TL;DR
- Lighthouse 13.5, released on 18 September 2026, adds an ard-schema audit that validates Agentic Resource Discovery (ARD) catalogs.
- Lighthouse checks robots.txt Agentmap, a link tag, an HTTP Link header, then /.well-known/ai-catalog.json; with nothing found it reports Not Applicable.
- ARD spec v0.91 (26 August 2026) moved the file to /.well-known/ard.json, which SEJ found Lighthouse 13.5 does not yet check.
- The audit sits in the experimental Agentic Browsing category, not SEO, and the release does not link it to Google Search.
Google's Lighthouse 13.5, released on 18 September 2026, adds an audit that checks whether a website publishes a valid catalog for Agentic Resource Discovery (ARD), a proposed standard for how AI agents find the tools and services an organisation offers. The Lighthouse team expects the release to reach Chrome DevTools in Chrome 156 and PageSpeed Insights within two weeks.
The audit sits in Lighthouse's experimental Agentic Browsing category, next to the existing llms.txt check, and is not part of the SEO category. Nothing in the release ties it to Google Search rankings. It is still the first time a mainstream Google site-quality tool has scored a site against an agent-discovery spec, and it will start appearing in reports many site owners already run.
What shipped in Lighthouse 13.5
The Lighthouse 13.5.0 release notes on GitHub list one new audit, "ard-schema: add Agent Resource Discovery audit". Two related changes ship alongside it. One groups the llms.txt and ARD audits together under an agent discovery heading. The other two adjust the WebMCP audits: the form coverage audit now passes instead of showing "not applicable" when every form is covered, and the registered tools audit warns when a page exceeds a recommended number of tools.
The release note states the expected rollout plainly: "We expect this release to ship in the DevTools of Chrome 156, and to PageSpeed Insights within 2 weeks." For most marketers, PageSpeed Insights is where they will first see the new check.
How the audit looks for a catalog
Search Engine Journal read the 13.5 source code and reported the order in which Lighthouse looks for an ARD catalog:
- An Agentmap line in robots.txt, a directive from the ARD spec that points to a catalog.
- A link tag in the page with the ai-catalog relation.
- An HTTP Link header with the same relation.
- If none of those exist, a request to /.well-known/ai-catalog.json.
If there is no pointer and nothing is served at the well-known path, the audit reports Not Applicable. It fails when the catalog does not match the ARD schema, or when a pointer leads to a catalog that cannot be loaded. The pull request describes the check as a "schema conformance audit", and Lighthouse built its validator from the ARD project's own conformance tests.
Google's Lighthouse documentation says the Agentic Browsing category does not produce a 0 to 100 score. It shows a pass ratio instead, because the standards for the agentic web are still emerging.
The file name has already moved
There is a mismatch between the audit and the current spec. The ARD specification page lists version 0.91, dated 26 August 2026, with status "Proposal". According to Search Engine Journal, v0.91 names /.well-known/ard.json as the manifest location and keeps ai-catalog.json only as an earlier path that readers may also check. The spec warns that a file available only at the old path "may not be found".
SEJ reported finding no reference to ard.json, or to the newer ard link relation, in the 13.5 source or on the project's main branch as of 21 September. It also noted that the spec describes discovery through in-page JSON-LD and DNS records, which 13.5 does not check.
The practical result is awkward for early adopters. A site that follows the latest spec, serves /.well-known/ard.json and uses the new relation, but has no Agentmap line and no ai-catalog file, would be reported as Not Applicable by Lighthouse 13.5. That does not mean the site is missing a manifest. It means Lighthouse is looking in the places the spec used earlier.
ARD, llms.txt and WebMCP compared
Three agent-facing formats now have Lighthouse audits, and they do different jobs. The table uses the descriptions from the ARD spec, the Lighthouse release and SEJ's reporting.
| Format | What it describes | Lighthouse check |
|---|---|---|
| ARD catalog | Callable resources: MCP tools, A2A agents, skills and other services | New in 13.5 (ard-schema) |
| llms.txt | A summary of the site's content for language models | Existing audit, now grouped with ARD |
| WebMCP | Structured actions a page exposes to an agent already on it | Existing audits, adjusted in 13.5 |
What ARD is trying to solve
The ARD spec credits three authors: Junjie Bu of Google, R.V. Guha of Microsoft and Shaun Smith of Hugging Face. Its motivation section describes a problem with how AI systems pick tools today. Most language models select tools by loading every tool description into the context window, which does not scale once there are thousands of agents and services. The spec proposes moving discovery out of the model and into a dedicated search service, where signals such as representative queries, publisher identity and compliance metadata can be used without spending context tokens.
