Guides › Google's Agentic Browsing Audit, Explained

Google's Agentic Browsing Audit, Explained

By Michael Emery · reviewed against the measurement record · 30 July 2026

Google's PageSpeed Insights and Lighthouse now include an "Agentic Browsing" category, new enough, as of this writing, that almost nothing has been written about what it actually checks. We verified it in-browser in July 2026 and are documenting it here in full, alongside the separate and equally documented fact that Google Search itself does not use any of this for rankings.

What the category is

Google's own description frames the category as working toward "high-quality, browsable websites for AI agents," and states that part of its purpose is validating the "correctness of WebMCP integrations." It is marked as under development and subject to change. This is early, experimental tooling, not a finished standard, and we're treating it that way in everything we say about it.

The six audits, one by one

1. Accessibility tree well-formed. Checks that the structural representation of your page that assistive technology and many agent renderers rely on is properly formed. Not a new check in spirit, but now scoped explicitly to agent-readiness rather than only human accessibility.
2. Cumulative Layout Shift. The same visual-stability metric Lighthouse has graded for years under Core Web Vitals, now also read as a signal for agent-facing quality. A shifting layout is presumably as disruptive to an automated reader as to a human one.
3. llms.txt follows recommendations. Checks for a Markdown file at /llms.txt containing at least one H1. Google's stated rationale: without it, LLMs "may not be able to understand how you want your website to be crawled or used for training." This is a format check, not a content check. It does not verify that the file's information is accurate or current.
4. WebMCP form coverage. Checks whether forms on the page are covered by WebMCP tool definitions, so an agent can act on them directly rather than inferring how to fill them out from page structure.
5. WebMCP tools registered. Checks that a site has registered callable WebMCP tools at all.
6. WebMCP schemas valid. Checks that any registered WebMCP tool definitions are schema-valid.

Three of the six audits concern WebMCP, an emerging, still-experimental protocol for a site to expose callable actions (like submitting a form or fetching a catalog) in a format an agent can call directly, rather than one it has to infer. That weighting tells you where Google is pointing next: the current version of the category is as much a preview of what it will check for later as a grade on what exists today.

What this does and does not mean for SEO

This is a page-quality and developer-tooling category, not a ranking signal, and Google has been explicit elsewhere that it isn't one. Google Search Console has no llms.txt report at all; its crawling settings cover robots.txt only. And Google's AI-optimization guidance, updated 15 June 2026, states directly: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them." Both facts are true at once: Lighthouse now grades a thing Google Search does not use to rank you.

That is not a contradiction unique to this feature. Google has done this before: Core Web Vitals existed inside Lighthouse as a quality signal before it became a ranking factor, and site owners who optimized for it early were ahead of the change rather than reacting to it. We are not promising the Agentic Browsing category follows the same path. That is a possibility worth watching, not something we're selling as a future ranking benefit. Treat that parallel as context, not a guarantee.

What site owners should do now

• Publish a properly formatted llms.txt: Markdown, one real H1. It's inexpensive, it's what the audit checks, and it's a genuine resource for the agents that do go looking for it. See our full guide to what the file is and isn't good for.
• Fix layout stability and accessibility-tree issues if you have them. This is established Lighthouse hygiene, now with a second reason to care.
• Treat WebMCP as a watch item, not an urgent build. It's early, experimental, and three of the six checks in this category concern it, worth tracking as the standard matures, not worth a rush implementation today.
• Do not confuse passing this audit with ranking better. We say exactly this on every page where the audit comes up: the package passes the llms.txt and page-structure checks of Google's agent-readiness audit; the WebMCP audits cover interactive tools a fact package deliberately does not install. We make no claim that any of it improves search position.

Why this matters more than the file check alone

The audit measures whether the right artifacts exist and are well-formed. It does not measure whether an AI agent, arriving with an actual question, can find and use what's inside them: a different and, for most businesses, more commercially relevant question. Our own testing found that an unpointed llms.txt was reached in 0 of 18 agent visits (Note 15), and in a separate test where the file contained a real price, six of ten runs fetched the file and zero of ten reported that price in their final answer (Note 19). The Agentic Browsing audit can be satisfied by a file no buyer's agent ever successfully reads. Passing it and being correctly understood by an agent are related, but they are not the same achievement, and the fix for the second one is different, and covered on our llms.txt guide and our AI SEO services page.

Get audited both ways

Run a free AI visibility audit to see what an agent actually recovers from your site, or see the full agent-readiness package, which is built to pass Google's Agentic Browsing audit and to measurably improve what agents recover from your pages.