Hotel AI visibility metrics: which ones matter and which are vanity

AI visibility means whether ChatGPT, Gemini, Claude, Perplexity, Google AI Overviews or Google AI Mode names your hotel when a traveller asks where to stay. These tools run on large language models (LLMs), software trained on huge amounts of text that writes an answer instead of returning a list of links. Their answers change from one run to the next, and that makes the topic easy to measure badly. A screenshot, a score or a traffic chart can each look convincing and still say nothing about whether guests hear your name before they book.

Our guide to measuring hotel AI visibility walks through the method step by step: which questions to ask, which engines, and how often. This article covers the layer above that. Once you have numbers coming in, which ones should a GM, owner or marketing manager put in the monthly report, and which ones should come out of it?

A metric is worth reporting if a change in it would change what you do. If it could go up or down and you would carry on exactly as before, it can stay out of the report.

Detection rate is the share of AI answers that name your hotel, out of all the answers checked. The part that matters is which questions go into it. A non-branded question is one a guest asks before they know you exist: "best boutique hotel in Prague for a weekend", "family hotel near the ski lifts in Pec pod Sněžkou". Those are the questions where an AI assistant decides whether you get considered at all.

Golden Well Hotel in Prague was named in 21 of 295 non-branded questions, 7.1%. That number is useful because of what sits behind it. The hotel was named for occasion questions such as "best hotel in Prague for a honeymoon" and missing from broad ones such as "good hotels to stay at in Prague". Mosaic House Design Hotel showed a similar split: named for identity questions like "unique hotels to stay in Prague", absent from location questions like "hotels in Prague city centre". Read at question level, the rate tells you which kind of guest finds you through AI and which kind does not.

For scale, Olympia Hotel in Mariánské Lázně scored 0% across 299 checks on five engines. Low rates are normal at the moment. What you are looking for is your own trend on a fixed set of questions.

A detection rate tells you how often you are named. Share of voice tells you who is named when you are not. Two hotels can sit at the same rate in very different positions: one in a market where AI rarely commits to any hotel, the other in a market where the same rival appears in nine answers out of ten.

The competitors worth tracking come from your own answers, which may differ from your rate-shop comp set. In Pec pod Sněžkou, the AI answers that left out Bouda Mama mostly cited other local hotels, Grandhotel Hradec among them, along with regional directories. In Prague, the gap went mostly to Booking.com and TripAdvisor. Losing to an OTA and losing to the hotel down the road call for different work, and only a share-of-voice view tells you which one you face. Our share of voice guide covers the arithmetic, including why the shares do not add up to 100%.

Each engine builds its answers from different sources, so one blended figure averages away the most useful detail. Golden Well's 7.1% was made of 9 of 60 answers on Gemini, 7 of 60 on Perplexity, 5 of 60 on ChatGPT, 0 of 60 on Claude and 0 of 55 on Google AI Overviews. Mosaic House was named in 11 of 40 Gemini answers and 0 of 40 on AI Overviews.

Across the four hotels in our case studies, three had 0% on Google AI Overviews for generic questions. A hotel that sees only a blended rate might decide its website is the problem everywhere. The engine split shows that the gap sits in one or two places, and those places get the attention.

Some engines show the pages they drew on. When those citations point at Booking.com instead of your own site, the guest who clicks lands on a page where you pay OTA commission. For Mosaic House, AI cited Booking.com, TripAdvisor, Expedia and thehotelguru.com far more often than the hotel's own website. For Golden Well, Booking.com and TripAdvisor were cited most, alongside a French travel blog and thehotelguru.com.

This metric points straight at work you can do. If your own site is rarely cited, the questions are whether AI crawlers can read it, whether the facts guests ask about are written on it, and whether your listings elsewhere agree with it. A rising share of citations to your own domain is one of the clearest signs that site work is landing.

An AI assistant can name your hotel and then give the wrong check-in time, a pool you closed two years ago or a room count that does not match your rooms. These errors are called hallucinations, and they reach a guest before you get a chance to correct them.

