Agentic Browsing in PageSpeed Insights: How to Make Your Website AI-Ready (2026)

TL;DR

• PageSpeed Insights now grades your site for AI agents, not just humans. Lighthouse 13.3 (May 7, 2026) added an Agentic Browsing category to the default config; PageSpeed Insights inherited it within two weeks. It sits next to Performance, Accessibility, Best Practices and SEO — and reports a ratio like 3/3, not a score out of 100.
• It checks three things by default: a clean accessibility tree, a stable layout (low CLS), and a valid llms.txt at your domain root. The wider Lighthouse category also audits WebMCP — annotated forms and registered agent tools.
• Google added it because agents are now real traffic. Operator, Computer Use, Project Mariner, Perplexity and ChatGPT's browse mode visit sites on a person's behalf. A growing share of requests hitting your server are software, not people.
• You don't fail for lacking AI features — the category is informational. But "AI-ready" is now a measurable, public number, and it's about to become a competitive one.
• Proof it's achievable: a real Lighthouse 13.4 run on my own site, umesh-malik.com, scores Agentic Browsing 3/3 and 100 on Accessibility, Best Practices and SEO — and it passes 7/7 on the isitagentready.com protocol-discovery checklist — using nothing but static files and one Cloudflare Worker. Here's exactly how, so you can copy it.

Your Website Is Now Being Graded for Robots

Open pagespeed.web.dev, run a report, and you'll see something that wasn't there a month ago: a category called Agentic Browsing, sitting right alongside Performance and SEO. It doesn't show a number out of 100. It shows a ratio — 2/3, 3/3 — and a short list of checks with names like "accessibility tree" and "llms.txt".

That ratio is not measuring how fast your page loads or whether your headings are in order. It's measuring how well an AI agent can read your page, understand it, and act on it — with no human in the loop.

This is the quiet half of a shift that's been building all year. We spent two decades optimizing for two audiences: humans who read, and crawlers that index. There's now a third, and it behaves like neither. An agent doesn't skim your hero copy or admire your animations. It wants structure it can parse, facts it can extract, and tools it can call. Google just turned "are you ready for that audience?" into a number anyone can pull up.

💡 Key insight: SEO made your site findable. GEO made it quotable. Agentic Browsing makes it usable by software. These are three different jobs, and the third one is now scored in the same tool you already use for Core Web Vitals.

What Is the Agentic Browsing Category in PageSpeed Insights?

Agentic Browsing is a Lighthouse category that scores how ready a page is for an AI agent to read it, understand it, and act on it without a human driving. It was introduced in Lighthouse 13.3 on May 7, 2026, moved straight into the default config, and PageSpeed Insights picked it up within a couple of weeks. As of mid-June 2026 it's live for everyone.

A few things make it behave differently from every other Lighthouse category:

• It reports a ratio, not a score out of 100. You'll see 3/3, not 92. It's a count of passed checks, not a weighted index.
• It's marked "under development." The exact audits and the way they're scored will change. Don't carve a 3/3 into your OKRs yet.
• It won't fail you for having no AI features. example.com — a page with almost nothing on it — earns a perfect ratio. This is a checklist of opportunities, not a penalty box.

In PageSpeed Insights, the default result is built from three checks. The broader Lighthouse category runs four audits:

Notice what's not there: nothing about keywords, backlinks, or meta descriptions. This audit is about machine comprehension and machine action, full stop.

Why Google Added It — Agents Are Real Traffic Now

The cynical read is "Google wants another number to chase." The accurate read is simpler: a meaningful and growing share of the requests hitting public web servers are agents, not humans — and the tooling finally caught up to that reality.

Look at who's browsing on a user's behalf in 2026: OpenAI's Operator, Anthropic's Computer Use, Google's own Project Mariner, Perplexity, and ChatGPT's browse mode. These don't issue a query and read ten blue links. They get a task — "compare these three products and book the cheapest one that ships by Friday" — and they execute it across multiple sites. To do that, they have to read your page, model what's on it, and act.

When an agent hits a page built only for human eyes, three things go wrong:

1. Comprehension is expensive. Feeding raw HTML or a screenshot into a model burns tokens and invites mistakes. A clean accessibility tree is an order of magnitude cheaper to reason over.
2. Action is fragile. If your "Add to cart" is a <div> with an onclick, an agent has to guess. Annotated forms and registered tools remove the guessing.
3. Discovery is a coin flip. Without an llms.txt or a tool manifest, the agent has to reverse-engineer your site's structure every single visit.

The honest caveat
This category is new and explicitly "under development." llms.txt in particular isn't yet widely consumed by AI tools — even the Lighthouse team says so. None of this is a guaranteed ranking lever today. It's a low-cost bet on where the web is obviously heading, made measurable a year or two before most sites will bother. That early-mover window is the whole point.

How It Helps Developers (and Agents)

For agents, the payoff is obvious: cheaper comprehension, reliable action, less hallucination about what your site does.

For developers, the wins are quieter but real:

How to Make Any Website AI-Ready

Here's where it gets practical. The PageSpeed category is the headline, but the fuller checklist lives at isitagentready.com — a free scanner that groups agent-readiness into five categories. I've used its taxonomy to structure the work below, because it maps cleanly onto what agents actually look for.

You don't need all five. A blog needs the first four and can ignore commerce entirely. A store needs all five. Work top-down — discoverability is the cheapest and highest-leverage.

💡 Key insight: The single highest-leverage change most sites can make is also the most boring — stop rendering text and diagrams as images. An image of a table is invisible to the accessibility tree. A real <table> (or a component that renders one) is readable by screen readers, crawlers, and agents in one shot. I'll come back to this, because it's exactly how I scored my own site.

