The question, asked the way a guest actually asks it
Hotel Savoy Bern wanted to know one thing: when someone asks ChatGPT, Perplexity, Gemini or Claude where to stay in Bern, does its own name come back. So we wrote ten questions the way a guest actually types them, not the way a hotel would, and asked each one in English, German and French through the engines’ own grounded search APIs. That is ten questions, three languages, four engines, 120 calls. Three answers cited no sources and were discarded rather than counted as answers naming nobody, and one call failed. 116 answers were kept, and every figure below comes from those 116.
The headline: named less than half the time, and nowhere near evenly
The Savoy was named in 54 of 116 answers, 46.6%. A guest asking one of these ten questions had a little under a coin flip’s chance of hearing its name. Hotel Schweizerhof, by comparison, appeared in 82 of the same 116.
That average hides a much sharper pattern. Two questions named the hotel every single time: “a good four star hotel” came back 12 of 12, and “a hotel near the station, walking distance” came back 11 of 11. One question named it never. “A hotel for a conference or event” returned 0 of 11, in the city that hosts the Swiss federal parliament and its conference traffic. Business travel was almost as weak, 2 of 12, and in German it was 0 of 4.
By language, German was the weakest at 37%, against 52% in English and 50% in French, an odd result for a hotel whose own website opens in German. By engine, Perplexity named it 60% of the time, Gemini 47%, Claude 41%, and the engine most people mean when they say AI search, OpenAI’s ChatGPT, named it least at 37%.
The diagnosis: not losing the comparison, missing the shortlist
Here is the part worth publishing on its own. When an engine answers a question like these, it quietly writes and runs a number of its own searches first, casting around for candidates and then checking a specific name once it is seriously considering it. We kept every one of those searches: from the 30 questions asked, the four engines ran 546 of their own.
The hotel’s name appeared in one of those follow-up searches 11 times out of 116 answers. For most of these questions, the hotel was not being compared and found wanting. It was not entering the round where the comparison happens at all, which is a different problem from a weak review score and it needs a different fix.
The pattern lines up with the question-level results above. Wherever an engine already held a firm, simple fact about the hotel, its location or its star class, it named the hotel every time. Wherever the answer depended on something true about the hotel that nobody had published in a form a machine could read, meeting space, quietness, a room that sleeps a family, the hotel dropped out of consideration.
You come up half the time for the questions you would win anyway, because someone who already wants a good four star hotel near the station was going to find you. The 46.6% is mostly answers that bring in little money, and none to the question that fills rooms on a Tuesday in November. What worries me more is the other number: the engines ran 546 searches of their own and your name appeared in 11 of them. That means AI decides who makes the shortlist before it ever looks you up, so when you’re absent from a question you’re not ranked low, you simply aren’t in the room. An average hides where you score zero, and that zero is where the money is.
Sebastián Ocampo, founder of Y8Y
What the guests already knew
Guests had already written the fact the engines were missing. One review called the hotel “possibly the quietest city hotel I have ever been in.” The German question for a quiet, central hotel named the hotel 0 of 4 times. The quality was real and documented in the hotel’s own review history. It simply lived in a guest’s prose, in a place none of the four engines read as a fact about the property.
What we expected to find, and did not
Before measuring, we expected a different cause: three third-party sites were found publishing machine-readable data about the hotel with a wrong postcode, a disconnected phone number and a rate quoted in the wrong currency. In this run, the four engines cited those three sites zero times. They remain worth fixing, and they are not what cost the hotel the conference question. We are saying so here because a report that quietly drops its own wrong guess at the start is not a report worth paying for.
On citations generally, hotelsavoy-bern.ch was read about as often as Bellevue Palace’s own site, 21 citations against 22. The site is being read. What it is not doing is stating the facts above in a form a machine can lift as fact rather than infer from marketing prose.
One comparison we can stand behind, and the one we will not make
A single comparison in this data holds up: one hotel in the same city, with a name distinctive enough to count without ambiguity, appeared in 82 of the same 116 answers against this hotel’s 54. That is a real, sourced gap between two named properties on the same thirty questions.
What we will not do is turn that into a ranking of Bern’s hotels. These ten questions were our choice, not a census of the market, and the answers behind them contain roughly sixty distinct hotel names, several of them the same hotel written two different ways. Resolving all of those to a real league table is a separate piece of work we have not done, and naming it as a ranking would be exactly the kind of invented precision this whole report is arguing against.
What we recommended
Four moves came directly out of the numbers above, in the order the evidence puts them.
Who can actually act on this list matters as much as the list. The Savoy belongs to BâleHotels Group, which is a division of Coop Genossenschaft, and its website runs on group tag infrastructure. So “put structured data on the site” is a group IT ticket rather than a decision the general manager can take on a Tuesday, while claiming a Google Business Profile description probably is theirs. A fix list that does not say which is which gets forwarded rather than actioned, and that distinction applies to every branded or group owned property reading this.
- State the facts the hotel already has, in a form a machine can read. Room count, the pedestrian street, the two-minute walk to the platform and the connecting rooms guests keep thanking the hotel for are nowhere in structured data on the hotel’s own site.
- Win the conference question. Zero of eleven, in a capital city, is the single most valuable gap in the report, and closing it means stating what the hotel can actually host, in the places the engines already look.
- Close the German gap. 37% against 52% in English is backwards for a hotel in a German-speaking canton, and two of the three answers discarded for citing nothing at all came from the same engine working in German, which suggests the sources it wanted simply were not there.
- Re-measure on a schedule. This is one Tuesday. The only way to tell a real fix from the engines simply having a different day is the same ten questions, asked again, on a stored date.
The honest limits
We chose the ten questions, not the hotel, so a session on the hotel’s actual best guest would change several of them. Two of the six engines this site tracks, Google AI Overviews and DeepSeek, have no usable grounded search API for this kind of run and are not in any of the figures above; both would need to be run by hand, and neither has been yet. And this is one measurement on one day. A second reading, not this one, is where a trend starts to exist.
Which one are you
Ask ChatGPT for a place to stay in your village or street, then ask for the thing you’re best at with no place name. If only the first returns you, you’re Nonna Giulina, because AI knows you exist and doesn’t know what you’re for. If you show up for both but not for the question that pays your bills, you’re the Savoy, because AI knows you well and describes you to the wrong guest.
Sebastián Ocampo, founder of Y8Y
See also
Up: Measure AI search visibility for hotels and Airbnbs. Across: How to report AI search visibility to hotel owners and The AI presence score explained. Down: AI presence score.
If a property wants this run done on its own name, in its own languages, the program that does it on a monthly cadence is at for hotels.