Agentic Browsing is a Lighthouse audit category that scores how well a page's accessibility tree, layout stability, llms.txt file, and WebMCP integration support AI agents operating the page — separate from Lighthouse's existing Performance, Accessibility, Best Practices, and SEO categories. It was introduced around Lighthouse 13.3 in mid-2026 (TODO(faizan): verify exact version and ship date). At the time this was written, Lighthouse's own category description read: "These checks ensure high-quality, browsable websites for AI agents and validate the correctness of WebMCP integrations. This category is still under development and subject to change."
What it actually checks
Running Lighthouse with --only-categories=agentic-browsing surfaces six audits:
agent-accessibility-tree— whether the page's accessibility tree is well-formed: valid ARIA usage, correct element nesting, accessible names on interactive elements. A page can look fine in a browser and still fail this if its markup is invalid — a<dl>with the wrong child elements, for instance.cumulative-layout-shift— the same CLS metric Core Web Vitals uses. A page that shifts under an agent mid-interaction is as much of a problem for an agent as for a human.llms-txt— whether anllms.txtfile exists and follows the basic expected format.webmcp-registered-tools,webmcp-form-coverage,webmcp-schema-validity— whether the page registers WebMCP tools (a way for a page to expose actions an agent can call directly, rather than an agent inferring them from the DOM) and whether those tool definitions are valid. Most sites don't register any WebMCP tools yet, so these typically score as not applicable rather than failing.
How it differs from SEO
Lighthouse's SEO category checks signals aimed at search engine crawlers and indexers — meta tags, crawlability, mobile-friendliness. Agentic Browsing checks something adjacent but distinct: whether an AI agent — software that reads a page's accessibility tree and, increasingly, calls page-exposed tools — can actually operate the page. A page can rank and index perfectly while still being hard for an agent to navigate, if its accessibility tree is malformed or its interactive elements lack accessible names.
How it differs from Core Web Vitals
Core Web Vitals (LCP, INP, CLS) measure loading speed, interactivity, and visual stability for human-perceived experience. Agentic Browsing reuses CLS directly — layout instability is exactly as disruptive to an agent's element targeting as it is to a human's clicking — but adds checks Core Web Vitals doesn't cover at all: accessibility-tree correctness and WebMCP tool validity.
How to pass
- Use valid, semantic HTML and correct ARIA. Invalid nesting (see the
<dl>example above) is a common, easy-to-miss failure mode — it isn't just a style issue, it breaks the accessibility tree an agent reads. - Give every interactive element (links, buttons, form fields) an accessible name — visible text, or
aria-labelfor icon-only controls. - Keep CLS at or near zero.
- Publish a well-formed
llms.txt.
On Shopify
Icon-only buttons in cart drawers, search toggles, and menu icons are the most common accessible-name gap. Apps that inject content after page load (reviews widgets, upsell banners) are the most common source of layout shift. llms.txt isn't generated automatically by Shopify today.
On WordPress
Page builders frequently generate icon buttons and custom form markup without labels — audit these specifically. Ads, cookie-consent banners, and lazy-loaded widgets are common CLS sources. A handful of SEO plugins have started adding llms.txt support; check whether yours does.
On Next.js
npm install ai-visibility, then use LLMSTextGenerator from ai-visibility/generators for llms.txt. Avoid client-only rendering that replaces server-rendered content after hydration, which can both shift layout and leave interactive elements briefly unlabeled.
See it in practice
Run a free scan to check a page's accessibility-tree signals directly, or see this site's own live result — including the accessibility bug this exact check caught during development, described in the accompanying blog post.