Reviews, not your website, decide whether AI recommends your hotel

Guests now open with a sentence, not a search box

A few years ago a booking started with a Google search and ended on a booking portal. Now it often starts with a sentence typed into an assistant: "family-friendly hotel in the Allgäu with an indoor pool that isn't packed." What comes back is a shortlist of a few names, and there is no second page to scroll. SparkToro and Similarweb measured US Google searches between January and April 2026 and found 68% ended with no click to any website, the fastest acceleration they have recorded. If you are not in the answer, you are not in the consideration set.

And here is the catch: an AI recommendation reads to a guest like a tip from a friend who did the research, not like an ad, which is exactly why it lands. A property that ChatGPT names in an answer carries weight no banner can buy. It earns that mention through what the open web already says about it, not through your media budget.

So the decision about your hotel is being made somewhere other than your website. You can polish every page and still lose, because the material the model works from sits on other people's domains. Google says as much in its own AI guidance: its generative features are rooted in the same ranking and quality systems as normal Search, and it explicitly warns against writing separate content for AI. There is no side door. GEO, short for generative engine optimization, starts exactly here: with what the open web already says about your hotel, and whether a model can retrieve it. For a hotel, most of that material is reviews.

Why what others write outranks what you write

A language model has no idea whether your breakfast is good. It has text. When guests keep writing about bread from the in-house bakery and staff who remember their name, those phrases accumulate into a pattern a model can match: this place means warmth and a serious breakfast. Ask it for a "welcoming hotel with a great breakfast" and you are in the candidate set. Nothing mystical about it. It is retrieval over other people's wording.

This is why other people's text beats your own copy. Every hotel calls itself warm and centrally located, so the claim carries nothing. When hundreds of unrelated guests independently say the same thing, it becomes a signal that is expensive to fake. Ahrefs analysed roughly 75,000 brands and found that how often a brand is mentioned across the web correlates with its AI citation rate at about 0.664, against roughly 0.218 for backlinks. Third-party mentions are the best-evidenced lever anyone has measured so far, and for a hotel that means reviews.

The star rating is the least useful part of all this. A 4.6 tells a model nothing about why. The free text is where the value sits, because it hands over the specific attributes a specific question needs. "Quiet room despite the city-centre location" or "step-free and still good-looking" are the sentences that win narrow queries, and narrow queries are what people type into an assistant.

Volume, recency, depth: which levers actually hold up

Volume builds confidence. A profile with two dozen reviews reads like an anecdote; several hundred reads like an established fact. You do not need the highest score in town, you need enough independent voices that the picture is stable. In practice that is a process rather than a hope: a QR code at checkout, an email two days after departure, a direct ask at reception while the guest is already saying nice things.

Recency matters, though not the way it is usually sold. Nobody has shown that answer engines prefer fresh pages; Ahrefs, looking at 1.4 million real ChatGPT prompts, found the pages it cited had a median age of around 500 days. What recency changes is the content of the record. A wall of praise from 2019 sitting above a recent run of complaints about a stalled refurbishment describes two different hotels, and the recent one is what gets summarised.

Depth beats stars. Ten reviews that say what your spa, your location and your restaurant actually deliver are worth more than a hundred rounds of "all great, would return." You can steer for that. Ask one concrete question in the request, such as what the guest used most in the room, and you get attribute-rich text instead of a shrug. Those attributes are the parts a recommendation gets assembled from.

The blind spot: your reviews live on more than one portal

Plenty of hoteliers keep the Google profile tidy, keep the Booking.com profile tidy, and call the job done. A model draws from a far wider field: Google, Booking.com, Tripadvisor, Expedia, HolidayCheck, travel blogs, Reddit threads, regional tourism sites. Ahrefs found only about 6 to 8% of the URLs ChatGPT cites also sit in Google's top ten for the same query, and roughly 80% do not rank in Google's top 100 at all. The sources these systems reach for are not the ones you have been optimising.

