Italiano

Atlas for Agents

An index of every hotel registration in one Italian region, and of what an AI agent could actually reach of each one on 30 and 31 August 2026.

The universe is defined by law rather than sampled: every registration in the Alberghiero segment of Puglia’s regional register was attempted, and every outcome is recorded, including the failures. Each value that has a basis carries it, and absences are records rather than gaps.

What it measured

1,117

registrations The whole segment, not a sample. Register snapshot 2026-08-29.

446

have a booking surface an agent can query 671 do not. The surface itself is served per record, with the evidence that identified it.

146

refuse this crawler Measured against a crawler that identifies itself, is contactable, honours robots.txt, and makes one request per host every three seconds. Whether they admit other crawlers is not measured, and stays unmeasured rather than guessed.

1,058

coordinates carry a stated precision Median 363 m, worst 18,529 m, across ten named methods. The remaining 59 say the precision is unknown, and say why.

147

have no web presence at all Not unreachable, not refusing: absent. The distinction is a field, not an inference.

87

corrections, published Each one says what was wrong, when it was found and with which instrument. Served values cite them by identifier. This is the number we are least tempted to hide and most willing to be judged on.

Why it exists

A machine-readable assertion does not carry what it is worth. A coordinate, a telephone number or a booking URL reaches an agent with no error rate, no scope and no date, and the agent has no way to doubt it. This index publishes that worth as fields: how strongly each value is established, where it is broken, for which question it may be used, at what date the label was assigned, and with which instrument.

What it is not

This is not a source of:

• room rates or any tariff
• availability for given dates
• booking conditions, cancellation terms or deposit rules

No record carries a price, a rate, an availability calendar or a set of terms, and no placeholder stands in for them. Searching the package for price finds nothing, and the package says so in the same place it says everything else.

What is here instead is the route to the answer: the booking surface itself, and whether an agent can put a question to it. The answer is obtained there, against the publisher, at the moment it is asked, which is also the only moment at which a price is true.

Take it

The deposit

The whole measurement is deposited as a single citable archive: the index, its 1,117 leaf records, the schemas, the licence texts and the documents that explain the boundaries. It is fixed by a digest computed over the index and every leaf, and the method for recomputing that digest is published inside the package, so a reader can check the bytes rather than trust them.

sha-256  3d040c227ba65b7c0a895b4eb0a1782b7533c5bc4b05141da128444bb3351b74

DOI 10.5281/zenodo.22335632 — this version, the archive whose digest is above. 10.5281/zenodo.22335631 resolves to the newest version instead: cite the version DOI for a figure, the concept DOI for the work. A figure belongs to the archive it was measured in.

The deposit carries the corrections the served data cite, written as what changed in the artefact you are holding. The full corpus, with every correction verbatim, travels with the complete copy, supplied on request.

Reach a person

If a record about your property is wrong, write to info@atlasforagents.com with the registration and what is wrong. Corrections are published, numbered and dated, and the served values cite them: there are 87 so far, and yours would be number 88 rather than a silent edit.

If you would rather we did not crawl you, the crawler is documented at /crawler, it honours robots.txt, and crawler@atlasforagents.com reaches a person. A property that refuses is recorded as refusing, which is a result and not a penalty.

This page

The same content is served as plain text at /llms.txt. If the two ever disagree, that is a defect and we would like to hear about it: a project that measures what machines can read from a page should not publish a page that says something different from what it serves.