If you want to go deeper on the callable layer, I wrote a full walkthrough of building a production MCP server and deploying it on Cloudflare Workers — that's the same server backing the numbers below.

Proof: How AI-Ready Is umesh-malik.com?

Talk is cheap, so here's the receipt. I ran my own site — umesh-malik.com — through Lighthouse 13.4 (the engine behind PageSpeed Insights) and the isitagentready.com checklist. This site is a SvelteKit SSG — fully static, one Cloudflare Worker, no special infrastructure. Everything below ships from files in a public repo.

The Agentic Browsing category comes back 3/3 — every weighted check passing:

Here's the honest, category-by-category breakdown — including what the site doesn't have:

.md and /llms.txt', tone: 'neutral' }] },
{ label: 'llms.txt + llms-full.txt', cells: [{ text: 'Yes', tone: 'positive' }, { text: 'Valid H1, description, ~3,500 words, every post linked', tone: 'neutral' }] },
{ label: 'MCP server card', cells: [{ text: 'Yes', tone: 'positive' }, { text: '/.well-known/mcp/server-card.json — 4 tools, Streamable HTTP', tone: 'neutral' }] },
{ label: 'Agent Skills + API catalog + WebMCP', cells: [{ text: 'Yes', tone: 'positive' }, { text: 'agent-skills/index.json, RFC 9727 linkset, navigator.modelContext', tone: 'neutral' }] },
{ label: 'OAuth discovery + auth.md', cells: [{ text: 'Yes', tone: 'positive' }, { text: 'Honestly declares the site as public/anonymous — no fake auth server', tone: 'neutral' }] },
{ label: 'DNS-AID + Web Bot Auth', cells: [{ text: 'Not yet', tone: 'negative' }, { text: 'On the roadmap — newer, lower-leverage for a content site', tone: 'neutral' }] },
{ label: 'Commerce (x402, MPP, UCP, ACP)', cells: [{ text: 'N/A', tone: 'neutral' }, { text: 'It\'s a portfolio + blog. Nothing to sell, nothing to fake.', tone: 'neutral' }] }
]}
/>

The result: the site passes every category that applies to it and scores 7/7 on protocol discovery. The 3/3 above is a real Lighthouse run, not a mockup — and I'm not pretending DNS-AID and Web Bot Auth are done, because they aren't. (Note the three WebMCP audits show as Not Applicable in the screenshot: the site registers WebMCP tools in code, but Lighthouse's WebMCP audits are still informational and didn't score them on this page — an honest nuance of a category that's openly "under development.") That candor is the point. Agent-readiness is a real engineering state, not a vanity badge.

How I Actually Got Here (and How You Can Copy It)

None of this required a backend rewrite. The whole agent-discovery layer is one Cloudflare Worker plus a handful of static files:

The most important line in that list is the one about diagrams as DOM components. This very post is the proof: every chart, table and step list you've scrolled past is a real Svelte component rendering semantic HTML — not a PNG. That's why an agent (or a screen reader) can read all of it, and it's a large part of why the accessibility-tree check passes. One decision, paid back across three audiences.

The reusable principle
You don't need Cloudflare, SvelteKit, or my stack. The pattern generalizes: reuse the content you already publish (your feed, your Markdown, your profile) and expose it through machine-readable surfaces — robots.txt, llms.txt, Link headers, an MCP endpoint. The data already exists. Agent-readiness is mostly about presenting it in formats agents understand, not creating new content.

Common Mistakes to Avoid

• Chasing the ratio instead of the readiness. example.com scores a perfect ratio with nothing real behind it. A green 3/3 on a site full of image-of-text content is a lie you're telling yourself. Optimize the underlying state, not the badge.
• Faking an auth server. If your site is public, say so in auth.md and OAuth discovery. Advertising endpoints that don't exist breaks the agents that trust them. Honest "anonymous, no auth" beats a fictional token endpoint.
• Treating llms.txt as a keyword dump. It's a map, not a meta-keywords tag. Give it a clear H1, a real description, and links to your genuinely best content. Stuffing it is the 2007 SEO mistake in a new file.
• Shipping images of text. The single most common thing that quietly fails the accessibility-tree check. If a human needs to read words in it, it should be real text, not a screenshot.
• Ignoring CLS because "it's just SEO." Layout that jumps confuses screenshot-based agents the same way it annoys users. Your Core Web Vitals work now pays an agentic dividend too.

The Bottom Line

Agentic Browsing in PageSpeed Insights is small today — a ratio, marked "under development," consumed by tools that are themselves a year young. It would be easy to dismiss. Don't.

The trajectory is unmistakable: agents are becoming a first-class audience for the web, Google just made their needs measurable, and the work to satisfy them is cheap, mostly static, and overlaps almost entirely with accessibility and good engineering you should be doing anyway. The sites that ship a clean accessibility tree, a real llms.txt, and a callable MCP endpoint now will be the ones agents reach for when the rest of the web is still serving them screenshots.

I made my own site agent-ready with a Worker and some static files, and I documented every move so you can do the same. Start with robots.txt and llms.txt this week. Then decide how deep into protocol discovery your site deserves to go.

Written for umesh-malik.com — no-fluff technical writing on AI, Web Dev, and Engineering. Curious how the callable layer works? Read How to Build a Production MCP Server next.

Originally published at umesh-malik.com

Keep reading on umesh-malik.com:

• AI Agents That Run the Business in 2026: Why 77% Never Reach Production (and What the 23% Do Differently)
• How to Build a Production MCP Server (I Added One to My Site)
• Build a RAG Pipeline From Scratch (Production Patterns That Actually Matter)