Off-site mentions and AI hotel recommendations
The sentence an assistant writes about your property is assembled from pages you do not control. Knowing which ones, and what they get wrong, is cheaper than any campaign to create new ones.
AI sessions measured across the sites we track, last 30 days.
Quick answer
Off-site mentions are what other people's pages say about your property, and they are most of what an AI assistant reads before recommending anything. In the corrected measurement window, 29 of the 57 questions we track, all about one cenote park in Punta Cana, produced answers built from 784 citations across 271 different domains. The site itself supplied 73 of those citations. More than 90 per cent of the material came from pages it does not own and, in most cases, has never seen.
Who the models were actually reading
The 271 domains in that window were not a level field. Ten booking and ticketing brands took 144 citations between them: Viator 41, GetYourGuide 38, TripAdvisor 36, Expedia 11, Atrapalo 5, Civitatis 5, Kayak 3, Trip.com 3, Airbnb 1 and Musement 1.
Put next to the 73 the site itself supplied, that is the shape of the problem. The pages the models leaned on hardest were commercial listings written by intermediaries, followed by a long tail of destination guides, local operators, blogs and press: 176 of the 271 domains were cited exactly once. Every one of those pages is a sentence about the place that the place did not write.
This is also why the mention count and the link count move separately. A site that publishes a perfect website has fixed the 9 per cent it controls directly.
The cheapest win is a correction, not a new page
When a third party page carries a wrong price, a closed season that reopened, or an amenity that no longer exists, the model repeats it with full confidence, and it repeats it in every answer built from that page. A domain cited 41 times is not one error, it is one error delivered 41 times.
So the first pass through off-site work is not outreach for new coverage. It is reading the source lists of the answers you already appear in, finding the pages that describe you wrongly, and asking for the correction. Most editors fix a factual error, and unlike a link request there is nothing to negotiate.
The second pass is the absence case: the questions where the assistant answered confidently and your name was nowhere. There the target list is not a generic outreach list, it is the specific domains that assistant already trusts for that question.
How to work from measured sources instead of a guess
All of this depends on knowing which pages were cited, which is not something you can infer from the answer text. reporte.ai stores the full response and its complete source list for every measurement, so a pass gives you the domains ranked by how often the models leaned on them for your questions.
Two caveats before you carry the ratio anywhere. The 7 domains measured here are our own travel websites and not hotel clients, so the 90 per cent is how the reading behaves rather than a number about hotels. And it is one pass: a single reading proves nothing, and the list of domains is worth rerunning before you build a quarter of work on it.
What it does turn off-site work into is a finite list. In the window described here it was 271 domains, of which a handful carried most of the weight. Working the top of that list is a different activity from generic outreach, and it is the one that changes what the answer says. The other half of the job, the pages you do control, is covered in content that AI cites.
FAQ
Because they are most of what the model reads. In the corrected window, the answers about one measured site were built from 784 citations across 271 domains, and the site itself accounted for 73 of them. The other 711 came from somebody else.
In that window, ten booking and ticketing brands took 144 citations between them, led by Viator with 41, GetYourGuide with 38 across its four country domains and TripAdvisor with 36 across its six. The rest of the 271 domains were destination guides, blogs, local operators and press, and 176 of them were cited exactly once.
You are probably already on them, which is the point. The gap is usually not presence, it is that the page describing you was written by somebody who visited once, so it repeats facts that are out of date and the model repeats them too.
Often yes, and it is the cheapest fix available, because a wrong price or a wrong opening time on a domain cited 41 times is repeated in every answer built from it. The work is finding which pages are actually cited, which is what the source list of each measured answer gives you.
They are read like any other page. What matters for a citation is whether the page states checkable facts in text, not whether it carries a rating, so a thin five star review contributes less than a detailed description.
No. Nothing here depends on the link passing authority. A page that names you correctly and is never linked to by anyone still feeds the model, and a bought link on a page that says nothing about you feeds nothing.
By reading the source list of the answers where you lost. Prompt tracking stores the full response and its sources for every measurement, so the target list is the domains the assistant already trusts for your question, and in that window there were 271 of them to choose from.
Treat it as the shape and not as your number. The 7 domains measured here are our own travel websites rather than hotel clients, and it is a single pass, so a single reading proves nothing. What transfers is that most of what a model reads about you is written by other people.