Chrome’s Lighthouse now audits your site for AI agents: what it scores and misses

Published June 28, 2026 · Last reviewed June 28, 2026

Google just turned “can an AI agent use this website” into a test you can run yourself. With Chrome 150 and Lighthouse 13.3, the audit tool built into Chrome DevTools added a new category called Agentic Browsing, and a companion product, Chrome DevTools for Agents, reached a stable 1.0 release.¹² Run Lighthouse on any page now and you get a section that grades how well your site is built for software that acts on a page, not just software that reads it.

This is a real signal, and not quite the one most teams will assume. The new category does not hand you a number out of 100, and Google is explicit that it is informational rather than a ranking at launch. It runs a handful of checks and reports how many your site passes, because the standards for the agentic web are still being written.³ That restraint is the honest part. It is also the tell: the audit grades how readable your site is to an agent, which is only half of what an agent actually needs from you.

Here is what Chrome shipped, what the audit measures, and the gap between a site an agent can read and a site an agent can reach.

What Chrome actually shipped

Two things landed together. The first is the Agentic Browsing category in Lighthouse, the open-source auditing engine that powers the report under the Chrome DevTools Lighthouse panel. It shipped in the default configuration with Lighthouse 13.3 in early May 2026, and on Chrome 150 or later it appears automatically, with no flag to flip.¹³ Open DevTools, run Lighthouse, and alongside Performance, Accessibility, and SEO you now see a section asking whether your site is legible to a machine acting for a person.

The second is Chrome DevTools for Agents, which Google moved to a stable 1.0. It comes in three forms: a Model Context Protocol server that connects a coding agent to DevTools, a command-line interface for batching actions into scripts, and a set of agent skills that teach an agent when to reach for a given tool.² It can list and execute the WebMCP tools a page exposes, and experimental WebMCP debugging tools, added to the Chrome 149 DevTools Application panel sidebar, let a developer watch tools register, validate their schemas, and inspect what an agent invoked.⁴

Read together, the message from the Chrome team is plain. Non-human interaction with a website is now a first-class thing to build for, test, and debug, with the same tooling weight that performance and accessibility have carried for years. That framing matters more than any single check, because it tells you where the platform is heading. The question is no longer whether agents will read your site. It is whether your site was built expecting them.

What the Agentic Browsing audit scores

The category checks four things, and the choice of four is itself a small map of how Google sees the agentic web today. It looks for an llms.txt file, for WebMCP tools the page registers, for a clean accessibility tree, and for layout stability measured as cumulative layout shift.³⁵ Two of those are about exposing structure to agents, and two are about not breaking when an agent reads the page the way assistive technology does.

The scoring is where the restraint shows. Unlike Performance or SEO, Agentic Browsing does not roll up to a weighted score from 0 to 100. It reports a fraction: how many of the readiness checks your site passes. Google’s own documentation says the goal at launch is to gather data and surface actionable signals, not to rank sites, because the underlying conventions are still moving.³ So a passing audit is a checklist completed, not a verdict that agents can do business with you.

The honest reading of each check helps here. The llms.txt convention has roughly 10 percent adoption, but an SE Ranking analysis across about 300,000 domains found it produced no measurable lift in AI citations, and no major model provider confirms reading it at runtime.⁵ Treat its presence as cheap hygiene the audit happens to reward, not a visibility lever. WebMCP is more substantial but earlier than the headlines suggest: it is a draft in the W3C Web Machine Learning Community Group, exposed through document.modelContext, and it is running as a Chrome 149 origin trial (announced May 19, 2026). It is not a ratified W3C standard, and it is not in stable Chrome by default.⁴ The two checks that will age best are the quiet ones. A clean accessibility tree and a stable layout are things an agent reads the same way a screen reader does, and they were good engineering long before agents arrived.

Readable is not the same as reachable

Pass every check in the new category and you have made your site agent-readable. That is worth doing. It is not the same as making your site agent-addressable, and the difference is the whole point.

Agent-readable means the structure is there to be parsed. An llms.txt index, WebMCP tools, a clean accessibility tree: these let an agent extract what your page already says. Agent-addressable means there is a representative an agent can find, talk to, and get a real answer from, one that can qualify whether the conversation is worth a human’s time and route it onward when it is. The first is about your markup. The second is about an interaction your markup cannot have.

