Chrome now audits your site for AI agents

In May, Chrome shipped an audit that grades your website on how well a piece of software can use it. Not read it. Use it. The category is called Agentic Browsing, it sits in the Lighthouse tab of the browser most of your buyers already have open, and you can run it against your own domain in about a minute.

I have been building for the web since 1994. I have watched this exact move three times now. Something the industry argues about becomes something the browser measures, and the argument ends. It happened with mobile. It happened with HTTPS. It happened with page speed, twice, badly the first time. The measurement always lands before the consensus does, and it always decides the consensus.

What the audit actually checks

Read the Chrome documentation before you read anyone's summary of it, mine included. Four things get examined. Whether your page registers WebMCP tools, which is how a page tells an agent what it can do. Whether the accessibility tree is intact, with real names and labels, since that tree is how an agent perceives the page at all. Cumulative Layout Shift, because a control that moves after render is a control an agent clicks and misses. And whether an llms.txt file exists at your domain root.

Two details matter more than the list. Chrome states that the category and its WebMCP support are experimental and based on proposed standards. And unlike every other Lighthouse category, this one does not produce a weighted score from 0 to 100. You get a fraction, a count of audits passed against audits run. That is a deliberate choice by people who know the number would be quoted in a pitch deck within a week of existing.

The part that confuses everybody

At the same time as this shipped, Google Search kept saying the opposite. Its guidance on AI features is blunt: you do not need to create new machine readable files, AI text files, markup, or Markdown to show up in generative AI search. John Mueller has called llms.txt a temporary crutch, useful for helping coding tools chew through developer docs and not much else.

I keep seeing this framed as Google contradicting itself. It is not. One team is answering “does this change your ranking,” and the answer is no. The other is answering “can an agent operate this page,” and the answer depends entirely on how the page was built. If you think those are the same question, you are still thinking about a web where the visitor is a person with eyes.

The honest position is unglamorous. Publish the file because it costs an afternoon and might matter later. Fix the accessibility tree because it definitely matters now, both for agents and for the humans it was built for in the first place. Do not let anyone charge you a retainer for either.

What nobody is doing yet

Here is the gap that bothers me. You can run this audit on your own site this afternoon and get a fraction back. Seven of eleven, say. And then you are stuck, because there is no published set of results for anyone else. Seven of eleven against what? Your competitor down the road might be at two. They might be at eleven. Nobody has looked, nobody has published, and a score with nothing to compare it against is not a finding. It is a number.

Every measurement layer the web has adopted followed the same path, and the comparison always arrived before the standard was finished. Accessibility had audits and lawsuits before it had conformance levels people agreed on. Performance had a public scoreboard before anyone settled what a good score meant. The scoreboard is what forces the definition, not the other way around.

So somebody is going to publish the first cross company comparison of agent readiness. It will be partial and it will be argued with. It will also become the definition, because the first published measurement usually does, and every firm named in it will have to respond to a number they did not choose.

What to do this week

Open DevTools on your own site, run the Agentic Browsing category, and write the fraction down with the date next to it. That is your baseline and it takes a minute. Then read the failures rather than the number, because the failures are specific and fixable in a way the fraction is not.

Then go a layer deeper than the audit goes. Lighthouse checks whether the file exists. It does not check whether your site tells a consistent story about who you are, which is the thing that actually determines what a model says about you. I wrote the longer version of that check in the machine readable website audit, and the next post in this series turns it into a scored rubric you can run yourself.

The uncomfortable part is that none of this tells you whether an AI system recommends you. Conformance and citation are separate measurements, and passing the first one guarantees nothing about the second. That is the third post.

Common questions