Chrome now grades your site for AI agents. Three of its six checks are for a standard almost nobody uses.

RBy Ronnie Huss, founder of SearchScore

Google shipped an agent-readiness category into the tool most of the web already uses to check itself. That is a larger event than the individual audits inside it, and the individual audits are what everyone has written about.

Here is what the category actually contains, read from a run rather than an announcement.

The interesting half is the half that does nothing

Three of the six audits test WebMCP, a proposal for websites to expose actions to an agent directly rather than having the agent read the page and work out which button to click. On any site that has not implemented it, all three return Not Applicable. They do not fail. They do not reduce the score. They sit there.

The same is true of the llms.txt audit when the file is absent, because publishing one is optional. A server error while fetching it is flagged, which is a different thing and a sensible one.

So the category as shipped cannot meaningfully penalise a site for lacking agent features. That is worth stating plainly, because it is the opposite of how a new Lighthouse category is usually read, and it changes what a sensible response looks like.

Measurement arrives before the requirement, and it always has

This is a familiar sequence and it is worth naming, because the timing is the actionable part rather than the audits themselves.

Core Web Vitals were measurable in Lighthouse well before they were a ranking signal. Mobile-friendliness was a Search Console report before it changed anything. HTTPS was reported before it was rewarded. In each case the gap between Google being able to see something and Google acting on it was long enough to be comfortable and short enough to matter, and the firms that moved during the gap were not the ones scrambling afterwards.

The inert checks are therefore the most informative part of the release. They are a published list of what Google intends to be able to grade. WebMCP occupying half of it is not a rounding detail: it is the largest single statement in the category about what Google thinks an agent-ready site looks like.

Two Google positions, and they are not in conflict

Google Search says an llms.txt is not needed. Its AI optimisation guidance of 15 May 2026 says so directly, for AI Overviews, AI Mode and every generative Search feature. Google Chrome now audits the file. Several write-ups this week have framed that as the company contradicting itself.

It is not a contradiction. It is two jobs.

Answering a question is retrieval. An engine fetches pages, selects passages and produces an answer. A summary file at the root does nothing for that, which is why Search says it is not needed and why no provider has confirmed reading one at inference time.

Completing a task is navigation. An agent has to work out what a site is, which pages matter, and how to act on them. A map at a known location is genuinely useful for that, and a set of declared tools is more useful still. That is the job Chrome is measuring.

Read that way, llms.txt is becoming a standard for the second job while remaining irrelevant to the first. The practical consequence is that it should be argued for on agent readiness and not sold as an AI search tactic.

What our own index says about the file

We measured llms.txt across the SAVI index panel, 102,873 domains weighted to the population, before Chrome shipped this category.

The bottom two rows are the ones that matter. A fifth of the least ready sites on the index already publish one, so as a marker of quality it barely discriminates. That was our finding in the pre-registered study before this release and it is unchanged by it: as a search signal the file is weak and getting weaker as it spreads. What has changed is that it now has a second job where the case is stronger, and a checker that grades it.

Those figures were measured on presence rather than compliance, because until this month there was no published definition of compliance to measure against. Chrome has now supplied one: Markdown, at least one H1, at least one link, and bare URLs do not count. How many of the existing 27.3% would satisfy that is not something anyone has published. We intend to.

What this does not show

Nothing here is evidence that an AI agent reads an llms.txt, or that anything consumes WebMCP at inference time. Lighthouse checks whether a site publishes these things. Whether anything uses them is a separate question, and on the current public record no major provider has confirmed that it does.

Treat this as a statement of intent by a browser vendor, which is worth acting on, rather than as a measured effect, which it is not. The distinction matters because the industry has spent this year selling the second while only the first is in evidence.

What is proportionate to do now

Publish an llms.txt and publish it correctly. It takes an afternoon, and correctly now has a definition: a Markdown document, one H1 naming the brand, a short summary, then your key pages as Markdown links. A file with bare URLs fails, which is not obvious and is the most common way to get it wrong. Our llms.txt guide has the format and worked examples.

Fix your accessibility tree, because that audit is not really about agents. A well-formed tree is what a screen reader uses, and it has been the right thing to do for twenty years. The agentic framing is new; the work is not.

Do not build WebMCP yet. Three inert audits are a signal to watch, not a specification to implement, and the standard has neither confirmed consumption nor meaningful adoption. Watching costs nothing. Building costs a quarter, and the audits will tell you when they start scoring.

The firms that will be ready are not the ones implementing an unratified standard this month. They are the ones who noticed that the browser started measuring, and who diarised the moment Not Applicable turns into a number. You can see where your own site currently sits with the free audit.

Method

Lighthouse 13.4.1 run headless against a live production site with --only-categories=agentic-browsing on 19 August 2026, and the audit list read from the resulting JSON report rather than from documentation. The category is marked as under development in its own description, so its contents may change; this is what it contained on that date. Anyone can reproduce it in one command on a machine with Chrome installed.

Panel figures are the SearchScore SAVI index panel, wave 1, dated 1 August 2026: 102,873 domains on scorer geo-145-2026-07, weighted to the population. The llms.txt measure there is presence, not compliance. Our measurement standards set out what these numbers are fit for.

Our founder ran the same audit against his own properties and wrote about what it found, including what it failed, why, and why one of the two failures is not going to be fixed.