Lighthouse now audits agentic browsing. What it checks, and what it says about AI answers

Lighthouse 13.5 shipped on 18 September 2026 with one new audit, and it lands in the one Lighthouse category built for a reader that is neither a person nor a search crawler.

It is called Agent Resource Discovery.

The release notes on GitHub list it in one line, next to another that groups it with the existing llms.txt audit "under agent discovery." They also say the release should reach Chrome 156's DevTools and PageSpeed Insights within two weeks.

So anyone who runs PageSpeed Insights in the next few days may meet a category they have never seen, with results that are hard to read.

Here is the complication. The new audit checks a file name that the standard it is based on has already moved away from.

I read the release, Chrome's documentation and the specification on 30 September. This is what each part checks, what the results mean, and what they do not.

Part 1: what the Agentic Browsing category is

Chrome's scoring page opens with a warning: the category and its WebMCP support are experimental and based on proposed standards.

It needs Chrome 150 or later, per the same page. The WebMCP audits also need you to register for an origin trial.

The documentation groups the audits under four headings: WebMCP integration, discoverability, accessibility for agents, and layout stability.

Read those as questions. Can an agent find the actions your site offers? Can it find a summary of what your site is? Can it read the page structure? And will the button still be where it was when the agent clicks it?

Those are fair questions for a site an agent might book, buy or fill a form on.

They are different questions from the ones GEO asks.

Part 2: why there is no score

Every other Lighthouse category gives you a number out of 100.

This one does not.

Chrome's scoring page says that because the standards for the agentic web are still emerging, the category shows a fractional score instead: a ratio of how many checks your site passes, plus pass or fail on specific audits.

That is Google telling you not to optimise for it yet.

Any tool, agency or dashboard that turns this ratio into an "AI readiness score" and puts a target on it is inventing precision the source refuses to give.

Part 3: the llms.txt audit is gentler than it looks

The llms.txt audit page is short and worth reading in full.

If your server throws an error when Lighthouse fetches /llms.txt, the audit flags it.

If the file does not exist and returns a 404, the audit is marked Not Applicable, because the page says providing the file is optional for now.

So a missing llms.txt does not fail you.

A broken one does. If your host returns a 500 or a login page for that path, fix the server response. Whether to publish the file at all is a separate decision.

We covered that decision in does llms.txt do anything, and the evidence there has not moved: Google's own guidance says Google Search does not use it. A Lighthouse audit that tolerates its absence is consistent with that.

Part 4: the new ARD audit, and the file-name problem

Agent Resource Discovery is a proposed standard for listing the callable things a site offers to AI agents.

The specification, version 0.91 dated 26 August 2026 and marked as a proposal, names authors from Google, Microsoft and Hugging Face. It describes entries for MCP tools, A2A agents and similar services, so that an agent can search for a capability instead of having every tool installed in advance.

Search Engine Journal's 21 September report by Matt G. Southern describes how the audit looks for a catalog. It checks robots.txt for an Agentmap line, then an HTML link tag with rel="ai-catalog", then an HTTP Link header, and falls back to /.well-known/ai-catalog.json.

Now read section 5.1 of the spec.

It says publishers publish at /.well-known/ard.json with rel="ard". It calls ai-catalog.json the predecessor path, says consumers may check it as a courtesy, and says a publisher on the old path should move.

So a site that follows the current spec to the letter could be invisible to the audit, according to SEJ's description of it, while a site on the old name passes.

I would not rush to publish both. I would wait for the audit to catch up with the spec, which Lighthouse's own release notes suggest is being tracked: they list a weekly check that syncs the ARD conformance script.

Standards this young will move again, and so will this post. Siddharth Khushalani wrote about the cost of that on the operator side in why every post is a promise to keep it true.

Part 5: Not Applicable is the right answer for most businesses

This is the part the category's arrival will get wrong in sales decks.

An ARD catalog lists tools an agent can call. A booking endpoint. A quote calculator exposed over MCP. An inventory lookup.

A dental clinic with a contact form has nothing to list. Neither does a café with a menu page, or a consultancy with a services page.

For those sites, Not Applicable is the correct result. It is not a gap.

NovaTechRay's own site returned a 404 for /.well-known/ard.json when I checked on 30 September, and that is right for it: it offers no callable tools. Its llms.txt returns a normal 200.

If someone quotes you to "fix" an N/A on this audit without first asking what tool your site would expose, they are selling the audit, not the outcome.

Part 6: what none of this measures

Being cited.

Nothing in the release notes or Chrome's documentation connects these audits to whether ChatGPT, Claude, Perplexity or Google's AI features name your business in an answer. SEJ's report says the same about Google Search.

Citation still runs on the things covered across this blog: whether answer engines can crawl you, whether your pages state facts plainly, and whether other sources talk about you. The agentic audits are about the next step, an agent acting on your site after it has already chosen you.

That step is real and it is coming. It is also a different job.

Part 7: what to do this week

Run PageSpeed Insights on your homepage once the category appears.

Read any Fail line first. Those are server errors and schema errors, and they are worth fixing because they are broken, not because of a ratio.

Leave Not Applicable alone unless you have a real tool to publish.

If you do run bookings or quotes through your site, the accessibility and layout-stability audits are where to spend time. An agent that cannot find the button fails the booking the same way a person with a screen reader does.

Then go back to the work that gets you named. This category grades whether an agent can use your site. It does not grade whether one will be sent there.

The short version

• Lighthouse 13.5, released 18 September 2026, added an Agent Resource Discovery audit and grouped it with llms.txt.
• The Agentic Browsing category is experimental and gives a pass ratio, not a score.
• A missing llms.txt is Not Applicable. A broken one is flagged.
• The ARD audit checks the older ai-catalog name; the spec now says ard.json.
• Not Applicable is correct for any site with no callable tools.
• None of these audits measures citation in AI answers.
• Fix Fail lines, ignore the ratio, and keep working on being named.

We read the source before we write the advice at novatechray.com.