Chrome's Agentic Browsing audit graded my site first. I ran it on 10 more AI-tool sites.

Chrome's Lighthouse tool failed our own site in July on its new Agentic Browsing check, one I didn't know existed, for a mistake that took less time to make than it did to notice. That was interesting enough that I ran the same audit again today, unedited, against ten more AI-tool websites: OpenAI, Anthropic, Perplexity, Notion, Jasper, Copy.ai, Zapier, ElevenLabs, Cursor, and Gamma. Only two of the ten passed every check that applied to them.

This is for anyone whose site might get visited by an AI agent instead of a human, which by now is most of us. By the end you'll have the exact command I used, the real failure messages from real company sites, and one honest catch: the score isn't the usual 0-100 Lighthouse number, and that distinction turned out to matter more than I expected.

Does Chrome's Agentic Browsing audit for AI agents still exist?

Yes: as of August 2026, it's a standard part of Lighthouse, not an experiment you opt into. It has shipped by default since Lighthouse 13.3, and I confirmed it's still there by running --list-all-audits on the current release, 13.4.1, under headless Chrome 150. The category checks four things. Your accessibility tree, because agents read the same DOM-derived tree that screen readers use to find buttons, links, and labels. Cumulative Layout Shift, so an agent doesn't click a button that moved. Your llms.txt file, a plain-text page description agents can read instead of guessing. And WebMCP, a proposed standard for declaring page functions, like a search box, as tools an agent can call directly (see Chrome's own scoring docs for the full spec).

One detail changes how you read every score below: this category doesn't average to 0-100 like the others. It's a ratio of the checks that actually apply to your site. Skip that detail and the numbers look inconsistent. Know it, and they tell a sharper story.

Where I went wrong first

Our own mistake was in the one file built specifically for AI agents, not a page any human reads. Our llms.txt listed key pages like this: - Build in a day: https://inaday.ai/build. Fine for a person. Fine for a normal crawler. The Agentic Browsing parser needs at least one real markdown link, [text](url), and a bare URL after a colon doesn't count as one. It counted our whole file as zero links and failed the check outright.

So I ran the same audit on 10 real AI-tool sites

The real test wasn't our own site. It was whether AI companies practice what their own products preach. I picked ten homepages, ran one command against each, and changed nothing between runs:

npx lighthouse@latest https://example.com --only-categories=agentic-browsing --output=json --chrome-flags="--headless=new"

No logins, no crawled subpages, just the front door every agent hits first. And that pass-ratio quirk from the section above turned out to be the sharpest finding of the day.

The results: who actually passes

Two sites passed everything Lighthouse checked. Two failed almost everything.

Eight of the ten failed the accessibility tree check, and the failures were specific, not vague. Anthropic's homepage flagged "Links must have discernible text" on a featured-article link. Copy.ai flagged "Certain ARIA roles must contain particular children" on a testimonial slider. Cursor flagged an ARIA-hidden element that was still focusable, a live product demo sitting behind a screen an agent can't see into.

None of this is cosmetic if agents are actually doing the browsing on your site instead of a human. We've already tested it from the agent's side, too: handing one our daily errands for a day showed how much an agent leans on clean, labeled structure just to get through a normal task list.

Layout stability, by contrast, is basically a non-issue right now: every one of the ten scored 1.0 or 0.99, and the single worst Cumulative Layout Shift value in the group was 0.052, comfortably inside Google's 0.1 "good" threshold.

The one check almost nobody has

WebMCP, not the accessibility tree, was the rarest pass of the ten. Only Zapier had a real one: its site registers an imperative tool called get_started_building, which returns install commands and starter prompts for its own developer SDK, straight from Lighthouse's audit output. Every other site, including the two that scored 100%, came back "not applicable" here, because WebMCP still requires signing up for an origin trial to test at all. Near-zero adoption isn't a company failing a check. It's a standard nobody's shipped yet.

If you built your own site with an AI coding agent in the first place, that's worth a second look: I walked through three different ways to build a website with AI, and none of the three routes generates an llms.txt or WebMCP registration for you automatically. You'd add both by hand, the same way I just did.

Would I do it this way again?

Yes, and next time I'd start here instead of ending here. Running this audit on ten sites before touching our own would have caught the bare-URL mistake before it shipped, not after Marco ran Lighthouse and found it by accident. Fair enough: nobody proofreads their own llms.txt for markdown syntax.

The bigger surprise wasn't my mistake. It was that eight of ten AI companies haven't closed gaps that are quick to find and quicker to fix. OpenAI's 100% score is a good example of why the number alone can mislead: it comes from passing two checks, not four, because it has no llms.txt to fail. Good enough ships, most of the time. For a check this cheap, that's not a great excuse.

At a glance