Google Says You Do Not Need llms.txt. Lighthouse Audits It Anyway.

Google's position on llms.txt depends entirely on which part of Google you ask. Search Relations has been consistent and blunt: the file does nothing for rankings and is not used by Google's AI features. Meanwhile Chrome shipped a Lighthouse audit that checks whether you have one.

This reads as a contradiction and is not one. They are answering different questions, and the gap between them is the most useful thing to understand about where the web is heading.

What Agentic Browsing is

Lighthouse gained a new category alongside Performance, Accessibility, Best Practices and SEO: Agentic Browsing. Where the SEO category asks whether a search engine can index your page, this one asks a different question — can a software agent operate this site on a person's behalf?

It is explicitly experimental, built on proposed standards, and requires a recent Chrome. It is a developer diagnostic, not a ranking report.

The llms.txt audit

Modest in what it does. It requests /llms.txt and:

• Not applicable if there is no file — publishing one is optional, and its absence is not a failure.
• Flags a server error if the request returns 5xx.
• Does not grade the file's structure or content.

So it will not tell you your sections are thin or your links are stale. What it does tell you is that Google considers the file a legitimate part of an agent-readiness checklist — which is a notable shift from treating it as noise.

The WebMCP audits

The more interesting half of the category. WebMCP is a proposed standard that lets a page expose structured tools to an agent, so the agent invokes a defined action instead of inferring what a button does from its label and hoping.

Two ways to declare a tool. Declaratively, on a form:

<form toolname="search-products"
      tooldescription="Search the product catalogue by keyword">
  <input name="q" type="search">
</form>

Or imperatively, in JavaScript:

navigator.modelContext.registerTool({
  name: 'search-products',
  description: 'Search the product catalogue by keyword',
  // ...
})

Lighthouse runs three checks:

• Registered tools — lists what the page exposes. Informational.
• Forms missing declarative WebMCP — forms an agent would struggle with. Informational.
• Schema validity — the only one that can actually fail.

The category also folds in two things you already care about: accessibility tree integrity and layout stability (CLS). That is not padding. An agent reads your page through the accessibility tree, and a layout that shifts under it causes the same misclicks it causes a human — which is a neat argument that accessibility work has always been agent-readiness work.

Reconciling the two Googles

Search is answering: does llms.txt affect how Google ranks or summarises you? The answer is no, and the evidence supports it — studies of crawler logs consistently show that GPTBot, ClaudeBot, PerplexityBot and the rest overwhelmingly fetch HTML directly and skip /llms.txt.

Chrome is answering: is this site built for software that browses on a person's behalf? There, a curated index of your best content is a reasonable thing to have, because an agent that finds one spends less time crawling to work out what your site is.

Both can be true because they are about different consumers at different times. Publishing llms.txt as an SEO tactic will disappoint you. Publishing it because agents that do read it get a better answer about you is defensible — and it costs an afternoon.

What to actually do

1. Make sure /llms.txt serves a 200, not a 5xx and not your SPA shell. That is the whole Lighthouse check.
2. Then make it good — Lighthouse will not, but agents that read it will notice. Descriptions matter more than ever under v2.
3. Fix your accessibility tree. Real users first; agents get it for free.
4. Watch WebMCP, do not rush it. It needs an origin trial and the spec is moving.
5. Do not expect rankings. Google has been clear, and pretending otherwise sets up a disappointment.

One trap worth knowing

A single-page app will often return 200 OK with Content-Type: text/html for any path, including /llms.txt. The audit sees a successful response. An agent sees your JavaScript shell where a Markdown index should be.

This is why our validator warns when a fetched file arrives as text/html — it is the clearest signal that a soft 404 has been served in place of the real file.