against a system a business exposes through MCP, a property or manager that offers its own booking as an MCP endpoint can be booked directly by the agent, without the aggregator in the middle. The last time an intermediary took the shelf, there was no way around it (what Booking.com does and does not do for your AI visibility is the current version of that trap). This time there is a way around, for whoever builds it before the incumbents do.
See also: The page that ranks first is not the page the AI names | How ChatGPT recommends hotels | What is entity salience | Does Booking.com help AI visibility
Questions this study answers
Will AI agents book hotels for you?
They can now. Since OpenAI’s Agents API went into public beta in September 2026, a developer can build an agent that searches, compares, checks availability and completes a booking without a person reading the list. How much of travel moves to agents, and how fast, is a forecast rather than something we have measured.
Do AI agents use Google rank to pick hotels?
Rank is a weak guide. In our study of 2,867 ChatGPT fan-out sequences, position mattered on the first query and then collapsed: by the middle of the sequence a page ranking first was cited 13.9 percent of the time against 4.7 percent for a page ranking tenth, and the overlap with Google’s results halved. Rank is a weak proxy for whether an AI names you.
Can an AI agent book my hotel directly?
It can if you expose something it can book against. The Agents API supports MCP, the protocol that lets an agent connect to an outside system, so a property or manager that offers its booking as an MCP endpoint can be booked by the agent without an aggregator in the middle. That is the long tail’s first real way around the OTA.