Free Report: What Really Drives Hotel Visibility in AI Search — Key Findings from the Stiplo Barcelona Study of 600+ Hotels

When travellers ask AI where to stay in Barcelona, one in five of the city’s registered hotels will never come up. Not in any answer, in any language. That’s the headline finding of a new study that asked the four leading AI assistants the questions real travellers ask, more than 27,000 times, to map exactly which hotels the machines recommend, which they ignore, and why.

Carlo Del Mistro, founder of London based Stiplo — a digital mystery shopper for hotels.

The research comes from Carlo Del Mistro, former Chief Digital Officer of Ennismore (the hospitality group behind brands including The Hoxton and Mondrian) and now Founder of London based Stiplo, a start-up that acts as a digital mystery shopper for hotels.

A regular at our London Travel Massive events, and a previous speaker, I sat down with Carlo to unpack what he found: why a perfectly rated guesthouse can be invisible to AI while a middling chain hotel is recommended hundreds of times, why a third of AI’s suggestions never touch the live web at all, and what hotels can actually do about it.

Why Barcelona, and why ask 27,000 times

Matthew: You asked AI assistants the same hotel questions real travellers ask over 27,000 times. What lead you to doing this kind of stress test? And what made you start with Barcelona?

Carlo: Over the last couple of years I have watched the question of what drives hotel AI visibility go from niche to one of the hottest topics in travel. There is a lot of good research on who is visible. What I could not find was proper scientific research on what actually drives that visibility. That gap is why we ran the study.

The Stiplo Barcelona Study measured 659 registered properties across 27,360 AI travel answers.

I wanted a closed context. Rather than sample loosely, we took a single city, looked at its entire hotel universe, and asked it the full range of questions real travellers ask. We seeded those from databases of real ChatGPT queries, so we were testing what people genuinely type, not what we imagined they would.

We asked more than a hundred questions, across four AI engines, in three languages, and we asked each one twenty times. That last part matters. Ask an AI the same question twice and you get different answers, so a single pass tells you little. We stopped seeing new hotels appear after about eighteen passes, so twenty gave us a stable picture.

As for why Barcelona. It is a wonderful city for tourists, which appealed to me. but more importantly, it is trilingual. A hotel there is expected to have content and to rank in Spanish, Catalan and English, and that made it a uniquely revealing place to start.

The one in five that AI never sees

Matthew: Let’s start with the number that stopped me: one in five registered hotels in Barcelona is completely invisible to AI - never recommended once, in any language, by any assistant. Were you expecting that?

Carlo: It was one of the findings that stopped me too. I had seen plenty of numbers thrown around on AI visibility, but most studies ask only a handful of variations of each question. When you ask 27,000 questions, you see the full picture, and it is stark.

The study discovered that one in five hotels in Barcelona are completely invisible on AI search.

One in five hotels in Barcelona is completely invisible. Never recommended once, in any language, by any engine. I was not expecting that. If anything I expected something closer to ninety or ninety-five per cent of hotels showing up somewhere.

What is just as striking is the concentration. Around twenty per cent of hotels capture more than eighty per cent of all the mentions, and then there is a very long tail. The top twenty hotels are named on average more than a thousand times each. The bottom half of the hotels that are visible at all are named roughly 10 times, and the long tail just once or twice. So once you fall out of that top group, you do not just slip down the list. You effectively disappear.

What actually moves the needle, and what’s just noise

Matthew: You tested the tactics that dominate the ‘generative engine optimisation’ conversation. Which ones held up, and which didn’t?

Carlo: This is where the study became genuinely exciting. It is one thing to see which hotels are visible. It is another to work out why. We matched the hotels the AI favours against everything we could measure about them, then used several statistical techniques to separate the signal from the noise.

What came out was clear. Four things genuinely drive visibility, and four things the industry talks about constantly turn out to do very little.

