Skip to main content
Back to Notes
·9 min read

What Is Agentic Browsing in PageSpeed Insights? (And How to Pass the llms.txt Check)

Google added an Agentic Browsing score to PageSpeed Insights in May 2026. Here is what it checks, why it exists, and how to pass the llms.txt audit.

Agentic BrowsingPageSpeed Insightsllms.txtGEOAI Search

Agentic Browsing is a category Google added to Lighthouse 13.3.0 on May 7, 2026, and to PageSpeed Insights soon after. It scores how ready your site is for AI agents, not human visitors. Instead of a 0 to 100 number, it reports a pass ratio across a few deterministic checks: an agent-readable accessibility tree, layout stability, and a valid llms.txt at your domain root.

I ran my own site through PageSpeed Insights, saw the new section, and passed it 3 of 3. That sent me looking into what Google expects from llms.txt, which is why I built a validator for it. This post is the full picture: what Agentic Browsing checks, when it shipped, why it exists, how it fits the shift from SEO to AI readiness, and how to pass the llms.txt audit.

Key Takeaways

  • Agentic Browsing is a Lighthouse and PageSpeed Insights category that scores how well AI agents can read and act on your site. It shipped in Lighthouse 13.3.0 on May 7, 2026.
  • It reports a pass ratio, not a 0 to 100 score, and it doesn't change your Performance, Accessibility, Best Practices, or SEO scores.
  • The llms.txt audit passes when the file has an H1, at least one Markdown link, and more than 50 characters of content.
  • Google Search does not use llms.txt for ranking or its AI surfaces. Passing the audit is about AI-agent readiness rather than Google visibility.
  • I built an open-source skill that validates and generates llms.txt and its companion files so the audit passes and everything follows the spec.

What is Agentic Browsing scoring?

Agentic Browsing evaluates how well your site is constructed for machine interaction through a set of deterministic audits, in Google's own words (Chrome for Developers, 2026). It sits next to Performance, Accessibility, Best Practices, and SEO in Lighthouse and PageSpeed Insights, and it scores something none of those touch: whether an AI agent can read your page, understand it, and act on it.

It shipped in Lighthouse 13.3.0 on May 7, 2026. The Lighthouse changelog notes a new agentic browsing category added to the default config, and it reached PageSpeed Insights soon after. Google marks the whole category experimental, so expect it to change.

There is no 0 to 100 score here. Agentic Browsing reports a fractional pass ratio, like 3 of 3, in its own section of the report. It doesn't feed into any of your other Lighthouse scores, which is why a site with a perfect 100 for SEO can still fail it.

Why does Agentic Browsing exist?

It exists because a growing share of the traffic hitting websites is now agents, not people. In 2026, tools like OpenAI's Operator, Anthropic's Computer Use, Google's Project Mariner, Perplexity, and ChatGPT's browse mode visit sites on a person's behalf. Google built Agentic Browsing to measure whether those agents can use your page.

Agents do not see your site the way a person does. Google's documentation puts it directly: agents rely on the accessibility tree as their primary data model (Chrome for Developers, 2026). If your buttons have no names, your roles are broken, or your layout jumps around while the agent reads it, the agent clicks the wrong thing or gives up.

What does Agentic Browsing check, and how do you run it?

The category runs deterministic audits in a few areas: agent accessibility, layout stability, llms.txt, and WebMCP. To run it, enter your URL at pagespeed.web.dev and PageSpeed Insights runs the audits for you; the Agentic Browsing section shows a pass or fail for each check. If you would rather run the audits yourself in Chrome DevTools or the Lighthouse CLI, that needs Chrome 150 or later, and the WebMCP audits need the WebMCP origin trial.

  • Agent accessibility. Every interactive element needs a programmatic name, valid roles and parent-child relationships, and content that is not hidden from the accessibility tree while it is still interactive.
  • Layout stability. Cumulative Layout Shift measures whether the page stays still. Elements that move while an agent reads them make it act on the wrong target.
  • llms.txt. Checks for a valid machine-readable summary at your domain root. The exact pass conditions are below.
  • WebMCP. Checks for registered WebMCP tools, forms with declarative WebMCP, and valid schemas. These audits need the WebMCP origin trial, so a standard PageSpeed scan may not run them.

When I ran my site, it passed 3 of 3 on the checks the scan ran. Google marks the category experimental, so not every audit applies to every site yet.

The shift from SEO to GEO and AEO

Agentic Browsing is one signal of a bigger change. For twenty years, SEO meant making pages rank for human searchers. Now a second audience reads your site: AI assistants and agents. Optimizing for them has its own names, generative engine optimization (GEO) and answer engine optimization (AEO).

