Lighthouse's AI catalog audit rewards publishing nothing

Lighthouse is Google’s open-source tool for auditing web pages. Version 13.5 added a check that grades an agent catalog, a JSON file listing what a site offers to AI agents.¹ I ran it against this site in four states. With no catalog at all, its Agentic Browsing category read 4 of 4. With the catalog a content site would reasonably write first, listing its llms.txt file and its RSS (web syndication) feed, the category read 4 of 5. So the audit rewards a site for publishing nothing and punishes the honest first attempt. Google expects the check to reach PageSpeed Insights within two weeks of the September 18 release, which puts that lost point in reports people act on.¹

Lighthouse 13.5.0 scored this site 4 of 4 on Agentic Browsing with no agent catalog, and 4 of 5 with the catalog a content site would write first.

What does Lighthouse’s AI catalog audit check?

Lighthouse’s AI catalog audit, named ard-schema in the tool, finds a site’s agent catalog and validates it against the Agentic Resource Discovery specification.² That specification is a proposal, written by authors at Google, Microsoft and Hugging Face, for describing tools and agents so that other agents can find them.³ The audit reads the first catalog it can locate and reports every schema error and warning in it.

Lighthouse looks for a catalog in four places, in this order:⁴

1. An Agentmap: line in robots.txt
2. A <link rel="ai-catalog"> tag in the page head
3. An ai-catalog entry in the HTTP Link header
4. The default address, /.well-known/ai-catalog.json

The first one that resolves is the file it fetches. A catalog that loads with any error scores 0, one with warnings only scores 0.9, and a clean one scores 1.²

The audit belongs to a category called Agentic Browsing, which holds seven equally weighted audits.⁵ Lighthouse describes the whole category as “still under development and subject to change.”⁵ It reports a fraction, the audits passed over the audits that applied, instead of a score out of 100.⁶

Three of the seven cover WebMCP, a draft browser interface that lets a page hand callable functions to an agent. One checks the accessibility tree, the structure an agent navigates instead of pixels. One is Cumulative Layout Shift, the Core Web Vitals measure of how much a page jumps while it loads. Two sit under an Agent Discoverability heading, the catalog audit and an llms.txt check.

That llms.txt check looks for a heading, at least one link and at least 50 characters,⁷ which is a shape test with the limits I covered in the piece on my own llms.txt file.

Why does publishing no catalog score better than a first attempt?

Lighthouse’s catalog audit marks a site with no catalog as not applicable, and a not-applicable audit is removed from the Agentic Browsing count. The source is explicit. When no catalog loads and nothing points to one, the audit returns notApplicable: true,² and the scoring code sets that audit’s weight to zero, so it “will not count.”⁸ A catalog with one error scores 0, and that zero counts.

I ran Lighthouse 13.5.0 against a local build of this site on September 22, once per catalog state, changing nothing else. Each run had Chrome’s WebMCP feature switched on with the --enable-features=WebMCP flag, because this site registers WebMCP tools only when the browser exposes that interface.

A default run reads one audit lower in every state, 3 of 3 with no catalog, while the catalog audit’s own result does not change. The four other applicable audits passed every time, so the catalog audit alone moved the fraction. The live homepage returned the same 4 of 4 as the local run with no catalog.

What the site served	Catalog audit	Agentic Browsing
No catalog, and nothing pointing to one	Not applicable	4 of 4
A catalog listing llms.txt and the RSS feed	0	4 of 5
A catalog listing only the site’s two agent skills	1	5 of 5
A robots.txt line pointing at a catalog that returns 404	0	4 of 5

A site owner who sees the new heading and writes the obvious file loses a point that a site owner who ignored the heading keeps. Only a flawless catalog matches doing nothing, and a pointer to a missing file costs the same point as a broken file.

I have already argued that a check validating structure cannot see a false value, in the piece on agent-readiness scanners. This audit adds a different problem, a scoring rule under which silence is the safe choice.

Why can’t a content site’s first catalog pass?

A content site’s first agent catalog fails on two structural errors, and even once those are fixed it stays mislabeled. The labels are the deeper problem, because the specification accepts nine media types, and every one describes an agent, a tool server, an agent skill or a registry of those.

None describes an article, a feed or a text file.⁹ The two files a publisher most wants machines to find, llms.txt and the RSS feed, have no correct label, and Lighthouse reports that only as a warning.

Lighthouse reported two errors on it. The feed’s identifier, brandonlazovic.dev/pulse, did not match the required formal name, urn:air:<publisher>:<namespace>:<agent-name>. The llms.txt entry carried one example query, where the schema requires two to five. Four low-severity warnings followed, two of them saying that text/markdown and application/rss+xml are “not one of standard discovery types.”⁹

Fixing only the two errors flips the result. I gave the same two files formal identifiers and two example queries each, and the catalog audit scored 0.9. The category read 5 of 5, and the audit’s title changed to “ai-catalog.json schema is valid,” with both type warnings still listed.

Lighthouse counts any audit scoring 0.9 or more as passed when it builds the fraction,⁶ so warnings never reach the headline number. The category’s underlying score did fall, from 1 with a clean catalog to 0.98.

My read: a 0.9 for those two files is a pass for formatting. The catalog now tells agents about two resources the specification does not recognize, and the report calls it valid.

Which catalog file does Lighthouse check, and which does the specification name?

Lighthouse 13.5 fetches /.well-known/ai-catalog.json and follows the ai-catalog link relation, but the specification’s current revision names a different file. Version 0.91 of the Agentic Resource Discovery specification, dated August 26, 2026, says a consumer “MUST fetch /.well-known/ard.json” and calls ai-catalog.json the path of “ARD’s predecessor.”³ Lighthouse’s discovery code never requests ard.json.⁴

The specification also addresses publishers directly: “a publisher on the predecessor path SHOULD move to ard.json.”³ A site that follows that advice to the letter is invisible to Lighthouse 13.5, which then finds no catalog and marks the audit not applicable. The most faithful catalog and no catalog at all produce the same report.

The validator this site keeps for its own unpublished catalog follows Lighthouse’s path, because Lighthouse is the tool people will run. Before anything ships here I would check both names again, since the specification is a proposal and both documents are moving.

Should a content site publish an agent catalog for Lighthouse?

A content site should publish an agent catalog only when it has something of an accepted type to list. Those types cover agent skills, agent cards that describe an AI agent, and server cards that describe a tool server, such as a Model Context Protocol server exposing tools to an AI model.

With nothing like that, the correct catalog is none, and Lighthouse 13.5 agrees by leaving the audit out. Listing llms.txt and a feed costs a point first and earns a hollow pass after two fixes.

This site has exactly two things that fit. It already publishes two agent skills under /.well-known/agent-skills/, one for querying the Pulse corpus and one for reading articles as markdown. A catalog listing only those two scored 1, and the category read 5 of 5. It stays unpublished for now. My read: a clean score alone does not justify a second hand-kept list that repeats the skills index the site already serves.

For a site weighing the same decision, four checks settle it:

1. Compare what you would list against the nine media types Lighthouse prints in its warning text. If nothing matches, stop there.
2. Validate the file before it ships. One error scores 0, and Lighthouse keeps its conformance tester in its own repository.⁹
3. Add an Agentmap: line, a link tag or a Link header only after the file it names returns a 200 status. In my run, a pointer to a missing file scored 0 with “Catalog file could not be loaded for schema validation.”²
4. Choose the file name on purpose. Lighthouse 13.5 reads ai-catalog.json, and the specification’s consumers are required to read ard.json.