Lighthouse 13.5, released September 18 according to its GitHub release notes, ships a new audit called ard-schema for Agentic Resource Discovery (ARD). It sits in the experimental “Agentic Browsing” category, grouped with the existing llms.txt check, and on its face it reads as Lighthouse telling you whether a site is ready for AI agents. The audit does something narrower: it looks for the manifest layout the ARD spec used before its current revision, so a site built to the spec published August 26 can score “Not Applicable,” the same result Lighthouse gives a site with no agent manifest at all.
What Does Lighthouse’s ARD Audit Check?
The ard-schema audit looks for a machine-readable catalog of a site’s callable resources (MCP tools, A2A agents, skills) in four places, in this order, and validates whatever it finds against a schema:
Agentmap:directive inrobots.txt<link rel="ai-catalog">in the page head- an HTTP
Linkheader carryingrel="ai-catalog" - the fallback path
/.well-known/ai-catalog.json
The outcome depends on what it finds. With no Agentmap line, link tag or header, and no file at the fallback path, the audit marks itself Not Applicable rather than failing. If one of those signals exists but points at a file that won’t load, the audit scores 0 with the note “Catalog file could not be loaded for schema validation.” The header check runs only in navigation mode.
A Spec That Moved Before the Audit Shipped
The ARD specification‘s own text, dated August 26 and labeled v0.91, tells consumers to fetch /.well-known/ard.json and to honor a rel="ard" link, and says publishers only need to serve ard.json and emit that link. It treats the older ai-catalog.json path and the ai-catalog link relation as optional for consumers to check, calling them “a courtesy to resources published before this revision.” The Lighthouse 13.5 source, read at the tagged release, never mentions ard.json or rel="ard" anywhere, a gap Search Engine Journal also flagged when it covered the release. The audit’s own UI strings still name the predecessor file: a passing result reads “ai-catalog.json schema is valid.” A site that did exactly what the current spec asks for, and nothing else, is invisible to the check.
The validator’s supporting library points at the same gap from a different angle. Its pinned upstream commit is dated September 12, more than two weeks after the spec moved to v0.91, and that upstream repository now carries both an ai-catalog.schema.json file and an ard-entry.schema.json file. Lighthouse’s audit is wired to the older of the two.
Where Agentmap Fits
Nothing in the source rules out a hybrid approach, and the gatherer’s own check order suggests one. Agentmap in robots.txt is checked first, and the code takes whatever URL follows that directive without matching the filename against a fixed list. A robots.txt line pointing at an ard.json file would still be fetched and validated today. That reading comes from the gatherer’s logic, not from guidance either the Lighthouse or the ARD project has published, and it only applies to sites that add the directive: the rel="ard" link tag and the bare well-known path still go unchecked. The spec itself lists Agentmap as a discovery mechanism, but its binding consumer rules require only ard.json and rel="ard", so the directive is an addition to those, not a replacement.
No Score, No Search Tie-In, Not Yet in PageSpeed
Agentic Browsing carries no 0-100 weighted score the way Performance or SEO do. Chrome’s own scoring documentation for the category says it reports a fractional pass ratio instead, because “the standards for the agentic web are still emerging.” The release notes don’t connect the audit to Google Search in any way, and the category is explicitly experimental. The Lighthouse team says it expects the release to ship in Chrome 156 DevTools and in PageSpeed Insights within two weeks. Anyone reading a client report with a fresh “Agentic Browsing” line should treat a Not Applicable result as a statement about which draft the tool implements, not necessarily about the site.
Chrome tooling has shipped with visible limits before: the four experimental ad metrics added to CrUX came without a JavaScript API for RUM tools. Robots.txt, meanwhile, keeps collecting agent-facing signals, as with Cloudflare’s bot preference sync. For ARD, an Agentmap line is the one current-spec mechanism Lighthouse 13.5 also checks.