How hotels get into AI recommendations and win more direct bookings

The question we hear most from hoteliers: why does an AI assistant recommend the hotel two streets away and not us? The answer usually has almost nothing to do with artificial intelligence. It is about what the hotel has written on its own website.

Language models do not invent recommendations. They assemble them from what is written about you: on your site, on maps, in aggregators, in reviews. A hotel with precise facts, stated the same way everywhere, makes it into the answer. A hotel promising "maximum comfort" while every directory lists something different does not.

Here is the whole thing. At the end there is a checklist to download: fifteen checks and five ready-made prompts that tell you in an afternoon what ChatGPT, Gemini and Perplexity know about you.

Your next guest may never open a search engine

There are two routes to the same booking today, and both run at the same time.

The old one: search engine → aggregator → comparison → booking. The guest types "hotel Cesky Krumlov parking", opens Booking.com, works through filters, compares, and books. Usually on the aggregator, at 15 to 20 percent commission.

The new one: question → options → follow-up → booking. The guest asks an assistant to "recommend five hotels in the Giant Mountains for a family with a three-year-old, we need breakfast and parking", gets three to five names with reasons, asks about cancellation terms, and goes straight to the site of whichever one answered best.

Notice what did not happen in the second case: any comparison. It happened earlier, without you. And if the model knows nothing specific about your hotel, there is nowhere for you to show up, because you were never among those five names to begin with.

Ten things people ask AI about hotels

These are the query categories hotels get picked on in generative answers. Substitute your own city and your own conditions.

They are nearly the same queries people type into a search engine, so the work you do for AI pays you back in ordinary local SEO too.

Why one hotel gets mentioned and another does not

When we audit a hotel site for AI visibility we ask five things. The answers are usually enough to show where the problem sits.

1. Is the site reachable at all? Surprisingly often we find a site closed off in robots.txt, or noindex on half the pages. The owner wonders why nobody ever finds them.
2. Does the hotel match what the guest is asking for? If you target families, the site has to cover what a family worries about. Not a general paragraph about an unforgettable experience.
3. Are the exact conditions and limits on the site? Check-in and check-out times, surcharges, dog rules, cots, parking. Numbers, not adjectives.
4. Does the data match across platforms? If your site says you have a pool and one directory never mentions it, the model has a conflict. Resolving a conflict costs more than skipping you. So it skips you.
5. Is the site structured usably? Sensible headings, meaningful URLs, sections that map to what a guest is looking for.

Where AI learns about a hotel

From four places. Your own site and its service pages weigh most, because that is the only source you fully control. Then maps and directories, so Google Business Profile, local map services and regional tourism portals. Then aggregators and reviews: Booking.com, Tripadvisor, Google ratings. And finally everything somebody else wrote about you, from articles to partner sites.

The more places say the same thing about you, the more confident the model is. The fewer, the more likely it replaces you with somebody it knows better.

Same prompt, different answer

We asked three assistants the same thing, "recommend five hotels in Prague with the best value for money". Each returned a different list, which we expected. What we did not expect was that each one compared on different criteria. One leaned on aggregators and reviews, one on maps and hotel websites, one mixed them and cited its sources.

So there is no single "first place in AI". There are several models drawing on different sources, and your position is computed differently in each. You have to measure them separately, and improving one channel is not enough.

What to do on the website

This is the part nobody expects when they start reading about neural networks. Most of the work is completely ordinary and can be done this week.

Write numbers, not adjectives. A model will not guess you have parking. "Best service and maximum comfort" carries no information. "25 free parking spaces, covered garage for 8 cars at 250 CZK per night" does. Same for dogs: weight, size, fee, restrictions. For check-in, an actual time. For a cot, the price and the conditions.

Then there is the front desk, quietly doing content research you could not buy anywhere else. Have them log for a week what guests ask by email and phone; whatever turns up on that list is what the site is missing. Put those questions into an FAQ word for word and answer them properly.

And build pages around guest needs, not around your org chart. Staying with a dog, holidays with kids, spa, conferences, parking and transport. A whole page about parking sounds like pedantry. It works better than almost anything else: photos of the entrance, bay dimensions, number of spaces, how to get in and out, the price. You can link to it, and a model can lift an exact answer out of it.

One question, one precise answer. "Can I bring a dog?" Yes, up to 15 kg, 350 CZK per night, not in the restaurant, bowls at reception. Language models quote that format most readily, because nothing has to be inferred from it.

And shorten the path. The era of websites where something rotates as you scroll is over. What wins now is a site where the guest is one or two clicks from the information and where booking works on a phone. Speed is a ranking signal too, and we covered the mechanics in why slow pages lose you customers.

Technical prep: five tasks for a developer

1. Check that the site is open to crawlers. That means robots.txt, noindex and error states. And specifically for AI, do not blanket-block GPTBot, ClaudeBot, PerplexityBot or Google-Extended unless you have a reason.
2. Add an llms.txt file. It is a new and still voluntary standard, a short map of your site for language models. Fifteen minutes of work, and worth having before your competitors do.
3. Get the content into HTML. A price list published as an image is invisible to a model. The same goes for opening hours and house rules.
4. Add structured data: schema.org Hotel and FAQPage, clear headings, clean URLs, a title and a description on every page. The full list is in our technical SEO checklist.
5. Run a full booking on a phone, all the way to the confirmation email. Most hotels never have.