To make that work, a resource is described once, on the publisher's own domain, and registries index it. The spec says registries must expose a standard HTTP REST search interface so any system can take part. Version 0.91 restates entries as JSON-LD nodes, which lets a publisher add terms from other vocabularies without waiting for the spec to change.
In plain terms, ARD is closer to a sitemap for callable services than to a content file. It matters to organisations that actually offer something an agent can call, such as a booking API, a product search tool or an MCP server. A brochure site with no callable services has little to put in a catalog.
Who should act now, and who can wait
Most business websites in Thailand do not publish MCP servers or agent endpoints, and for them the new audit will show Not Applicable. That is a neutral result, not a failure, and there is nothing in the release to suggest it affects search visibility.
Businesses that do expose services to agents, such as travel and hotel booking engines, e-commerce catalogs with product APIs, or SaaS tools with an MCP server, have a reason to look closer. For them, the questions are practical:
- Do you publish any agent-callable resource today? If not, there is nothing to catalog yet.
- If you do, which spec version are you following? If you publish only /.well-known/ard.json, Lighthouse 13.5 may not see it. Serving the older ai-catalog.json path as well, or adding an Agentmap line to robots.txt, covers both readers while the tooling catches up.
- Does your catalog validate? A schema failure is reported as a failed audit, which is more visible than Not Applicable in a client-facing PageSpeed report.
- Is your robots.txt managed by a CMS or plugin? An Agentmap line has to survive plugin updates.
What the release does not say
The release does not connect the audit to Google Search, AI Overviews, AI Mode or Gemini. It does not say whether Google's own agents read ARD catalogs, or when Lighthouse will follow the spec's move to ard.json. The ARD spec is a proposal, not a ratified standard, and its authors warn that some media types it uses are still pending formal registration. Treat the audit as a signal that Google's tooling team is watching agent discovery, not as a ranking factor.
What this means for Thai marketers
For most Thai brands, the useful takeaway is about reporting rather than action. When a new row appears in a PageSpeed Insights or Lighthouse report under agent discoverability, it will usually read Not Applicable, and that is fine. It is worth telling clients and stakeholders this before they see it, so a new line in a familiar report does not turn into a false alarm.
For brands whose visibility in AI answers matters, the bigger work is still content that AI systems can cite and trust, which is what a GEO programme focuses on. Agent catalogs become relevant only once a business has services worth calling. A technical SEO audit can include robots.txt hygiene and a check of which agent-facing files a site already serves, so nothing conflicts when the time comes.
FAQ
What is the new Lighthouse ARD audit?
It is a schema check for Agentic Resource Discovery catalogs, added in Lighthouse 13.5 on 18 September 2026. It looks for a catalog pointer or a file at /.well-known/ai-catalog.json and validates it against the ARD schema.
Does the ARD audit affect Google rankings?
Nothing in the release says so. The audit sits in Lighthouse's experimental Agentic Browsing category, separate from the SEO audits, and the release notes do not mention Google Search.
My site shows Not Applicable. Is that a problem?
No. Not Applicable means Lighthouse found no ARD catalog where it looks. For a site with no agent-callable services, that is the expected result.
Should I publish ard.json or ai-catalog.json?
The current spec, v0.91, uses /.well-known/ard.json, but Lighthouse 13.5 looks for ai-catalog.json and an Agentmap line, according to SEJ's reading of the code. A site that wants both to find its catalog can serve both paths until the tooling is updated.
When will this appear in PageSpeed Insights?
The Lighthouse team said it expects 13.5 to reach PageSpeed Insights within two weeks of the 18 September release and to ship in Chrome 156 DevTools.
Lighthouse 13.5 does not change what makes a site rank. It adds a check for an agent discovery format that is still a draft, and for now it looks in the older location. Sites with services worth exposing should publish at both paths and validate; everyone else can note the new row and move on.
