How the AI Hotel Index is measured

Personas and themes

Every question is written from one of eight traveller personas, or as a general discovery question, and tagged with one theme. A hotel's persona profile counts the 10 questions written for each persona, asked in both engines; the theme profile counts the questions tagged with each theme the same way.

The 100 questions

20 general discovery questions and 80 persona questions, asked once each per engine, with the destination name substituted. Published in full for both markets; the DE set is a native rewrite, not a translation. Region pages (Cyprus, Crete) use a separate set of 12 questions per market.

Engines and settings

Two engines, as a traveller would use them. ChatGPT: the public app's default model, web search on, no custom instructions, a fresh conversation per question. Google AI Mode: signed-out, default settings. Both run with the market's location and language: a UK location with English (UK) for the UK set, a German location with German for the DE set.

One run a month

Each destination and market is run once a month, in the first week, over at most three days. Answers are not re-asked if a hotel is missing; an answer that fails to load is retried once and otherwise counted as an answer with no hotels named. The edition is named after the month of the run and is shown on every page.

Scoring

Score is the share of the 200 answers in which the hotel is named, as a number from 0 to 100. A hotel named in 42 answers scores 21. A hotel counts once per answer however many times the answer repeats its name. First choice is the share of answers in which the hotel is the first one named. ChatGPT and AI Mode columns are the answer counts out of 100 each. General discovery is the count out of the 40 destination-level answers.

Precise ranks are shown for scores of 10 and above, ordered by score, then by ChatGPT answers. Below 10 the hotels are grouped into bands, listed alphabetically: named in 6 to 19 answers, named in 1 to 5 answers, not named in any answer. Bands always carry their text label.

How citations are counted

For ChatGPT only the explicit source list under an answer is counted; links inside the answer text are not. The share of answers that had a source list is published next to the citation shares. A hotel's own site counts as cited when its domain is in the source list of an answer that names it. For Google AI Mode, each reference is attributed to the surface that backed it: Business Profile card, web page, or other Google card. Domains are grouped as hotel websites, Booking, TripAdvisor, Expedia, other review sites and other.

The universe rule

A hotel is in the index when it is classified 4 or 5 star by the national authority (Cyprus Deputy Ministry of Tourism, Greek Chamber of Hotels). Where no official classification is published, the Google star class is used, and where Google carries none, the class listed by the major booking sites. Apartment hotels are included if classified; villas are not. Hotels can ask for inclusion or a correction at any time; the change takes effect at the next run. The universe is reviewed every edition, so a hotel can enter or leave it as classifications, openings and closures change.

Reading the numbers

The index is directional, not a guarantee. AI answers are stable in the short term: asked the same question again within the hour, ChatGPT and Google AI Mode name the same leading hotels, with Google AI Mode varying more than ChatGPT further down the list. From our observations they change over time, including the leading choices. Each edition records what the engines said in one run of 200 answers per destination and market; the next run can place a hotel higher or lower without anything about the hotel having changed. Read a score as a level and a direction over several editions, never as a precise rank for a single month.

Which hotels an engine includes in an answer, or leaves out, is the engine's doing. We do not choose, weight or filter the hotels named; we ask the frozen questions, store the answers and count the names against the universe. A hotel absent from an answer was not named by the engine, nothing more. We measure the stability between editions for every destination and quote it whenever we present the numbers.

Comments and corrections

There is no pre-notice before publication: an edition goes live for every hotel of a destination at the same time. Every hotel is free to send us a comment on its row, or a correction to its name, star class, destination, website or aliases, at any time. Write to cosmin@tharro.io with the hotel name in the subject, or use the contact form. Hotels cannot opt out of the index; the measurement covers the whole universe or it is not a measurement.

Use of the data