Business of Experience Perspective – Meriç Murat Erdinç, Founder, Stay Indexed
/The AI Shortlist: Why Most Hotels Never Make It Into the Answer
By Meriç Murat Erdinç, Founder, Stay Indexed
A traveller types a question into ChatGPT: where should I stay in London for a quiet long weekend? The answer comes back in seconds. Three hotels, maybe five, each with a line of reasoning attached. No second page, no map with two hundred pins. That short list is quietly becoming the new shelf. Search gave every hotel a place somewhere in the results, even if it was page four. An AI assistant gives a handful of names and stops. If you are on it, you are in the conversation. If you are not, the guest may never know you were an option. I wanted to see how that shelf actually behaves, so I measured it.
How the study worked
I took 40 hotels across five cities: New York, London, Dubai, Sydney and Milan. Eight per city, split evenly between four-star and five-star. Each hotel went through the same scan. Eight questions a real traveller might ask about that city, covering general stays, romantic getaways, family trips, luxury, budget and last-minute bookings, each run through ChatGPT, Gemini, Claude and Perplexity. That produced 1,277 AI answers. For each one I recorded whether the hotel being scanned was recommended, where it sat in the list, and which hotel the assistant suggested instead. The hotels are anonymised throughout. The point is the pattern, not any one property.
Most hotels were never mentioned at all
Across all 1,277 answers, the hotel being scanned was recommended just under 12 percent of the time. That average hides how uneven the picture is. 19 of the 40 hotels were not recommended in a single answer. Not once, across four assistants and eight questions. Another 11 appeared between one and four times. Only four hotels were recommended 15 times or more. Put differently, more than half of the hotels in the study were recommended once or not at all. These are not obscure properties. They are established four and five-star hotels in five major international cities.
The shortlist is winner-takes-most
The mentions that did happen were concentrated in very few hands. The four most-recommended hotels collected 84 of the 149 mentions in the study, a little over half. The top eight took almost four in every five. The same concentration showed up from the other direction. Every scan also records which hotel the assistants recommended most often instead of the one being scanned. In a typical scan, that leading alternative was named around 10 times out of 32. The typical hotel in the study was named once or not at all. So the realistic picture for most hotels is not "we rank fifth instead of second". It is that a small number of names in each city absorb most of the recommendations, and the assistants repeat them question after question.
Stars did not buy a place on the list
I expected five-star hotels to dominate. They did slightly better in aggregate, with 78 mentions against 71 for the four-star group. Beyond that, the tier made surprisingly little difference. Nine of the 20 five-star hotels were never recommended. Ten of the 20 four-star hotels were never recommended. Almost the same. More telling, the two most visible hotels in the entire study were both four-star. One scored 84 out of 100 on visibility, recommended in 30 of its 32 answers. The other scored 71. Both sat well ahead of every five-star property in the sample. The most visible hotel of all also had the lowest technical readiness score of the 40, at 53 out of 100. Its website was, by the scan's measure, the least prepared for AI of any hotel I looked at. It was recommended almost every time anyway. That is a single case and I would not build a rule on it. It does fit a broader pattern in the data, though: being readable to AI and being chosen by it look like two different problems.
The four assistants do not agree with each other
In aggregate, the four assistants recommended the scanned hotels at broadly similar rates. Gemini named 16 different hotels from the sample, Claude 15, ChatGPT and Perplexity 11 each. They just did not name the same ones. Of the 21 hotels that were recommended at least once, only eight appeared on all four assistants. Seven appeared on just one. Four of those seven were visible only on Claude, three only on Gemini.
For a hotel, this matters more than it sounds. If your team checks visibility by asking ChatGPT a couple of questions and finds nothing, you might be visible on Gemini. If you check Gemini and find yourself, you may be absent everywhere else. A single spot check on a single assistant tells you very little. Guests do not all use the same one, and the assistants are not drawing from one shared list.
Visibility depends on the question
The scan breaks results down by the type of question asked, and the gaps were consistent. Among the hotels that were visible at all, the most common weak spot was romantic getaway searches. It was the weakest category for 10 of those 21 hotels. Family trip searches were the weakest for another five. At the other end, last-minute searches were the strongest category for eight hotels, followed by general "where should I stay" and luxury stays. This is useful because it is specific. A hotel with a well-reviewed spa and couples packages that never appears for romantic getaway questions has a gap it can name and work on. "We are not visible in AI" is not actionable. "We are not visible when people ask where to go for an anniversary" is.
So what gets a hotel onto the list?
The honest answer is that nobody outside the AI companies knows the full mechanism, and anyone selling certainty on this is ahead of the evidence. My working explanation is that assistants recommend the hotels other people write about. Across the scans, the sources the assistants relied on were overwhelmingly third-party: travel guides, editorial round-ups, review platforms, local press. Hotels' own websites barely featured. Technical work on your own site helps a model read the property correctly once it has chosen to mention it. It does not appear to be what gets a property chosen in the first place. That shifts where the effort should go.
A practical checklist for operators
If I were running commercial or marketing for a hotel today, these are the five things I would do first.
Check all four assistants, not one. Ask ChatGPT, Gemini, Claude and Perplexity the same handful of questions a guest would ask about your city. Write down who appears. The disagreement between them is part of the answer.
Ask the questions by intent. Run separate questions for couples, families, business travel, luxury and last-minute stays. Most hotels are visible for some of these and invisible for others, and the gaps are where the work is.
Learn who sits on your shortlist. Note which hotel is recommended instead of you most often. That property is your real competitor in AI answers, and it may not be the one your revenue team benchmarks against.
Audit what is written about you elsewhere. Look at the guides, round-ups, review platforms and press that cover your city. Are you in them? Is the information current? An out-of-date description in a well-read guide can become the assistant's version of your hotel.
Keep the technical basics in place, but do not stop there. Structured data, crawler access and clear descriptions are cheap and worth doing. They help an assistant describe you correctly once it has decided to mention you. They are unlikely to get you mentioned on their own.
What this study does not tell you
It covers 40 hotels in five international cities. It does not include Indian markets, and the pattern there may well differ, which is worth testing rather than assuming. AI answers are live and they move. A scan run next month could produce different placements for individual hotels, though the broad shape, a short list dominated by a few names, held across every city and both tiers. It also measures recommendations, not bookings. There is still no clean way to trace a reservation back to a specific AI answer, and I would be wary of anyone who claims otherwise.
Where to start
Every hotel's position is its own, and the only way to know yours is to look. Stay Indexed runs the same scan used in this study on a single property: the four assistants, the eight traveller questions, the hotel most often recommended instead of you, and your weakest question type. The free scan takes about two minutes at stayindexed.com. If you run it on your own hotel, I would be glad to hear what came back, especially from operators in markets this study did not cover. I am at info@stayindexed.com.
Images: courtesy of Meriç Murat Erdinç / Stay Indexed