Praising yourself is not enough

"We have the best spa in the region" on your own site is worth nothing unless something backs it up. What backs it up is guest reviews, photos visitors uploaded to maps, independent articles, press mentions.

Answer your reviews, especially the bad ones. Models quote recurring complaints as readily as praise, so if ten people write that parking is impossible, that lands in the answer. Fix it, let new guests confirm the fix, and the picture eventually gets rewritten.

What matters more than the number of mentions is their quality. Ten invented five-star reviews do not outweigh one specific account from someone who actually stayed with their dog.

A mention is not a booking

Getting into the answer is the first step, not the result. The full path: AI answer → interest in the hotel → a click or a branded search → your site → comparing the room and the terms → booking. You can lose it at any of them.

The last step depends entirely on whether you give the guest a reason to book direct. It does not have to be a discount, and because of rate parity it often should not be. What works is something extra for the same money: breakfast, a transfer from the station, a late check-out, free parking. Something cheap for you with obvious value for the guest. Just write down every limit on it, or it turns into a complaint.

The rest is craft. A booking engine on your own site, connected to your system, with current rates and a clear final amount before payment. For a hotelier the difference between a direct booking and an aggregator is 15 to 20 percent of every night. We go into it on our page about websites for hotels and accommodation.

How to measure whether it works

The worst possible measurement is a screenshot of ChatGPT mentioning you once. Answers shift from day to day and a single prompt proves nothing.

Do it like this instead. Keep a fixed set of prompts, ten is plenty, and ask them the same way every time in a fresh chat. Log the date, the service, the hotels mentioned and the sources. Track visits from AI separately in analytics. Citation tracking tools have existed for about a year and they cost money, so for a handful of prompts a spreadsheet still wins. We compared both in our review of AI search tracking tools. And count mentions, visits and confirmed bookings apart, because they are three different things.

Then there is a trap that catches experienced people too. As AI visibility grows, branded search usually grows with it. The guest hears about you from an assistant, then googles you anyway and arrives "directly". Good news for the business, bad news for attribution. A rise in direct traffic on its own does not prove language models caused it. Honestly, this is the weakest part of the whole discipline and nobody has properly solved it.

Try it in ten minutes

Open ChatGPT, Gemini or Perplexity, turn on web search, and run these five prompts. The first three without naming your own hotel, each in a fresh chat.

1. A hotel with children and parking in your city
2. Where can I stay with a dog?
3. Compare my hotel with its nearest competitor
4. What is known about this hotel's pros and cons?
5. Where can I officially book a room?

Add one line to each: "Cite your sources. If you do not have the data, say so." Ten minutes of that will tell you more than most audits. You will see what the internet says about you, and more to the point, who is standing where you should be.

Download the hotel AI-readiness checklist (PDF, 2 pages): fifteen checks in five areas, five prompts to run, and a table for three tasks with an owner and a deadline.

Frequently asked questions

Do people really book hotels through AI assistants?

They book through your site or an aggregator. AI is the step before that, replacing the phase where the guest searched and compared. The result appears as a branded search or a direct visit, not as a booking "from ChatGPT". That is why it is so hard to measure and so easy to underestimate.

Do I need a new website for this?

Usually not. Eighty percent of the work is content: precise facts, an FAQ section, separate pages for guest needs, and aligning your data with maps and aggregators. Consider a new site only when the current one is slow on mobile or cannot be extended sensibly.

How long before changes show up?

Technical fixes and content land in indexes within weeks. Reputation and external mentions take months. Expect the first measurable shift in two to three months, assuming the content does not quietly revert in the meantime.

Should I block AI crawlers so they do not steal my content?

For a hotel it makes no sense. The content you would be protecting is your room descriptions and house rules, which is exactly what you want the model to know. Blocking AI crawlers is a choice for publishers who sell content, not for accommodation that wants to be recommended.

Is a good Booking.com profile enough?

It helps your visibility, not your direct bookings. Your Booking.com page ranks for Booking.com. When a guest has decided and comes to your site to verify the details, those details have to be there. Otherwise they go back to the aggregator, and you pay commission on a guest AI brought you.

What does this preparation cost?

The content side is mostly your team's time plus a few hours of copywriting. The technical side, meaning structured data, llms.txt, speed and indexation checks, is a few hours of developer work on an existing site. If you want us to look at your hotel specifically, get in touch and we will send you this same checklist filled in with your numbers.

The bottom line

There is nothing magic about any of this. Language models reward hotels that state precisely what they offer, state it the same way everywhere, and run a site you can book through without friction. Which is, incidentally, exactly the hotel a human would pick.

Compared with SEO it has one advantage for now: almost nobody is fighting over it. How long that lasts is anyone's guess, but I would not bet on more than a year or two.