Picture the visit. A buyer’s agent arrives with a specific question its human asked thirty seconds ago: do you support SSO for a 40-person team, is there regional data residency, what does onboarding look like. A site with a perfect Agentic Browsing score still answers that question the only way a static page can, which is by hoping the answer happens to be written somewhere the agent can reach. If it is not, the agent has two dead ends. It can try the contact form, which acting agents either cannot complete or should not, or it can wait for sales follow-up, and an agent comparing several vendors in parallel does not wait. The reading surface was immaculate. The answering surface was never there.

This is not a flaw in Lighthouse. The audit grades what it can grade, which is the page. It is a reminder of what a page-level audit structurally cannot see.

What still matters once an agent is on your site

Once an agent can read and even operate your site, three questions move to the front, and none of them is a markup property.

The first is identity. When an agent reaches a tool or a form on your domain, who does that site speak for? Cloudflare’s own April 2026 readiness scan of the 200,000 most-visited domains found that only about 4 percent declare any AI preferences in robots.txt and fewer than 15 expose a machine-readable capability card, so the practical answer today is usually nobody in particular.⁶ A human-readable name tied to a real person or company is what lets the agent on the other side know whose interests it is dealing with.

The second is consent. A fully authenticated, fully capable agent can still be one you do not want to hear from, because sending is nearly free. The scarce thing on an agentic web is not access, it is permission to reach the human behind the site. When millions of agents can knock politely and correctly, a site needs a way to say yes to the worthwhile ones without exposing a person to all the rest.

The third is routing. Turning an agent’s question into a qualified conversation, and escalating to a human only when the fit is real, is work that happens after the page loads and after any tool runs. It depends on context the agent brought and judgment about whether a person should step in. Lighthouse cannot score any of this, because identity, consent, and routing live in the interaction, not the HTML. They are exactly the half that an agent-readable site leaves open.

How this connects to Tobira

Tobira works on the half a page audit cannot see. A Tobira @handle gives the agent on your site a human-readable identity tied to a real person or company, and a Site Agent makes the site addressable: it can answer a visiting agent’s questions, qualify fit, and route a worthwhile conversation to the owner, with mutual-reveal consent so a real person is reached only when both sides agree. That is complementary to the readable layers Lighthouse now checks, not a replacement for them. A high Agentic Browsing score and an addressable representative solve different halves of the same visit, and the clearest way to see why both matter is to follow what agentic browsers actually change for your website.

FAQ

What is the Lighthouse Agentic Browsing category?

It is a new category in Chrome’s Lighthouse audit tool that checks how ready a page is for AI agents to read and act on. It shipped in the default configuration with Lighthouse 13.3 in early May 2026 and appears automatically on Chrome 150 or later. Instead of a score from 0 to 100, it reports a fraction of readiness checks passed, and Google describes it as informational rather than a ranking because agentic-web standards are still emerging.

Does a high Agentic Browsing score mean AI agents can use my site?

It means agents can read your site well, which is not the same as being able to do business with it. The category grades readable signals such as llms.txt, WebMCP, the accessibility tree, and layout stability. It cannot see whether a visiting agent can ask a real question and get an answer, qualify fit, or reach a human. A clean score and an answering layer solve different problems.

Is WebMCP a web standard yet?

No. WebMCP is a draft in the W3C Web Machine Learning Community Group, exposed through document.modelContext, and it is running as a Chrome 149 origin trial (announced May 19, 2026). It is not a ratified W3C standard and not enabled in stable Chrome by default. Build with it as an early preview, not a settled platform feature.

Does adding llms.txt improve my AI visibility?

The new audit rewards having one, but the evidence on visibility is weak. Adoption sits near 10 percent of domains, an SE Ranking analysis across roughly 300,000 domains found no measurable lift in AI citations, and no major provider confirms reading it at runtime. Treat it as cheap hygiene that the Lighthouse check happens to credit, not as a lever that changes what agents do.

What is the difference between agent-readable and agent-addressable?

Agent-readable is about how agents read your site: structure, tools, and clean markup, which is what Lighthouse audits. Agent-addressable adds the other half, a representative agents can find and talk to that answers, qualifies, and routes to a person, tied to a human-readable name so a real person is reached only with consent. The two are complementary layers, and neither substitutes for the other.

Sources