Google now measures whether your site works for AI agents

Chrome's Lighthouse added an Agentic Browsing category that checks llms.txt, WebMCP, accessibility and layout stability. What it changes for your site.

What the Agentic Browsing category is

Lighthouse is the audit that runs inside Chrome DevTools and behind PageSpeed Insights. Until now it measured performance, accessibility, best practices and SEO. The new category, documented by Google at developer.chrome.com, measures something else: whether an AI agent can understand and use your page.

One detail almost every write-up on this gets wrong: there is no 0–100 agentic browsing score. Google reports a fraction of checks passed plus a pass/fail list. The documentation says why — the standards for the agentic web are still emerging, so the goal for now is to gather data and give actionable signals rather than produce a ranking.

If someone sells you a "Google agentic browsing score", be sceptical.

The four things it checks

1. llms.txt — a text file at the domain root, alongside robots.txt, holding a machine-readable summary of what the company does and which pages matter. It is the easiest of the four to fix, and the one almost nobody has.

2. WebMCP — an API that lets a site explicitly expose its logic and forms to an agent instead of making it guess from the interface. It requires Chrome 150+ and origin trial registration, which makes it the longest-horizon item here.

3. The accessibility tree — Lighthouse filters the subset of accessibility checks that matter for machine interaction: every interactive element needs a programmatic name, roles and parent-child relationships must be valid, and nothing may be interactive while hidden from the accessibility tree.

4. Layout stability (CLS) — the Core Web Vitals metric you already know, here for a different reason: an agent depends on elements being where it expects them.

Item 3 is the interesting one, and the most misread

Accessibility usually gets treated as a compliance topic — something you do because of WCAG, the ADA, a lawsuit. What this new category makes visible is that it is the same data.

An AI agent reads your page through the same channel a screen reader does. If your quote button is an unlabelled icon, the screen reader announces "button" and the agent has no idea what it does either. These are not two jobs. It is one job with two beneficiaries.

What to do this week

Run Lighthouse on your site. Chrome DevTools, Lighthouse tab. It is free and takes under a minute.

Publish an llms.txt. Plain text, at the domain root, served as text/plain. A summary of what the company does, for whom, where it operates, and the list of key pages. The most common failure on SPA-built sites: the server returns the application HTML instead of the text file — if yourdomain.com/llms.txt opens your homepage, it does not count.

Write only what is true. An llms.txt advertising a service the company does not offer is worse than none: you are handing the model a false claim about your brand in structured, easy-to-quote form.

Name your interactive elements. Icon buttons, bare "learn more" links, form fields with no label. An aria-label fixes it, and fixes it for both audiences.

Where PlaceFirstAI fits

The free PlaceFirstAI diagnosis checks whether your domain publishes an llms.txt that opens without JavaScript. That feeds the readiness score, alongside search and citation crawler reach.

On the Growth plan you also get the llms.txt file already written from what the audit verified on your site — pages that actually responded, services found in your own content — ready to publish.

And an honest caveat: what Lighthouse measures is whether an agent can operate your page. What PlaceFirstAI measures is whether your brand gets cited when a buyer asks ChatGPT, Gemini, Claude or Perplexity. Those are neighbouring problems, not the same one. What is new is that the border between them is shrinking — and Google now ships a tool that shows it.

Run the free diagnosis for your domain and see where your brand stands today.