Independent research measuring how much Italian accommodation an AI agent can actually reach.

What this is

A measurement, not a product. Not a booking site, portal or agency.

The question is not whether a property can be booked. Obviously all of them can — through the large OTA platforms, through direct online booking, by email or phone, or some combination of those.

What this project studies is which of those channels each property actually has, and which of them an AI agent can use directly in order to:

establish the basic facts about the property on its own — CIN (the Italian national identifier for tourist accommodation), location, type
find its way to further information, starting with the property's own website
know which channel, if any, it can use to learn real availability, prices and conditions

Nobody has measured that. This project is measuring it, over a universe defined by law rather than by distribution: the accommodation entered in Italy's regional registers — hotels, guesthouses and the rest — published as open data. The first sample is the hotels of one Italian region (Puglia).

Who is included

Every property properly entered in the register. All of them, in the covered area, without selection and without exception.

Atlas is agnostic about which properties matter. It does not rank them, does not choose between them, and has no way to tell a customer from a non-customer. The register does the selecting, and the register is public.

That agnosticism is what makes the result a measurement rather than a listing. Commercial accommodation databases are fed by suppliers: a property enters because someone sells it. That makes them structurally unable to answer this question — a property with no distribution has no reason to appear at all, and a property with distribution but no direct channel looks identical to one with both.

A register-anchored population can answer it, because absences are visible.

What is collected

For each registered property: whether it declares a reachable booking path — and where it does not, the absence itself is recorded — which booking system serves it where identifiable, that page's address, and the date of observation.

Not collected: prices, availability, reviews, ratings, customer or reservation data, or the content of pages beyond what is needed to locate the booking path.

What remains, for each property, is the answer to one question: is there a direct way to book that a machine can find, in order to query availability and potentially book autonomously or with assistance? Atlas records whether that channel exists. It does not use it.

Full crawler conduct, and how to be excluded, at atlasforagents.com/crawler.

Method commitments

These are stated because they constrain the result, not because they are required:

Inferred and observed values never share a field without a discriminator
Every produced value carries how it was obtained and how far it can be trusted
Confidence figures are derived from a measured error rate, or the field is absent
Absences are published rather than omitted
Attribution and licence travel with every record, so a consuming agent can pass them to its own user

Related work

LAR — Layered Agentic Retrieval: a proposed convention letting a property declare to agents what it is and how it should be read, rather than being inspected approximately and probably misread, or ignored altogether. The paper · the specification.

This site publishes a LAR manifest at /.well-known/lar.json. A project arguing that publishers should declare their surfaces has no standing if it does not declare its own.

Who

Francesco Marinoni Moretto, independent researcher and author of Selling to Agents, a playbook on agentic commerce. No commercial relationship with OTAs, booking engines or property management systems.