Atlas

A dated record of machine-readable capabilities the open web does not provide.

When software acts on someone's behalf it has to read a site rather than look at it: find the terms of use, locate a price, discover an interface. Often there is nothing there to read. Atlas measures that absence across a defined set of sites, repeatedly, and keeps an append-only signed record of what was missing and when — including seven years reconstructed from a public archive.

21dated epochs, 2019–2026

14capability classes measured

signedappend-only, anchored off-site

What this sells

A signed assay is one capability class across one declared population at one declared date, delivered with the method so the result can be recomputed against us. Scope is agreed in writing and invoiced in euros. The tiers, what each one costs, the terms, the limits of the instrument and the three things that are not for sale at any price are set out on one page.

Before deciding whether it is worth paying for, read one: a complete specimen assay is published in full — the rate, its interval, the evidence count, what was excluded and why, and the command that recomputes it.

Read a specimen assay See the engagements Request a scope

The question an assay answers is not whether a given site has an AI policy. It is what the state of the art was, across a defined population, on a given date — the form the question takes when a dispute turns on whether a machine-readable reservation of rights existed at the time of use, or when a party selling a capability needs to show the gap it closes is real.

Findings

The measurements are the point, and they are readable without calling anything.

When did websites start blocking AI crawlers? — 21 dated epochs, January 2019 to September 2026. 0.0% through January 2021 (95% CI 0.0–4.9%, n=74), holding across five consecutive annual epochs; first movement October 2023 at 3.8%; 14.1% by September 2026. Sample, not census, and the method is written down so it can be recomputed against us.
What share of sites serve an llms.txt? — 6 published answers to one question, from 5.86% to 47.49%, a spread of 8.1×, and none of it is error. Three parameters decide the number: which population was sampled, whether an HTTP 200 was validated as a real file, and how unreachable hosts are counted. Every measurer states its own; the fault is in quoting a single figure as the rate.

Where a dispute turns on whether a reservation of rights was machine-readable at the time of use, the test is the state of the art at that past date. The policy literature names the standards and publishes no adoption figures. Of the capability classes measured here, only the robots.txt AI stance has a series reaching back before 2026 — and we say so rather than implying the rest do.

This site is also the worked example

An index of missing machine-readable surfaces should have one itself. Everything below is live, and every link is a link — not prose describing one.

Call it

There is a real MCP server at /mcp, JSON-RPC 2.0 over HTTP, protocol 2025-06-18. A GET returns 405 with a JSON-RPC envelope, which is what live servers do; post to it instead.

curl -s https://atlas.geosensor.tech/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

absence_rate — measured absence per class, with its denominator and the sample it came from. Open, no credential.
adoption_curve — the dated series itself: 21 single-vantage epochs with confidence intervals. Open, no credential. No arguments gives the trend; epoch gives one point, series: true gives all of them, and the caveats come back every time.
host_report — needs a bearer token. Findings about named third parties are not published, so this returns that refusal rather than data.
bulk_export — needs a token and payment. A receive-only address is configured, so the challenge names a real recipient, and a payment is verified and settled through a facilitator before anything is served.

The gate is meant to be walkable

Of the gated endpoints we measured, 15.3% were shut to any client that had not met them before: refused, with nothing machine-readable saying where to authenticate. This one says. Calling host_report without a credential returns 401 with WWW-Authenticate: … resource_metadata=… — the pointer RFC 9728 requires and the MCP authorization spec makes a MUST — which leads to the resource metadata, then to the authorization server, then to a token:

curl -s -X POST https://atlas.geosensor.tech/oauth/token \
  -d 'grant_type=client_credentials&client_id=you'

Issuance is open: the advertised endpoint actually issues.

Documents

/free/curve — the dated adoption series itself: 21 single-vantage epochs with confidence intervals and its caveats attached. Free, no credential, no payment.
/llms.txt — orientation, in prose
/offer.html — what is sold, priced, with the terms
/.well-known/agent-card.json — A2A agent card
/openapi.json — interface description
/.well-known/x402 — payment discovery; the manifest names the network it settles on
/.well-known/tdmrep.json — TDM reservation policy
/.well-known/security.txt — RFC 9116 contact
/robots.txt and /sitemap.xml — AI crawlers named, and a sitemap they can find

Status

Research in progress. The machine-payable routes work and stay up: a payment is verified and then settled on-chain before anything is served, there are no freebies beyond the published free sample, and a subscription can be bought by software at /paid/subscribe without a human in the loop, at the rates the offer page lists. They are a demonstration that an agent can buy data here unattended, and they are not what this operation sells.

TESTNET only (Base Sepolia): settlement is on-chain but moves test tokens with no monetary value. That sentence is generated from this service's own configuration, so it cannot go stale.

What this site observes is published in aggregate only: it makes no claims about an individual named company's own site or conduct. Measurements that others have published themselves are public claims, and are cited and examined by name where that is the subject.