Read

What Lighthouse’s Agentic Browsing Score Actually Measures

If you’ve run Lighthouse or PageSpeed Insights on your site since May and noticed a new category sitting under Performance, Accessibility, Best Practices and SEO — you’re not imagining it. Agentic Browsing shipped in Lighthouse 13.3 on 7 May 2026, and it’s in the default configuration.

It’s also unlike every other Lighthouse category, which is why it confuses people. Here’s what it’s actually doing.

It’s a fraction, not a score

Every other Lighthouse category gives you a weighted average from 0 to 100. Agentic Browsing doesn’t. Instead you get a ratio — how many agentic readiness checks your site passes, out of the total run.

Google has been explicit about why. From the official documentation:

“Because the standards for the agentic web are still emerging, the current focus is to gather data and provide actionable signals rather than a definitive ranking.”

In other words: Google isn’t confident enough yet to tell you that a 73 is good and a 41 is bad. It’s telling you which specific things you pass and which you don’t, and leaving the judgement to you.

That’s a reasonable position, and it’s worth holding onto when someone tries to sell you a “score improvement” package. There isn’t a score to improve. There’s a list of things that either work or don’t.

Before you run it

Two practical gotchas that catch people out:

You need Chrome 150 or later. Run it in an older build and the category either won’t appear or won’t populate properly.

The WebMCP checks need an origin trial registration. WebMCP is still experimental, so unless you’ve registered your origin for the WebMCP origin trial, those particular audits won’t report meaningfully. If you’re seeing blanks rather than failures, this is usually why.

The whole category is flagged experimental by Google. That’s not a reason to ignore it — but it is a reason not to panic about it, and a reason to expect the checks to change.

What it actually checks

The audits fall into four groups.

1. WebMCP integration

WebMCP is a proposed standard that lets a website hand an AI agent a structured list of things it can do, rather than making the agent screenshot the page and guess where to click. Lighthouse checks three things here:

  • Registered WebMCP tools — has your site declared any at all?
  • Forms missing declarative WebMCP — you have forms an agent might reasonably want to complete, but haven’t exposed them.
  • WebMCP schema validity — you’ve declared tools, but the schema is malformed.

Lighthouse hooks into the Chrome DevTools Protocol to watch tool registration events, and it checks both declarative tools (defined in your HTML) and imperative ones (registered via JavaScript).

For most sites today, all three of these will fail simply because you haven’t implemented WebMCP. That’s not a crisis — see post 3 in this series for whether you should.

2. Agent-centric accessibility

This is the group that matters most, and the one nobody talks about.

An AI agent doesn’t see your page the way you do. Its primary data model is the accessibility tree — the same structure a screen reader uses. Lighthouse runs a filtered subset of the accessibility audits that are critical for machine interaction:

  • Names and labels — does every interactive element have a programmatic name?
  • Tree integrity — are your roles and parent-child relationships valid?
  • Visibility — is content interactive while hidden from the accessibility tree?

If you’ve done serious accessibility work, you are already most of the way there. That’s the genuinely good news buried in this whole topic: agent readiness and accessibility are largely the same job. The work you did to serve users with screen readers is the work that now serves the machines reading your site on someone’s behalf.

If you haven’t done accessibility work, this is where your fraction is losing points, and it’s where the fix has the broadest payoff.

3. Layout stability

Cumulative Layout Shift — the same CLS you already know from Core Web Vitals — appears here for a different reason.

For a human, layout shift is annoying: you go to click a button and an ad pushes it down. For an agent, it’s fatal. The agent identifies an element’s position, then attempts to interact with it, and in between those two moments the element has moved. The action fails or, worse, hits the wrong thing.

Ads, images without declared dimensions, and content injected after load are the usual culprits.

4. Discoverability (llms.txt)

Lighthouse checks whether you have an llms.txt file at your domain root.

I have a lot to say about this one, and it’s the subject of the next post in this series. The short version: add the file, because it costs ten minutes and this audit looks for it — but don’t build a strategy on it. The evidence that it does anything else is very weak.

Why your results might fluctuate

The audits are deterministic, but people still see their fraction move between runs. Three usual causes:

  • Timing of dynamic tool registration. If you register WebMCP tools via JavaScript, whether Lighthouse catches them depends on when they fire relative to the snapshot.
  • Accessibility tree variability. Big changes in DOM size or complexity change the tree, which changes the audit.
  • CLS. Layout shift has always been variable between runs. It still is.

If your number bounces around, run it three times and take the pattern rather than the reading.

What to actually do about it

In order of value:

  1. Fix your accessibility tree. Semantic HTML, proper labelling, valid ARIA. Biggest win, broadest benefit, and it helps real people too.
  2. Fix your layout shift. Declare image dimensions, reserve space for injected content. Helps your Core Web Vitals as well.
  3. Add an llms.txt. Ten minutes. Tick the box. Move on.
  4. Consider WebMCP — but only if your site actually does something an agent would want to do. More on that in a fortnight.

Notice that items 1 and 2 are things you should have been doing anyway, and both pay off in areas that have nothing to do with AI. That’s usually how it goes with this stuff: the genuinely useful work is unglamorous and already on your list.


Want to know where you stand? Our Agent Readiness Audit runs these checks properly and explains what each result means for your business — rather than leaving you with a fraction and a shrug.

Let’s find out where you stand →