The single strongest driver, by a distance, was the number of Google reviews a hotel has. Stronger than anything technical, stronger even than how authoritative the website is. After that: who links to you from around the web, then whether you already rank in Google’s own results, which the AI leans on, and finally just having pages a machine can actually read that answer real guest questions. On that last one the twist is that having the content matters far more than how much of it there is.

And the four things the industry sells hardest did almost nothing.

• An llms.txt file, the tactic of the moment: nothing.
• Opening or closing your site to the AI crawlers: no measurable difference to whether you get named.
• FAQ and schema markup: no independent effect.
• Freshness and page-speed tuning: nothing once you account for the fact that big professional sites are simply fast to begin with.

The uncomfortable takeaway is that the technical GEO checklist being sold right now, llms.txt, bot rules, schema, freshness, is measurably not what moves the needle. Reputation and readable content are.

The new gatekeepers: listicles and Reddit

Matthew: When AI recommends a hotel, you traced where it’s actually getting its information. Barely any of it comes from the hotels themselves or official tourism boards. Over three-quarters comes from guides and listicles. Who are these new gatekeepers, and should we trust them?

Carlo: It is fascinating to trace. When you ask an AI a question it does two things: it runs its own internet search, usually leaning on Google, and it draws on what it was trained on months or years earlier. It blends the two. What is striking is how little of the final answer comes from the hotels themselves. More than two-thirds of it, about 68 per cent, is assembled from guides and listicles. Only around 7 per cent comes from the hotels’ own sites, and under 2 per cent from official tourism boards.

And they are not the gatekeepers you’d hope for. The listicles doing the work are mostly local SEO guide sites, barcelona.com, thehotelguru, foreverbarcelona, not the New York Times or Condé Nast.

The source the models trust most is Reddit. It only appears in about one in eight ai web searches, so it is not the most common source. But when it does appear, the model actually cites it as a source about two-thirds of the time, a higher rate than any OTA or guide. So a machine is now assembling your reputation partly from commercial guide sites and partly from Reddit threads, and most hotels have no idea those are the pages deciding whether they get recommended.

When AI still sells the hotel you used to be

Matthew: The Grand Hyatt was recommended under its old name, Hotel Sofia, over 200 times in your data. Hotels spend millions on rebrands. Your data suggests AI can keep selling the old identity for years afterwards. Is there anything a hotel can actually do to update a machine’s memory?

Carlo: What the Grand Hyatt shows is how long it takes an AI to catch up with reality. We forget that these models are trained months, sometimes years, in advance, and then they carry a fixed cut-off. The hotel changed its name from the Hotel Sofia, but in the model’s memory it is still the Sofia, and it keeps recommending it that way.

Blocking AI crawlers can’t help here either. It won’t stop the model recommending you from memory, and it does shut the door on the systems that might otherwise pick up your new identity. To be clear, opening the crawlers won’t get you recommended more, that isn’t what drives visibility, but it’s what lets the model read the new reality when it does look.

The encouraging part is that much of this is in the hotel’s hands. When you rebrand, or move information around your site, the priority is to stay readable to AI. Let the inference bots in, and make sure your schema and metadata reflect the new reality. The Grand Hyatt is a vivid example because it is such a prominent property, but I see the same thing constantly on hotels nobody would notice.

Why the machines skew to the top end

Matthew: I thought it was fascinating that four-star-plus hotels are about a third of Barcelona’s register but take over three-quarters of AI’s recommendations. Meanwhile nearly half of the budget properties never appear at all. Why does AI skew so hard toward the top end?

Carlo: There are a few reasons the upper end pulls so far ahead. First, resourcing. Four-star-plus hotels are more likely to be part of a chain, with real teams behind the website and the online presence, so they tick the boxes AI visibility rewards. Second, coverage. They are the hotels the guides and listicles tend to write about, and that third-party content is exactly what the AI leans on. Third, they simply produce more of their own content, where a smaller hotel rarely has anyone whose job that is.

But the more encouraging story is at the other end.

Some of the smaller three-star and budget hotels punch well above their weight, and they do it by owning a niche.

