What an AI agent is, and what the score measures
An AI agent is an assistant that does not just answer from memory, it actually opens web pages and acts on them, reading the content, following links and clicking buttons, the way a person would but much faster. ChatGPT, Gemini, Claude and Perplexity all now do versions of this.
Agentic Browsing is a score inside Google's PageSpeed Insights (the free tool at pagespeed.web.dev) that estimates how well one of those agents can read and use your page. Unlike the other scores, it is not out of 100. It shows a ratio, like 3/3, a simple count of how many readiness checks your page passes. Google has labelled it experimental, so treat it as an early signal, not a settled rule.
What it checks
It looks at a handful of things, all about whether an agent can make sense of your page.
- llms.txt: a plain text file at your site's root that hands AI a short, clean guide to your most important pages. We explain it in full in a separate guide.
- A clear structure, what developers call the accessibility tree: the plain outline of your page that marks which part is the header, which is the menu, which is the main content, and which things are real buttons. Screen readers for blind users have relied on this map for years, and now AI agents use the same one.
- A stable layout: the page should not jump around as it loads. An agent reads and clicks faster than a person, so a button that shifts at the last moment can make it click the wrong thing.
- WebMCP: a brand-new, still-rare way for a site to offer an agent a set of ready-made actions. Almost no site has this yet, so you can safely ignore it for now.
Why it matters, even though it is experimental
The experimental label is fair, and you should not lose sleep over a single ratio. But the direction is the real story. Google, the biggest name in search, has decided that whether an AI can use your site is now worth measuring, right alongside speed and SEO. When Google starts scoring something, it is telling you where it thinks the web is going.
It also quietly settles an old argument. People who work on AEO (getting your site read and cited by AI) have said for a while that you should prepare your site for AI, and plenty of others called it hype. A dedicated score from Google is about as clear a signal as you get that at least the basics are real.
How to pass it, without a developer
Most of this is either already handled for you or a small, one-time tidy.
- llms.txt: this is a file, and it is one of the things we generate for you as part of a fix. You upload it once and you are done.
- Structure: if your site is on Wix, Squarespace, Shopify or a modern WordPress theme, the frame (header, menu, main area, footer) is mostly built for you. Your job is to use the real blocks inside it, a Heading block for titles, a real Button for actions, and alt text on images, instead of big bold text and clickable pictures. If your site was custom-built, this is a setting in how the page is coded, and any developer can add it in an afternoon.
- Stable layout: this usually comes from images or banners loading without space reserved for them. If your page visibly jumps as it loads, that is the thing to fix. If it does not, you are already fine.
- WebMCP: skip it for now. It is too new to worry about.
The honest version
Here is the balanced take. For where you rank on Google, page structure barely moves the needle, and we say so plainly in our guide to SEO jargon. For whether an AI agent can actually read and use your page, it now counts, and Google has started measuring it. Both of those are true, and the gap between them is the whole point of Agentic Browsing: it is not about ranking, it is about being usable by the assistants people are starting to send in their place.
So do the basics, once. Add the file, use real headings and buttons, keep the page from jumping around, and then get back to your actual work. A free scan will tell you where you stand on both the classic signals and the newer ones, without you having to decode any of this yourself.