How we measure
The method behind every number Tharro publishes. Written out in full so it can be checked, argued with, and reproduced.
The unbranded-only rule
Every competitive measurement Tharro publishes uses unbranded queries only.
Branded queries β searches containing a hotel's name β return accurate, positive answers for almost every property, because the model has been handed the answer inside the question. They measure nothing about competition. Unbranded queries are the ones where an assistant must choose between hotels, and they are the only ones that reveal standing.
Branded performance is measured separately and reported as brand defence, never blended into a visibility score.
Prompt design
Each property is measured against a fixed prompt set built from three inputs: the destination, the property class, and the traveller personas present in that market. Prompts are phrased as a traveller would phrase them, not as a marketer would.
The set is fixed across runs. When a prompt is added, it is versioned and reported separately until it has enough history to sit in a trend. Changing prompts silently would make every trend meaningless.
Personas
A market's prompt set spans the traveller types that actually book there β typically couples, families, business, wellness, groups, luxury and value. Persona coverage is reported separately rather than averaged, because aggregate scores hide the finding: standing is almost always uneven, and a persona gap points at a specific content or reputation problem.
Engines
Measurement runs across three surfaces β ChatGPT, Google AI Mode and Perplexity β using the same prompts, on the same day.
They are reported separately as well as together, because they disagree. Strong standing on one is not standing in general, and the three have different audiences: Google AI Mode sits inside the search box where most travel research begins, ChatGPT carries longer planning conversations, Perplexity skews toward users who want sources.
Mentions, citations and share
Three distinct measurements, recorded separately.
Mention β the property is named in the answer.
Citation β a source is credited, and it is identified by type: the hotel's own domain, an OTA, a review platform, an editorial site or a destination site.
Share of answer β of all hotels named across the prompt set, the proportion that are this property, measured against a defined competitive set.
A property can be highly mentioned and barely cited. Reporting only the first number is the most common way an AI visibility score misleads.
Competitive set
All comparative figures are read against a defined set of four to six properties β the hotels a traveller genuinely compares against β not an industry average or a regional benchmark. The set is stated on every comparative view.
Cadence and variance
Prompt sets run weekly. Assistants are non-deterministic: the same prompt can return different hotels on different runs. A single run is a sample, not a measurement, so figures are reported across runs and single-run movement is not treated as a trend.
Collection gaps are labelled on the chart rather than interpolated. A week with no collection is shown as a gap.
What we do not claim
AI visibility measures discovery, not bookings. Tharro does not claim a causal link between a visibility score and revenue, and does not model one. The commercially interpretable figure is the citation gap: recommendations routed through intermediaries carry a commission that direct citation does not.
Nor is a score comparable across markets. A 40% mention rate in a market with eight competing hotels means something different from the same figure in a market with sixty.
Frequently asked questions
Can I see the prompt set for my property?+β
Yes. It is visible in the product, and it does not change without being versioned.
Why only unbranded queries for competitive scoring?+β
Branded queries flatter every property equally and measure nothing about competition.
Why do the engines disagree?+β
They retrieve from different sources at different moments and weight them differently. The disagreement is a finding, not an error.
How large is the prompt set?+β
It varies with market size and persona spread. It is stated per property rather than fixed globally.
Do you use the public chat interfaces?+β
Measurement runs programmatically with consistent parameters, so runs are comparable to each other.
What counts as a citation?+β
A source credited in the answer, recorded with its domain and type. Being named without a source is a mention, not a citation.
Questions about the method
Questions about the method are welcome and answered directly β this page exists to be argued with.