So the work is coherence, not parity. You do not need the same review count everywhere. You need every source to describe the same hotel. A property known across all of them for the same two or three things, say family-friendly and a quiet street, gives a model an easy, low-risk answer. Contradiction is what costs you, because when the evidence conflicts the model reaches for the hotel it can describe with confidence.

So audit the channels you never look at. That is where you find photos from two refurbishments ago, the wrong restaurant hours, a complaint about something you fixed two seasons back. And these systems repeat what they find without hedging: Columbia's Tow Center ran 1,600 queries across eight AI search tools and got the source wrong in more than 60% of responses. ChatGPT misidentified 134 of 200 articles while signalling any uncertainty only 15 times. Stale detail about your hotel gets restated as fact.

Critical reviews help you, if you answer them

A wall of five stars with no dissent anywhere in it reads as bought, to people and to models. A normal spread with some critical voices reads as real. The variable that matters is not whether criticism exists but whether you answer it, and a specific, unembarrassed reply adds text to the record showing that this hotel reads its mail and does something about it.

Your replies are part of the record too. A guest complains about thin walls; you answer that those rooms were soundproofed over the winter. Both facts are now public, the problem and the fix, and a model summarising your hotel has the second one available. Stay silent and only the complaint survives to be quoted.

Answer every review with anything substantive in it, the good ones included, and answer in your guests' vocabulary. If a compliment gets a reply that repeats the concrete thing, "glad the breakfast terrace and the Alpine view worked for you," you have just doubled the number of places that phrase appears next to your name.

Two Alpine hotels, same rating, different outcome

Take two four-star hotels in the same Alpine village, both sitting at 4.5. Hotel A has around 180 reviews, many from the last few months, full of specifics: the sauna, the ski-bus stop, the fact that dogs are genuinely welcome. Hotel B has more reviews, but most are two years old and most say some version of "lovely, thank you." Side by side on a comparison page, the two look identical.

Now someone asks for a "ski hotel with a sauna you can reach with a dog and without a car." Hotel A wins that answer, and not because it is rated higher. It wins because its reviews contain all four attributes from the question, in recent plain language a retrieval system can match. Hotel B may well be the better hotel. It just never got written down.

That is the whole mechanic. No tricks, and no secret file either: Google's guidance says its generative features need no AI-specific markup, and John Mueller confirmed in 2025 that no Google Search system reads llms.txt. The job is making your real strengths legible enough, in public, that a machine can match them to a real question. Do that steadily and you win bookings your competitor never sees being lost.

Build a review habit that runs without you

Start with the ask, because everything downstream depends on volume you can rely on. Pick one moment and one channel, write the wording down, and train the team on it so it stops depending on who is on shift. The best moment costs nothing: a guest praising the stay at checkout is a guest who will say the same thing online if you ask them right there.

Then steer the vocabulary. If what sets you apart is how you handle sustainability, talk about it in the building, explain what you actually changed, and guests will use those words in their reviews. Nothing invented, just the true thing said out loud often enough to reach the record. Getting your structured data right serves the same end: schema has not been a ranking factor since Google said so in 2018, but it makes prices, hours and location machine-readable, which is how you stop a model from guessing them.

Reputation is the distribution channel now

The old arithmetic was best website plus biggest budget equals visibility. The new one is that the richest, most current and most consistent public record gets recommended. That is unusually good news for a hotel, because this is the one currency you cannot buy; it is earned by hospitality people choose to write down. And the audience is not niche: ChatGPT reported 900 million weekly users in February 2026, and Google's AI Overviews passed two billion users a month.

Start narrow, but start this week: list every platform your hotel appears on, fix what is out of date, put one review request into the checkout routine, and answer everything that comes back. All four feed the same picture, the one ChatGPT, Gemini and Perplexity assemble when the next guest asks where to stay. That picture, not your homepage, decides whether your name is in the answer.