Methodology

How agenticweb.wiki is built, what counts as a row, how state is assessed, and where the gaps are.

The nine steps

The map is organised around nine steps an agent and a site pass through, each with a classic-web analog.

How records are produced

Records are drafted by AI research agents that read each primary source — spec text, datatracker page, or vendor documentation — under a fixed brief published at /methodology/verify-brief/. Every evidence label is assigned per source, not per record: a claim is marked VERIFIED only when an agent read a primary source or two independent secondaries itself. External URLs in the dataset are checked on a recurring schedule, with HTTP status codes recorded in data/linkcheck.json. An adversarial review re-fetches a sample of sources before publish and corrects or downgrades anything it cannot confirm.

Inclusion

A row needs a public specification text or a fixed artifact; products without a spec are not rows.

Anyone can ask to be listed, at /submit/. An origin instance (a domain publishing one of the standards here) is verified by running our own probe rule against the origin, so it is listed as measured; a harness, venue, authority, or agent instance is verified by fetching its primary specification, repository, or documentation source; a resource (SDK, validator, registry, scanner) is verified by fetching it; a new standard goes through the same brief as every other row.

Facet vocabularies

Parties

Form

Track

Scope

Phase

Implementation

Adoption

Evidence

Freshness

Each row is re-checked on a 30-day cadence. Freshness state:

Blanks

No step has three or fewer standards as of 2026-10-30. A step is marked thin only while it has three or fewer records; the count on each category page is the live total.

Named gaps: no published standard was found for agent attention or ranking signals, for agent-side audit logs beyond attestation proposals, or for agent-to-human consent receipts. These are recorded as blanks, not omissions.

What the occupants list

Checked 2026-10-25. For each reference-map occupant we track, rows we have that they do not, and rows they have that we do not.

agentswelcome

lumar

agent-ready

What answer engines read

As of October 2026, no answer engine — Google AI Overviews or AI Mode, ChatGPT Search, Perplexity, Claude, or Bing Copilot — documents consuming llms.txt, ARD catalogs, Agentmap, AIPREF, or TDMRep. Google's own AI optimisation guide states that no special markup or llms.txt is required for a page to appear in AI features. This site serves those files because it maps them, not because they affect ranking. Source: developers.google.com/search/docs/fundamentals/ai-optimization-guide.

How this stays fresh

Watchers run every six hours against package registries, repositories and feeds (github, npm, pypi, feed) and daily against specification texts and standards bodies (ietf, w3c, page, erc). Changes they find by machine are published automatically with the evidence attached, and records drafted this way carry a visible Candidate badge until a person has reviewed them. Each morning the engine also re-probes the /check/ readiness panel and publishes the daily digest with cited primary sources and mechanically verified numbers, alongside a weekly review digest. The data lives in the public repository github.com/Belgravia-Labs/agenticweb-data, where anyone can open an issue or pull request. Every automated update is listed on /changes/.

Probes

Where a standard defines a checkable request, the protocol row carries an executable probe: method, path, what to parse, and what to record when the path is absent. Adoption counts on this site are marked measured only when we have run that probe ourselves; otherwise adoption is claimed, observed, or unmeasured.