Six independent Italian hotels. 172 questions put to AI assistants, every answer read by hand. The full figures, including the uncomfortable ones.
47 of 172
Questions in which the hotel was named — just over one in four (27 %).
Put differently: when a traveller asks an AI assistant for advice and it answers by naming two or three places, in three cases out of four the hotel we measured is not one of them.
Same method, same weeks, same four families of questions:
| Hotel | Named | |
|---|---|---|
| A | 1 / 28 · 4 % | |
| B | 3 / 32 · 9 % | |
| C | 3 / 18 · 17 % | |
| D | 8 / 30 · 27 % | |
| E | 9 / 32 · 28 % | |
| F | 23 / 32 · 72 % |
Between 18 and 32 questions per hotel, depending on how many families of questions applied.
This is the figure that matters most. It is not true that assistants ignore independent hotels: one of the six is named in almost three answers out of four. The problem is not the category — it is what the AI finds, or doesn't find, about each hotel.
We grouped all 172 questions by family:
| Family of questions | Named |
|---|---|
| Type and town — “hotels in [town]” | 11 / 52 · 21 % |
| Type of guest — “where to stay with children” | 10 / 38 · 26 % |
| Feature — what sets it apart | 10 / 38 · 26 % |
| Nearby place — “near [landmark]” | 16 / 44 · 36 % |
The simplest and most frequent question — which hotels are there in this town — is the one where hotels appear least of all. It is also the one with the most competition: the assistant has to pick two or three names out of many, and it picks those it finds consistent information about.
Two hotels out of six were not named even once when we simply asked which hotels there are in their town — zero answers out of eight attempts, in two languages, on four different assistants.
Both have a working website and published reviews. It is not a question of the hotel's quality: it is a question of what the AI manages to read.
Six hotels are few. These figures describe what we measured, not the whole Italian market, and we don't claim otherwise.
And the sample is not random. These are hotels that wrote to us after seeing an ad: people who had already started to wonder. A sample drawn at random from all Italian hotels would give different numbers — we don't know whether better or worse, and anyone who says they know without measuring is guessing.
We have not yet published “before and after” figures, because the second measurement, twenty-one days after the fixes, has not yet been made on any hotel. When it exists, we will publish it — even if it says nothing changed.
What we can say is more modest and more useful: the result varies enormously from hotel to hotel, and it can be measured in an afternoon. Until you know which of the six situations you are in, any discussion of what to do is guesswork.
The questions are fixed before looking at any result, and put to ChatGPT, Claude, Gemini and Perplexity, in Italian and in English, several times each. Every answer is read by a person, because names look alike and an automatic count inflates the result. The full method is on this page.