Accuracy is best tracked as a list of specific claims, each checked against facts you have confirmed, with the source the AI seems to have used. A count of errors with no detail gives you nothing to fix. A short list such as "AI says breakfast is included, it is not, and the wording matches an old OTA listing" tells you which listing to update. Our article on AI hallucinations covers the usual causes.

Every metric above is a proxy. The owner will ask what it is worth. Revenue at stake turns your own booking figures (monthly bookings, average booking value, OTA share and the commission you pay) into an estimate of the direct revenue and commission involved when guests who ask AI do not hear about you.

It is an estimate, and any honest version says so and lists its assumptions. Treat it as a way to size the problem and to judge what monitoring is worth paying for. It is a poor scoreboard, because it moves when your inputs move as well as when your visibility does. Our revenue calculator shows the inputs and assumptions side by side.

Branded-query mentions come first. Ask ChatGPT "tell me about Hotel X in Prague" and it will usually describe Hotel X. That shows the model has read about you. It does not show that a guest who has never heard of you will be told about you, which is the whole point. Keep branded questions out of the headline rate, or report them separately.

One-off screenshots are the next trap. Someone on the team asks ChatGPT a question on their phone, sees the hotel named and shares it. The same question asked an hour later can return a different list. A screenshot records one run of a system that changes from run to run, in either direction. It is fine as a talking point and poor as a measurement.

A single AI score with no questions behind it is the third. A number out of 100 looks tidy. If you cannot see which questions were asked, on which engines, how many times, and which answers named you, you cannot tell a real improvement from a change in what was measured. Ask any tool, ours included, to show the questions and answers behind every score.

The fourth is AI referral clicks read on their own. GA4 can show visits arriving from chatgpt.com or perplexity.ai, and it is worth setting up. An AI answer can name a hotel without a clickable link, though, and a guest who reads your name and then searches for it on Google shows up as organic or direct traffic. Referral clicks are a floor. Our GA4 tracking guide explains the setup and how to read the number against your detection rate.

Bouda Mama, a family mountain hotel in Pec pod Sněžkou, and Mosaic House in Prague both have a detection rate of 12.5%. On a dashboard showing one number each, they look equally visible. They are not comparable.

Bouda Mama's 12.5% is 3 detections out of 24 checks on two engines, ChatGPT and Google AI Overviews, the engines its Radar plan covered at the time. Mosaic House's 12.5% is 25 detections out of 200 checks on five engines. On Bouda Mama's panel, one extra detection would move the rate by about 4 points. On Mosaic House's panel, one detection moves it by half a point. The smaller panel produces a number that can swing a long way from one week to the next without anything changing at the hotel.

The comparison that does hold is the engine both hotels were tracked on. On ChatGPT, Mosaic House was named in 20% of checks and Bouda Mama in 16.7% (2 of 12). Bouda Mama also holds the only Google AI Overviews mention among the four hotels, 1 of 12 checks. That is a real result and a single one, so it should be reported as one mention and left at that.

The lesson applies to any report. A percentage without the number of checks, the engines and the question set behind it cannot be compared with anything, including your own figure from last quarter if the question set changed in between.

A one-page monthly report can hold all six meaningful metrics. Start with the detection rate on non-branded questions, with the number of checks beside it and last month's figure next to it. Add the same rate split by engine, the competitors named in place of you, the share of citations pointing at your own site, a list of any wrong facts found and whether they were fixed, and the revenue estimate with its inputs. Keep the question set fixed between reports, and when you change it, mark the change so nobody reads it as progress.

Nobody can guarantee that an AI assistant will recommend your hotel, and a report should not promise it. What a report can do is show whether the work you put in is moving the right numbers, on the right questions, on the engines your guests use.

If you want a starting baseline, Staylight's free scan runs one Radar scan for your hotel: 15 guest questions across three engines, with no card required.