Updated September 2026 to reflect Google’s latest Lighthouse documentation.
Google’s Lighthouse tool now has a fifth category next to Performance, Accessibility, Best Practices and SEO. It’s called Agentic Browsing. And it measures something none of the others do: whether an AI agent can actually use your website.
If that sounds abstract, stick with us. A growing share of the traffic hitting your site won’t be a person clicking around. It’ll be an AI assistant booking an appointment, comparing prices or filling in a contact form on someone’s behalf. Google is already scoring for it.
What is agentic browsing in PageSpeed Insights?
Agentic Browsing is a Lighthouse category that checks how well AI agents can read, understand and act on your web page. It arrived in Lighthouse 13.3 in May 2026. Because PageSpeed Insights runs on Lighthouse, many PageSpeed Insights reports now show it too, as a fifth result after SEO.
Google labels the category experimental. The standards behind it are still being written. So treat it as an early diagnostic, not a final grade.
If your PageSpeed Insights report doesn’t show it yet, don’t worry. See how to check your own site below.
How the agentic browsing score works
Unlike Performance or SEO, this category doesn’t give you a score out of 100. Google deliberately avoided a weighted score because the agentic web is still taking shape.
Instead, you get three things:
- A pass ratio. How many of the checks that apply to your page it passed. For example, 3/3.
- Pass or fail on each audit. With specific errors or warnings where something is broken.
- Not applicable results. Checks that don’t apply to your site are left out of the ratio.
That last point matters. Most business websites haven’t built WebMCP tools yet, so those checks usually drop out. A typical site sees something like 2/2 or 3/3, not a score out of six. A small ratio doesn’t mean you’re behind.
The six agentic browsing checks
Google’s documentation groups six audits into four areas.
1. Accessibility for agents
Every web page gets turned by the browser into two things: the visual layout people see, and a structural map underneath called the accessibility tree. Screen readers have used this tree for decades. AI agents now read the same tree to work out what’s on the page and what they can do with it.
Google calls the accessibility tree an agent’s primary data model. Lighthouse checks that interactive elements have proper names and labels, that roles and parent-child relationships are valid, and that nothing interactive is hidden from the tree.
2. Layout stability
This is Cumulative Layout Shift (CLS), the same metric behind Core Web Vitals. Agents often take a snapshot of a page and then act on it. If a button moves after the agent has found it, the agent can click the wrong thing. Agents move faster than people, so shifts during loading hit them harder.
3. llms.txt
An llms.txt file is a short, machine-readable summary of your website, placed at your domain root. Lighthouse checks whether you have one and whether it’s well formed. Our own is at the-lighthouse.co/llms.txt if you want to see what one looks like.
The file is optional. If you don’t have one, the check is typically marked not applicable rather than failed.
4 to 6. WebMCP integration
Three audits cover WebMCP, a proposed standard that lets a website expose its own actions directly to AI agents. Think “book a table” or “request a quote” as a tool an agent can call, instead of an agent guessing its way through your form. Lighthouse checks for registered WebMCP tools, forms that could declare WebMCP support but don’t, and whether your WebMCP schema is valid.
WebMCP is very early. These audits need Chrome 150 or later and registration for Google’s WebMCP origin trial. For most businesses, it’s one to watch, not one to act on yet.
Why “accessibility tree is not well-formed” is the one to fix first
Of all the checks, this is the one worth your immediate attention. Layout stability is something you’ve probably tackled already for Core Web Vitals. llms.txt and WebMCP are optional extras still finding their feet. The accessibility tree check is different. It fails for concrete, fixable reasons. And fixing it helps real people too, not just agents.
Common causes include:
- Icon-only buttons and links with no accessible name. A social media icon or a hamburger menu that’s just an image, with nothing telling a screen reader (or an agent) what it does.
- ARIA attributes with nothing behind them. A search box carrying
aria-expandedandaria-haspopupwithout therole="combobox"that’s supposed to justify them. Visually, nothing looks wrong. Structurally, it’s a contradiction. - Empty elements with labels. A
<div aria-label="Promotional banner">with nothing inside it. The label promises content that isn’t there. - Broken parent-child relationships. A list item outside a list, or a menu item outside a menu. Small nesting errors that don’t change how a page looks, but do change how it’s read.
None of these are exotic. They’re the same issues WCAG accessibility audits have flagged for years. If your site has already had accessibility-focused UX/UI work, you’re likely ahead.
How to check your own site
There are three ways to run an agentic browsing audit today.
- PageSpeed Insights. Run your page at pagespeed.web.dev and look for Agentic Browsing after the SEO result. It’s the quickest option, if it appears for you.
- Chrome DevTools. In Chrome 150 or later, open DevTools, go to the Lighthouse panel and run a report with Agentic Browsing ticked. To inspect the tree itself, open the Elements panel, then the Accessibility pane, and switch on the full-page accessibility tree view.
- The Lighthouse command line. For developers. Install the latest Lighthouse with
npm install -g lighthouse@latest, then run it against your URL to get the full report.
Whichever you use, you’ll get a list of the exact elements causing a failure, not just a vague category result.
Scores can shift a little between runs. Layout shifts, changes to the page’s structure and when scripts load can all move the result. Run it two or three times before drawing conclusions.
Should you panic about this?
No. Google has been clear that the category is under development, and that the aim right now is gathering signals, not ranking sites. Google hasn’t described Agentic Browsing as a search ranking factor. You won’t lose visibility overnight because of a low result.
Treat it like any early warning instead. The businesses that clean up their accessibility tree and layout stability now will have done the work by the time agent traffic becomes a real slice of the pie. The ones that wait will be retrofitting under pressure.
And the honest upside is that almost everything this category rewards, you should be doing anyway. Named buttons. Valid ARIA. Stable layouts. That’s not a new checklist written for robots. It’s the web accessibility standard that’s existed for years, with a new audience reading it.
Agentic browsing FAQs
Is agentic browsing a Google ranking factor?
Not as far as Google has said. It’s an experimental diagnostic in Lighthouse and PageSpeed Insights. The fixes it points to, like better accessibility and a stable layout, are good for your visitors and your site either way.
What’s a good agentic browsing score?
Passing every check that applies to your page. Because the ratio only counts applicable checks, 2/2 and 3/3 are both full marks. Focus on fixing any failed audit, not on the size of the numbers.
Do I need an llms.txt file?
It’s optional. It’s quick to create and does no harm, as long as it’s an accurate summary of your services and key pages. Don’t fill it with marketing fluff.
Do I need WebMCP?
Not yet, for most businesses. It’s an early proposed standard that needs a Chrome origin trial to test. Worth watching if you run bookings, quotes or e-commerce at scale.
How is this different from AI search and GEO?
They’re related but not the same. AI search and GEO are about getting your business mentioned in answers from tools like ChatGPT and Google’s AI Overviews. Agentic browsing is about whether an AI agent can actually operate your site once it gets there. You want both.
Want a second pair of eyes?
If you want to know where your site stands, our free website audit covers agentic readiness alongside the usual performance and SEO checks. You can also check your layout stability in seconds with our free speed test.
And if you’re ready to build for AI agents properly, from WebMCP tools to automated workflows, see our AI automation and agentic growth services, or get in touch.