Google Chrome has added a category to Lighthouse, the audit tool built into every Chrome browser, called Agentic Browsing. It scores how ready a site is for AI agents to read and operate. It arrived with little announcement, it is marked experimental, and it is already running against your site whenever anyone audits you. If you have optimised for Lighthouse for a decade, this is the first genuinely new thing it has measured in years, and it is worth understanding precisely, because it behaves unlike every other category, and because passing it means both less and more than it appears.
It is not scored like the rest of Lighthouse, and that tells you something
Performance, Accessibility, Best Practices and SEO each give you a number from 0 to 100. Agentic Browsing does not. It reports a fraction, how many readiness checks your site passes, with no weighted score and no ranking. Google has been explicit about why: the standards for the agentic web are still forming, so the category exists to gather data and surface signals rather than to rank you. The header shows a pass ratio so you can watch progress at a glance.
Read the intent honestly. This is Google’s Chrome team building observability for a web that does not fully exist yet. It is not a ranking factor, it is not a promise, and a green result is not a citation. What it is: the first standardised, deterministic, reproducible measure of whether an AI agent can mechanically read and act on your pages. That alone makes it worth passing, and worth not over-reading.
What it actually checks
The audits are deterministic, which means they are reproducible and safe to wire into a CI/CD pipeline. They fall into three groups.
WebMCP integration. Lighthouse watches, through the Chrome DevTools Protocol, whether your site registers tools an agent can call, both declarative tools defined in HTML and imperative ones defined in JavaScript. This only applies if your site exposes tools, and the relevant checks require a recent Chrome build with the experimental WebMCP origin trial enabled. If you do not expose tools, the check registers as not-applicable rather than a failure. A content site that does not run an MCP server is not penalised for it, and should not claim a WebMCP pass it has not earned. The practical how-to for exposing a form to agents is covered in our WebMCP guide.
Agent-centric accessibility. Agents rely on the accessibility tree as their primary data model, it is, literally, the machine-eye view of your page. Lighthouse pulls the subset of accessibility audits that matter most for machine interaction: every interactive element has a programmatic name, roles and parent-child relationships are valid, and nothing is hidden from the accessibility tree while still being interactive. This is why your Accessibility score and your Agentic Browsing result move together. Clean semantic HTML and correct ARIA are not an ethical nicety here; they are the data the agent reads.
Stability and discoverability. Two checks. Cumulative Layout Shift, because an agent that identifies an element and then acts on it needs that element to stay where it was, layout shift from ads, unsized images or injected content moves the target. And the presence of an llms.txt file at your domain root: a machine-readable summary of your most important content. One important mechanic here: a missing llms.txt is currently marked not-applicable because the file is optional, so it drops out of the denominator rather than counting against you, while a malformed file can be flagged as a fail. That is why a clean 2/2 can sit ahead of a 2/3: more denominator is more surface being graded, not more won.
The contradiction it exposes, and how to read it
Google’s own Search team has said you do not need llms.txt or similar files to appear in its AI search, and that Search ignores them. Google’s Chrome team then shipped an audit that checks for one. Same company, opposite-looking signals.
They only conflict if you treat “AI” as one thing. They are measuring different floors. Search is talking about ranking, and it is right, the file will not change your Google position. Chrome is talking about agent operation, whether an autonomous agent can efficiently parse and act on your site, and there the file helps. The Lighthouse documentation even frames it as an efficiency hypothesis: without a machine-readable summary, agents may spend more time crawling the site. That is an operation claim, not a ranking claim, and not evidence that Google’s own products read the file. The full resolution sits in our Four-Floor Model: the audit measures the fourth floor, agent execution. It cannot see the floors below it, which is where citation and recommendation are actually decided.
So ship the llms.txt, because it is cheap, it passes this audit, and the surfaces that do read it, Perplexity, coding agents, exist. Just do not let anyone sell you the audit pass as a citation lever. It is not one. The reason it cannot be is that the file is a self-declaration, and an AI system cannot use what a site says about itself to differentiate it from every other site that says the same thing. What differentiates a site is corroboration, not self-description, the mechanism we cover in how AI learns your brand.
Why your results fluctuate
The audits are deterministic, but your result can still move between runs, and the documentation names the causes. If your tools register through JavaScript (the imperative WebMCP API), the timing of registration can affect whether Lighthouse captures them in its snapshot. Significant changes to DOM size or complexity reshape the accessibility tree. And layout shift can move elements between the moment an agent identifies one and the moment it acts. In practice: re-run before you cite a score, and treat a single pass as a reading, not a guarantee.
What a passing score is worth
Necessary, not sufficient. Passing means an agent can read and operate your site cleanly, the mechanical precondition for being used by AI at all. It says nothing about whether an AI will choose to cite or recommend you, because that is decided by content quality and, above all, by off-page authority the audit cannot measure. Treat it as the floor of the building, audited; not the building. The work that turns “readable” into “recommended” is corroboration, and it is covered in the best-of problem and our work on editorial selection. You can see a worked, verifiable pass on our own agent-readiness proof page.
A falsification note, in the spirit of treating this as a probe rather than gospel: if, over the next several quarters, AI assistants demonstrably begin grounding answers in llms.txt at query time for non-coding queries, the “agent operation only” reading of this audit weakens and the file moves closer to a genuine retrieval signal. We will update this guide when the evidence moves. As of this writing, that grounding is not observable.
