Lighthouse 13.3 (May 2026) added an "Agentic Browsing" category alongside the familiar Performance, Accessibility, SEO, and Best Practices scores — see the official scoring reference from the Chrome DevTools team. That is a genuinely new kind of audience for a website to optimize for: not a human with a mouse, and not strictly a search-engine crawler, but an AI agent that navigates a page, reads its structure, and acts on what it finds — clicking, filling forms, extracting data — on behalf of a user who may never look at the raw page itself.
What the category actually checks#
Agentic Browsing runs three checks, and they are narrower and more concrete than the category name suggests:
- Accessibility tree well-formedness — every interactive element needs a programmatic name and label, roles and relationships in the tree have to be valid, and nothing can be interactive while hidden from the tree. This is the same underlying tree screen readers use, which is not a coincidence — an agent navigating programmatically and a screen reader navigating for a blind user are solving a similar problem.
- Cumulative Layout Shift stability — a page that keeps shifting elements around is unreliable for an agent that clicked based on a coordinate or a snapshot taken a moment earlier. A human visually adjusts to shifted content without thinking about it; an agent's action can miss entirely if the target moved.
llms.txtvalidity — a plain-text file at the site root describing the site's purpose and structure for AI systems. Search Engine Land's coverage of the check confirms it specifically verifies the file contains real Markdown-formatted links ([label](url)), not bare URLs — bare URLs are harder for a language model to parse confidently as "this text names that destination."
ℹWhy the accessibility overlap is not a coincidence
Every one of these checks was already valuable for a completely different audience before agents existed — a well-formed accessibility tree serves screen reader users, layout stability serves anyone on a slow connection or older device, and a clear site map in plain language serves any new visitor trying to understand what a site actually is. Agentic Browsing did not invent new work; it gave existing good practice a second, concrete reason to exist.
What this means in practice#
If your site already scores well on Accessibility, you likely already pass most of the Agentic Browsing checks — that is the actual relationship between the two categories, not a coincidence of naming. The genuinely new piece of work is llms.txt: a small file that did not need to exist before this category, and that no other Lighthouse category checks for at all.
A minimal one just needs the shape a Markdown-link-parsing model can confidently read: a short description of what the site is, who runs it, and a set of real Markdown links to the primary sections. Mine lives at devstash.me/llms.txt if you want a working example of the shape.
⚠A mistake worth naming directly
A bare URL like https://example.com/blog in an llms.txt file is not obviously wrong to a human
reader, but it fails the Markdown-link check specifically — the Lighthouse audit is checking for
the [text](url) syntax, not just the presence of a URL string. If you generate this file with an
LLM, tell it explicitly to use real Markdown links, since a model asked for "a list of important
pages" will often default to plain URLs.
The broader shift this signals#
Search engines built their own crawling and ranking logic over roughly two decades. Agentic Browsing is an early, narrow instance of a browser vendor building tooling for a different kind of automated visitor — one that acts on the page rather than just indexing it. The three checks that exist today are a reasonable starting set, not a finished spec, and I would expect this category to grow as agentic browsing itself becomes more common and its failure modes become better understood.
For now, the practical takeaway is small and concrete: keep your accessibility tree clean, keep your layout stable, and give your site an actual llms.txt with real links in it. All three were good practice already. This category just gave them a number you can watch.
Adesh Shukla
Frontend developer with a design background. Building DevStash — a developer ecosystem covering automation, AI workflows, and modern frontend systems.