Until now, PageSpeed Insights measured how fast and pleasant your site is for people. In May 2026, Google quietly added a fifth score to the tool (specifically to Lighthouse 13.3) - Agentic Browsing. It no longer measures how a human experiences the page, but how well an AI agent can read and use it.
The best part? Google can't even agree with itself about it. But first, the basics.
What is agentic browsing
More and more often your site is opened not by a person, but by an agent - ChatGPT, Gemini, Perplexity, Claude - which, on the user's behalf, compares products, fills in forms or adds items to the cart. An agent doesn't "look at" a page, it reads its structure. If that structure is a mess, your site is invisible to the agent, even if it looks great to humans.
A score that isn't a score
The Agentic Browsing category sits alongside Performance, Accessibility, Best Practices and SEO. The twist: it doesn't return a number from 0 to 100. Instead it shows a ratio of passed checks (e.g. 3/3). Google admits the standards are still emerging, so it gathers signals rather than handing out grades. The category is officially "under development" and does not affect your other scores.
It checks a handful of things:
- A clean accessibility tree - the agent "sees" through the accessibility tree, so every button and field needs a clear name (semantic HTML and ARIA).
- A stable layout (low CLS) - if elements jump around, the agent clicks the wrong thing.
- llms.txt - a machine-readable summary of the page for language models.
- WebMCP - annotated forms and "tools" through which an agent can reliably operate your site.
The plot twist: Google vs Google
Here is where it gets fun. Chrome's team rewards you for llms.txt via Lighthouse. Google's Search team, meanwhile, has spent the past year saying the opposite. John Mueller compared llms.txt to the long-dead "keywords" meta tag, and on Bluesky, asked whether Google pages hosting llms.txt count as an endorsement, he replied: "I'm tempted to say something snarky since this comes up so often, but to be direct, no." Another time he noted that with the file you are essentially telling these systems "I have the best website ever".
The numbers back him up. Ahrefs looked at 137,000 domains: 97% of llms.txt files received zero requests in May 2026, and only 28% of domains published one. The cherry on top: on 3 December 2025 an llms.txt briefly appeared in Google's own developer docs - and was deleted the same day. Brett Tabke, founder of the Pubcon conference, put it bluntly: "We just don't need people thinking LLMs are different from any other spider."
The finest irony: many people generate their llms.txt using an LLM that reads their HTML - so that another LLM will not have to read it. A robot writing a note for a robot about a page the robot can already read.
So - should you care?
More than you would think, but not because of llms.txt.
Two of the four checks (the accessibility tree and a stable layout) are things you should fix anyway - they also help humans, SEO and conversions. WebMCP is the direction Google is actually pointing to. llms.txt is a cheap "nice to have": for documentation read by coding assistants (Cursor, Claude Code, Copilot) it has a small, real niche, but for a typical shop do not expect miracles today.
Common sense: do not chase the 3/3 ratio as a fetish. Get the basics right (accessibility, speed, semantics, structured data), keep an eye on WebMCP, and add llms.txt because it costs you nothing. Agent traffic is coming - the only question is whether it finds you ready.
How we can help
At Degriz we technically prepare online stores for this shift - accessibility, speed, semantic structure, structured data. If you would like to know how your site scores on agentic browsing and what to improve, we will run an audit for you.
