Opens in a new tabSkip to content
Agent LighthouseAgent Lighthouse

    Searches the text of every published page. The evidence sources themselves are not in this index — search all of them on the trusted sources page.

    GitHub ↗
    Browse checks and page contents
    agent-interfaces/ai-catalog-metadata

    AI Catalog complete metadata

    What it checks

    Complete metadata helps AI agents understand who owns the service, when it was last updated, and what it can do. Missing fields reduce agent confidence in your services and may cause them to skip your site in favor of better-documented alternatives.

    Why it matters

    A discovery client picks which catalog entry answers a query by matching text drawn from named entry fields. An entry that omits those fields is therefore present in the manifest and invisible to the query. The manifest lists it, and no search surfaces it.

    Evidence

    ARD entry metadata read by a discovery client

    • Hugging Face ships hf-discover, an ARD-compliant client whose navigate mode performs “automatic .well-known/ai-catalog.json discovery from a website” and follows federated registries — github.com/…/hf-discover (verified 2026-08-24)
    • Its navigation code builds the text it matches a query against from five entry fields: displayName, description, tags, capabilities and representativeQueries. displayName is already mandatory under ARD §4.2, so the four this audit scores are exactly the optional keys that decide whether a query surfaces an entry at all.
    • ARD §4.1 makes specVersion, host and entries the required top-level fields, and defines version, updatedAt, tags, metadata and trustManifest as optional enrichment; identity is expressed through host.identifier (a DID) and the optional trustManifest — github.com/…/ard-spec (verified 2026-08-24)
    • ARD is a Linux Foundation working-group specification with Google, Microsoft and Hugging Face among its contributors, published 2026-06-17 under Apache 2.0 — developers.googleblog.com/announcing-the-agentic-re… (verified 2026-08-24)

    Limits

    No crawler is documented to downrank a site for thin catalog metadata. The consequence is mechanical — the client matches less text — rather than a published ranking signal. The hosted Hugging Face server does not fetch arbitrary well-known files at all: “Navigation is intentionally not exposed by the hosted server”. The consuming path is therefore a user-driven CLI, not a background crawler. The specification is also a draft and says so.

    Historically this audit scored an invented field list: owner, contact, lastUpdated and services. None of those appears in any revision of the spec, or in any of the four live manifests checked — the ARD conformance example, neon.com, weaviate.io and the Shopware core template. A spec-perfect manifest therefore scored zero until the 2026-08-22 rewrite.

    How it scores

    The consuming code is first-party, public and readable, and the fields it matches on are exactly the ones this audit scores. It is not grade A because ARD is a draft (v0.9) rather than a ratified standard, and because the behaviour is documented in a client’s source rather than in a vendor statement about a hosted crawler.

    Sources