Google Lighthouse Now Checks Your Site for AI Agents
Google released Lighthouse Agentic Browsing, an experimental audit that checks whether AI agents can navigate, understand and take action on your website. Four checks: pass or fail.
TL;DR
- Google Lighthouse added an "Agentic Browsing" category (experimental)
- It checks four things: WebMCP tools, accessibility tree, layout stability (CLS) and llms.txt
- The result is pass/fail fractions, not 0-100
- The Causabi check covers some closely related points and shows what to fix
What is Lighthouse Agentic Browsing?
Lighthouse is Google's tool that scores a site on Performance, Accessibility, Best Practices and SEO. Now there's a fifth category: Agentic Browsing.
The premise is simple: AI agents (like ChatGPT, Gemini, Perplexity, and browser-integrated AI assistants) interact with websites differently than humans do. They can rely on the accessibility tree instead of visual layout, and they need stable layouts to reliably click buttons and fill forms.
Lighthouse Agentic Browsing shows how ready a site is for this new type of visitor.
The 4 Agentic Browsing Audits
1. WebMCP Tool Registration
NEWWebMCP is a Chrome-proposed standard that lets your site declare its interactive capabilities to AI agents. Instead of an AI agent guessing how to book an appointment or add something to a cart, the site's forms announce what they do.
Lighthouse, per Chrome's documentation, audits both the declarative version (HTML attributes on forms) and the imperative version (registration through registerTool() in JavaScript). The audit lists registered tools; it requires joining the WebMCP Origin Trial.
2. Agent-Focused Accessibility
PASS/FAILAI agents can navigate a site using the accessibility tree, the same structure screen readers use. If interactive elements lack labels, it is hard for an agent to find or use them.
Lighthouse, per Chrome's documentation, checks three things: names and labels on interactive elements, accessibility tree integrity (roles and parent-child relationships) and visibility of interactive content. If an element has no label, it is harder for an agent (as for a screen-reader user) to find it.
3. Layout Stability (CLS)
PASS/FAILCumulative Layout Shift (CLS) has long been a Core Web Vital. It matters for AI agents too: if a button moves between when the agent finds it and when it clicks, the action may fail.
Target: CLS < 0.1. Common causes of failures: unsized images, late-loaded ads or cookie banners, and injected content above existing elements.
4. llms.txt Discoverability
PASS/FAILChrome's documentation (llms.txt audit) calls llms.txt "an emerging convention used to provide a machine-readable summary of a website's content" for LLMs and AI agents. It is a proposed format (llmstxt.org), not an adopted standard.
The audit checks for the file at your domain root. Server errors are flagged as issues; 404 (file not found) is marked "Not Applicable" since the file is optional. There is no evidence that major assistants read llms.txt.
How the Score Works (It's Not 0-100)
Unlike every other Lighthouse category, Agentic Browsing does not produce a weighted 0-100 score. Instead, you get a fractional ratio: how many of the readiness checks pass.
The reason, as reported: agentic web standards are still being developed. The category is in data-collection mode — providing useful signals without locking in a scoring methodology that might change as WebMCP and related standards mature.
This also means the score can fluctuate: dynamic WebMCP tool registration via JavaScript may be missed during Lighthouse snapshots, and accessibility tree changes from DOM modifications affect results run-to-run.
What You Can Do
Add llms.txt to your domain root
Create a Markdown file at /llms.txt with your business description, key pages and contact information. The format is described at llmstxt.org: it is a proposal, not a standard. It takes about 20 minutes. Lighthouse looks for it; that assistants read it is not proven.
Optional: WebMCP declarations on your forms
Add toolname and tooldescription attributes to key <form> elements — your contact form, booking form, search, add-to-cart. This is declarative WebMCP and Lighthouse should detect it. The step is optional and experimental: the standard is not finalized, and our WebMCP article suggests waiting.
Audit your image alt text
Every <img> needs a descriptive alt attribute. Not empty alt="" (which hides images from agents), but actual descriptions. Agents may use alt text to understand visual content.
Fix semantic HTML landmarks
Ensure your pages use <header>, <main>, <nav> and <footer>. The accessibility tree agents rely on is built partly from these elements.
Measure and fix CLS
Run PageSpeed Insights on your homepage. If CLS is above 0.1, what usually helps: add width/height to images, avoid late-injected content above the fold, preload critical resources.
What Causabi Checks of This
The Causabi check includes points close to some of the Lighthouse Agentic Browsing audits. The result is shown alongside your AI score (0-100):
- +llms.txt presence
- +Image alt text
- +Labels and headings in the accessibility tree
Every failing check comes with what to fix.
Free check: can AI read your site?
See your AI score before you read on
About 90 seconds · No email required · Free
The Bigger Picture
Beyond the audits themselves, the release shows that Google has started measuring how usable sites are for AI agents. The category is still experimental and may change.
When Google adds something to Lighthouse, it usually enters the vocabulary of developers and SEO tools. Whether WebMCP and llms.txt become baseline requirements is not known yet. Accessible markup is worth having either way: it helps people and agents alike.
Checking a site against these points fits into one evening.
Related Articles
FAQ
What is Lighthouse Agentic Browsing?
Lighthouse Agentic Browsing is an experimental audit category in Google Lighthouse that checks how ready a website is for AI agents. Unlike the usual 0-100 Lighthouse score, it uses a fractional pass/fail system across four checks: WebMCP tool registration, agent-focused accessibility, layout stability (CLS) and llms.txt presence. Per Chrome's documentation, the category is experimental and needs Chrome 150 or later, and the WebMCP checks require joining the Origin Trial.
How is the Lighthouse Agentic score calculated?
Unlike traditional Lighthouse categories, the Agentic Browsing category does not produce a 0-100 weighted score. Instead, it shows a fractional ratio (e.g., 3/5 checks passing), pass/fail status per audit, and informational counts. The stated reason, as reported: agentic web standards are still under development.
What is WebMCP and how do I add it to my site?
WebMCP is a Chrome-proposed standard (not yet adopted) that lets websites declare their interactive capabilities to AI agents. The declarative version requires adding toolname and tooldescription HTML attributes to your <form> elements — for example: <form toolname='book_appointment' tooldescription='Book a consultation with our team'>. The imperative version registers a tool with registerTool() in JavaScript (the name of the object it is called on has changed between versions, so check Chrome's documentation). Source: developer.chrome.com/docs/ai/webmcp.
Does llms.txt affect the Lighthouse Agentic score?
Yes, going by the audit description: llms.txt is listed under the Discoverability category. The audit checks whether a summary file exists at your domain root. Server errors are flagged; 404 responses are marked 'Not Applicable' since the file is optional. This does not mean assistants such as ChatGPT read llms.txt: there is no evidence that they do.
How does Causabi help with Lighthouse Agentic Browsing?
The Causabi check includes points close to some of the Lighthouse Agentic Browsing audits: llms.txt presence and accessibility for agents (element labels, headings, image alt text). For each failing check it shows what to fix.