The goal shifts from ranking to being read and acted on. ChatGPT, Perplexity, and Claude pull answers from pages directly. Structured content, clean markup, and machine-readable files like llms.txt are how you make a site easy for them to parse.

One caveat is worth stating plainly, because a lot of posts get it wrong: Google Search does not use llms.txt. John Mueller has compared it to the old keywords meta tag, and Gary Illyes said at Google's Search Central Live in July 2025 that Google does not support it and is not planning to. So passing the Lighthouse llms.txt audit doesn't help your Google ranking or your visibility in Google's AI features. Its value is the audit itself, plus the non-Google AI tools that do read it.

How to pass the llms.txt check

The Lighthouse audit is named "llms.txt follows recommendations," and it passes when three things are true: the file has an H1, it contains at least one Markdown link, and it is longer than 50 characters (DebugBear, 2026). Bare URLs fail. The links have to use Markdown format, a linked title in brackets followed by the URL in parentheses.

So what is llms.txt? It is a Markdown file at the root of your site that gives AI tools a curated map of your content. It follows a spec proposed by Jeremy Howard of Answer.AI in September 2024: an H1 site name, a short summary, then sections of Markdown links. A companion file, llms-full.txt, popularized by Mintlify, holds your whole site's content in one file.

Beyond those two, there are llms-ctx.txt and llms-ctx-full.txt, which a tool generates from your llms.txt, and per-page .md versions of each page. The Agentic Browsing audit only checks llms.txt, but these other files are what give AI tools your full content and per-page context.

Why I built a skill for it

After I passed the audit, I wanted to know exactly what Google expects, and I did not want to eyeball it every time. So I built an open-source Claude skill that validates, repairs, and generates llms.txt and its companion files against the llmstxt.org and Mintlify specs, then I ran it on my own files.

The skill ships a dependency-free validator that flags the mistakes the audit and the spec care about: a missing H1, bare labels instead of Markdown links, prose sitting in a links-only section, and links that point to pages an AI tool cannot fetch. It also generates the companion files and gives per-stack guidance for serving them, whether you run Next.js, Astro, Hugo, WordPress, or plain HTML.

  • Put a valid llms.txt at your domain root: an H1 with your site name, a one-line summary, then sections of Markdown links to your key pages.
  • Use Markdown links, not bare URLs, and keep the file substantive, well past 50 characters.
  • Validate it before you ship, so a missing H1 or a stray bare URL does not fail the audit.

The skill is on GitHub as claude-skill-llms-txt, MIT licensed. It is one of several open-source Claude skills I build and use.

Frequently asked questions

What is Agentic Browsing in PageSpeed Insights?

It is a Lighthouse category, added in version 13.3.0 on May 7, 2026, that scores how ready your site is for AI agents. It checks things like your accessibility tree, layout stability, and llms.txt, and it reports a pass ratio rather than a 0 to 100 score.

When did Google release Agentic Browsing?

Lighthouse 13.3.0 shipped on May 7, 2026, with the agentic browsing category added to the default config, and it reached PageSpeed Insights a couple of weeks later. Google marks it experimental, so the exact audits can still change as the standards settle.

Does llms.txt help my Google ranking?

No. Google Search does not use llms.txt. John Mueller compared it to the keywords meta tag, and Gary Illyes said at Search Central Live in July 2025 that Google does not support it. It helps you pass the Lighthouse audit and it is read by some non-Google AI tools, not by Google Search.

How do I pass the llms.txt audit?

Serve a valid llms.txt at your domain root with an H1, at least one Markdown link rather than a bare URL, and more than 50 characters of content. Following the llmstxt.org spec covers all three conditions.

What is WebMCP?

WebMCP is a proposed standard for exposing a page's actions to AI agents as tools, so an agent can act on a page instead of only reading it. Some Agentic Browsing audits check for it, but they need the WebMCP origin trial, so a normal PageSpeed scan may not run them.

Is llms.txt the same as robots.txt?

No. robots.txt tells crawlers which URLs they may access, and it has been a web standard for decades. llms.txt is newer and gives AI tools a curated Markdown map of your key content. They solve different problems and can sit side by side at your domain root.

Sources

I'm Eric Forte, an AI automation engineer. The llms.txt skill is free and open source on GitHub: claude-skill-llms-txt. I build GoHighLevel and n8n automation at ericforte.com. For more field notes on AI search and automation, subscribe to the newsletter on the blog.

Eric Forte

Eric Forte

GoHighLevel + n8n integration engineer for GoHighLevel marketing agencies. JavaScript when no-code hits its ceiling.

More NotesGet a Free Audit

Subscribe

Get the Next One in Your Inbox

Build notes, breakdowns, and the patterns I'm using on client work I've shipped. One email when there's something worth sending. Unsubscribe in one click.