Methodology
How we calculate the score
The formula is public, and this page reads it from the code that applies it. They cannot tell two different stories.
Four things measured, one number
A 0 to 100 mark for how often an AI names this territory when someone asks where to go. A hundred means it always comes up, first. Zero means that for the AI it doesn't exist.
35% How often it's recommended + 25% How well they know it + 20% Who talks about it + 20% Whether they can read it
Illustrative example: numbers chosen to show how the four factors add up to the score, not any real territory's result.
How often it's recommended
Picture someone asking an AI “where should I go to the seaside in Italy”. This number says how often this territory shows up in the answer, and in which position.
worth 35 of 100
How well they know it
If someone asks “what is there to see here”, can the AI name real places, or does it stop at the two everybody knows? This measures how well it knows the place.
worth 25 of 100
Who talks about it
Where the AI gets its information when it talks about this territory. If it takes it from Booking rather than the official site, the destination is being described by someone who profits from it.
worth 20 of 100
Whether they can read it
Can the programs that read websites on behalf of AI get in, or do they find the door shut? An AI cannot recommend what it was never allowed to read.
worth 20 of 100
Three assistants, three different answers
Assistants don't draw on the same sources, so they don't give the same answers. If your territory does well on one and badly on another, the problem is specific and so is the fix.
ChatGPTOpenAI- GPTBot
- OAI-SearchBot
- ChatGPT-User
GeminiGoogle- Google-Extended
- Googlebot
PerplexityPerplexity- PerplexityBot
- Perplexity-User
One general score, one per assistant
The general one is the plain average of the three. No hidden weighting between assistants: if ChatGPT counted more than Perplexity, that would be our opinion dressed up as a measurement.
Why five measurements a month
The same question asked twice of the same model gives two different answers. One question isn't data, it's an anecdote. Every month we repeat everything five times per assistant and publish the average with how much it moved: that ± next to the number is the difference between an observatory and a screenshot.
Why it matters whether a site lets itself be read
An AI won't recommend what it couldn't read. Each assistant sends its own programs out to read websites: if a territory blocks ChatGPT's, its ChatGPT score drops. This isn't technical hygiene, it's visibility.
Two kinds of question
In the first, the territory competes with the others in its category. In the second, we only check whether the AI knows it. They are two different things and deserve separate measurement: there are famous territories AI cannot describe, and small ones it describes beautifully.
The visibility bands
Between 38th and 41st place there is no measurable difference. Between one band and the next there is. Bands are the honest way to read this ranking.
Demand scale
How many monthly searches the questions we track here actually move. It tells you whether a ranking position is worth thousands of people or a few dozen.
- High demand40,000+
- Medium demand8,000–40,000
- Low demand1,500–8,000
- Marginal demand0–1,500
Where the territory loses ground
A trip is decided in three moments: first you dream, then you plan, finally you book. A destination can be strong in the dream and disappear when it's time to choose where to sleep.
The traveller knows they're coming to Italy but not where. A destination either makes the list or doesn't exist.
Dates, constraints, itineraries. Here the AI needs real detail, and it shows immediately when it has none.
Where to stay. The stage where one mention is worth one booking, and where individual hotels enter the picture.
Out of scope: the traveller has arrived and the booking decision is already made.
Out of scope: post-trip storytelling, better measured on social platforms than on AI assistants.
answers analysed
Every month we ask the same questions five times to each assistant. The answers change, so we publish the average and how much they moved.
- prompts per destination
- 50
- runs a month
- 5
- engines
- 3
- answers analysed
- 75,000
- First measurement: August 2026
Data status
At this stage the numbers are generated. The schema, the formula and the pages are final: when real runs replace the simulated dataset, not a line of this methodology changes.
What we take for grantedKeep exploring
Every number in this observatory can be opened, broken down and checked. From here you can go wherever you need.