Why it is hard to tell when a hotel has been left out of an AI recommendation
Ask ChatGPT to recommend a hotel in a given city for a particular kind of stay and it returns three or four names with roughly a sentence each. Those sentences are assembled from whatever the model can find and corroborate, which is seldom the material an operator would have chosen. The properties left off the list never learn they were in contention.
That is the whole of the subject, and it is worth stating plainly, because most owners approach it sceptically after watching several acronyms arrive with a retainer attached. What follows is what absence costs a property, and what closing the gap involves.
A lost RFP is visible. A rate that loses to the comp set shows up in the shopper report. Discovery losses behave differently. No enquiry arrives, no booking is abandoned, nothing moves in the P&L. The demand never appears, and no report records that it was ever there.
That is why the issue sits unaddressed at most properties. Nothing surfaces to prompt the conversation. The first signal tends to be indirect: a competitor visibly winning a segment or a feeder market that a hotel had assumed was its own, with no obvious change in rate, product or sales effort.
The volume behind it is no longer marginal. Around 40 percent of travellers now use AI somewhere in their planning, rising to roughly 60 percent among younger leisure travellers, and most of those who have used it say they went on to book off the back of an AI recommendation.
.png)
What makes a hotel description usable to an AI, and what makes it invisible
AI models make no judgement about quality. They work from what they can cite, and most hotel copy gives them nothing to cite.
Consider a couple looking for a quiet weekend somewhere, with dinner good enough to stay in for. They ask an AI rather than scrolling listings, and they ask in roughly those words.
Somewhere in that market sits the hotel they are describing. Forty rooms, no coach groups, a restaurant the head chef has built a genuine local following for. Its website offers a warm welcome, comfortable rooms and a range of dining options. Every word true, none of it usable. Nothing there establishes whether the property is quiet or busy, forty rooms or three hundred, or whether dinner is a reason to come or simply a convenience.
So the recommendation goes to the hotel whose page states forty rooms, no groups or events, and a restaurant open to non-residents with a menu that changes weekly. That property may well be the weaker of the two. It is the one a model can describe with any confidence, and shortlists are built from confidence.
This is what catches good operators out. The strengths are real, the guests who find the place agree, and the description still lets it down, because it was written to sound appealing rather than to be specific.
Specificity is only half of the requirement. The other half is corroboration. A fact stated once on a hotel's own website carries limited weight. The same fact appearing on the site, in the guides, on the Google profile and in press coverage carries considerably more. Single-source claims get hedged. Corroborated ones get repeated.
Who actually writes the description an AI gives of a hotel
One study attributes around 71 percent of AI hotel recommendations to third-party material, guest reviews, guides and articles among them, and only about a quarter to the property's own website.
Read as an allocation of authorship, that is a striking split. Three quarters of how a model characterises a hotel comes from sources the hotel did not write, most of them optimised for something other than its positioning. An OTA amenity field exists to make properties comparable, which means it flattens whatever makes any one of them distinct. A review from 2022 describes a restaurant concept that has since been replaced.
None of this is solved by editing a homepage. It is solved by making the handful of facts that genuinely sell a property appear, in the same form, everywhere a model looks.
What an AI check reveals about competitors that an SEO report does not
There is a benefit here that operators tend to notice only afterwards, and it marks the clearest difference between this and a standard SEO audit.
An SEO report returns positions: keywords, rankings, the domains sitting above. An AI check returns reasoning, because the model answers in sentences rather than in a table. Asked to recommend a hotel for a particular kind of stay in a given market, it names three or four properties and explains what it believes each one is for.
That is a more useful read on a competitive set than a ranking. It shows which attributes the model has attached to each property, which kinds of stay they own, and where one entry is thinner than the others.
Some of what surfaces has little to do with marketing. Where a model consistently credits two nearby hotels with something a third does not appear to offer, there are two possibilities, and they lead to different meetings. Either the feature exists and nobody ever stated it in a form a model could find, which is a copy fix costing an afternoon. Or the competitors have built something that property has not, which belongs in a capex discussion rather than a content one. This is intelligence that ordinarily requires mystery shopping or a commissioned market study, and here it falls out of the same check.
Why hotels that act early hold on to the advantage
The descriptions forming now become the baseline. These systems reinforce what they already have evidence for. A property described well today accumulates corroboration and continues to be recommended. A property described from OTA fields continues to be described from OTA fields, and every month of that makes the record harder to shift.
Recovered demand carries no commission. Most travellers still book through familiar channels, but they increasingly discover through AI, and a discovery that leads a guest to approach a hotel directly involves no intermediary fee for the introduction. Across a season, on bookings a property would otherwise have paid to be introduced to, the difference is material.
How to check where a hotel currently stands, in about ten minutes
The gap is straightforward to document before anything is commissioned.
Five queries, covering a property's three highest-value segments, phrased as a guest would phrase them. Not "hotels in Valencia". Closer to "somewhere quiet in Valencia for a long weekend, walkable to the old town, with a restaurant worth staying in for".
Each query run in ChatGPT, Gemini and Perplexity, in a fresh session with memory and personalisation switched off. This matters, otherwise the results reflect what a model has learned about the person asking rather than what a stranger sees.
Four things recorded per query: whether the hotel appears, which attributes are credited to it, which of those are wrong or missing, and which properties appear in its place.
What all of this means in practice
Discovery is moving to a place where guests search by specifics, the answer is short, and anything left unstated gets filled in by someone else. The properties that take control of that record now are the ones these systems will keep recommending, in words their owners actually chose.
I would be interested to hear from other operators. Run those queries against your own hotel, and which competitor comes back instead?