We had hostels that were close to invisible for general questions. Ask about a backpacking trip or a weekend away with friends, though, and those same hostels were consistently in the top 10. We saw the same for football trips, people heading to the stadium for a match. If you truly own a category, it does not matter whether you have five stars or none. You win the questions that matter to you.

A third of answers come from memory, not the web

Matthew: Roughly a third of recommendations come from the model’s memory, not from searching the live web. What does that mean for a hotel that redesigns its website today?

Carlo: It means patience, and it means keeping the door open. If roughly a third of recommendations come from the model’s memory rather than a live search, then the beautiful new website you launch today is invisible to that third until the models retrain. A redesign is not a switch you flip. What you can do is make sure that when the AI does look, and when it does retrain, it can actually read you. Keep the site crawlable, keep the structured data accurate, and treat visibility as something you earn over months, not overnight.

Winning the mention, losing the booking

Matthew: Say a hotel does win the recommendation. You found that even then, the AI usually doesn’t send the traveller to the hotel’s own site. Where does the booking actually go?

Carlo: This, for me, is the finding that matters most commercially, and it is the one hoteliers least expect. Getting recommended and getting booked turn out to be two completely different games.

When an AI names a hotel, it links through to that hotel’s own website only about seven per cent of the time. More than nine times out of ten, the traveller is pointed somewhere else. The click lands, in order, on a local “best hotels in Barcelona” listicle first, then on an online travel agency, booking.com most often, with TripAdvisor acting as the review gateway.

A hotel can win the recommendation and still lose the booking, and with it the margin, to an intermediary.

I think of it as two separate battles. The first is being found: getting named when a traveller asks. The second is being chosen and keeping the guest on your website for the booking once you are named. Most of the industry conversation is stuck on the first, on visibility. But you can be perfectly visible and still watch the AI hand your guest to an OTA. Winning the mention is only half the job. The other half is making sure that when you are mentioned, the path actually leads back to you.

It answers in Catalan, but it searches in English

Matthew: You caught ChatGPT searching in English even when the traveller asks in Catalan. What should local hoteliers in Barcelona take from that?

Carlo: This was one of the results I least expected. I did not think ChatGPT would be doing its actual searching in English. This is a ChatGPT trait specifically. Ask ChatGPT in Catalan and roughly four in five of its underlying searches ran in English, with Catalan almost abandoned. Even in Spanish it searched in English about 60% of the time. It answers you in your language, but it goes looking in English. Perplexity, by contrast, tends to search in the language you asked in.

The lesson for a local hotelier is uncomfortable but simple. Your English content is not optional, even if your guests are local. You could have beautiful Catalan content and still be invisible, simply because the engine went looking in English and did not find you. This almost certainly holds for a lot of smaller languages, and it is one of the reasons we chose a trilingual city for the first study.

If you do one thing this quarter: Google reviews

Matthew: If a hotel owner reading this can only do one thing this quarter, what does your data say it should be?

Carlo: When we cut through everything, the single biggest driver of visibility was the number of Google reviews. Not OTA reviews, not TripAdvisor. Google.

So if a hotel owner does just one thing this quarter, they should look hard at where they send guests to leave a review. If you are pointing them to TripAdvisor or an OTA, you are handing that signal away. Redirect it to Google. And here is the part people find surprising. The quantity of reviews matters more than the score itself. Of course you want good reviews, but the sheer number counts for more than most people think.

— Thanks Carlo, for sharing your insights!

Download your free copy of The Stiplo Barcelona Study.

Carlo Del Mistro is the founder of Stiplo, a digital mystery shopper for hotels, and former Chief Digital Officer of Ennismore. Download the full study covering all 659 of Barcelona’s registered tourist accommodations, across four AI assistants and three languages including the complete methodology, engine-by-engine findings and what hotels should do next.

Matthew Gardiner Is the Director of Travel Massive London, a fractional marketing leader and growth partner to travel brands and destinations.