On this page
Chrome's Lighthouse — the audit built into DevTools that every web developer has run at least once — now has a category called Agentic Browsing. It grades one thing: how reliably an AI agent could use your page. As Chrome puts it, "for an AI agent to reliably complete a flow on your site, such as booking an appointment or placing an order, it needs predictable, machine-readable signals" (Chrome developers blog).
Here is what it checks, what the result means, and what to do with it — without needing a developer for most of it.
What the four checks actually test
| Check | What passes | Why an agent cares |
|---|---|---|
| Agent-centric accessibility | Every interactive element has a programmatic name; the accessibility tree is intact | The tree is how agents read a page — an unnamed icon button does not exist to them |
| Stability (CLS) | Content does not jump while loading | Agents click coordinates and read screenshots; a moving layout causes misclicks |
| WebMCP integration | Registered tools are present, forms carry declarative annotations, the schema is valid | This is what lets an agent do things, not only read |
| llms.txt | "a machine-readable summary at the domain root" (Chrome docs) | A one-page brief beats guessing which of your pages matters |
The llms.txt audit is stricter than "the file exists": it flags a file that is "missing an H1 header, is too short, or doesn't contain any links" (DebugBear).
Why there is no 0-100 score
Unlike Performance or SEO, this category does not produce a grade. Chrome returns "a fractional score: a ratio showing how many agentic readiness checks your site passes" plus a pass or fail per audit, and describes the category as informational and unbenchmarked — "actionable signals … rather than a definitive ranking".
Two honest consequences:
- A green is not a certificate. A plain brochure page with no AI features can pass everything it is eligible for; DebugBear's example is blunt — "example.com gets a green 2/2 score". The number is not the point; the failing rows are.
- The bar will move. The category is still marked under development, and the WebMCP audits only apply once your site actually registers tools.
Run it in five minutes
You need Chrome 150 or later. The command line version is the most reliable today, since some hosted runners still ship an older Lighthouse:
npm install -g lighthouse@latest
lighthouse https://[your-site] --only-categories=agentic-browsing --viewThat opens a report with the four checks and, more usefully, the exact elements that failed. If a command line is not your world, check any URL here — it reports the same agent-facing signals for a page, no install required, and inspect a site shows which tools a page exposes to an agent right now.
Fix in this order
- Name your controls. Icon-only buttons, unlabelled inputs, links that say "here". Highest value, smallest effort, and it fixes screen-reader accessibility at the same time.
- Stop the jumping. Reserve space for images, banners and cookie bars so the page does not shift under an agent mid-click.
- Publish an llms.txt with an H1, a real summary and links to the pages that matter. Ours is at /llms.txt if you want a shape to copy.
- Expose your key actions as tools. Booking, contact, order lookup — the adoption steps and the page tools guide cover both the declarative form annotations and the imperative version.
Ask your assistant to work the report
Do not read the JSON yourself. Paste this into any AI assistant that can browse:
Run or read a Lighthouse agentic browsing report for https://[your-site].
List every failing audit in plain English, sorted by how much it would hurt an
AI agent trying to [book an appointment / buy a product] on my site.
For each one, tell me exactly which element or file to change and give me the
replacement text or code. Mark anything that needs a developer with [CHECK].You will usually get a list of fewer than ten fixes, most of them text.
What a passing audit still does not prove
The audit proves your page is legible and stable. It cannot prove your page is useful: that the answers a shopper needs are actually written down, that your prices are current, that an agent can complete the booking rather than merely see the form. That part is content and capability, and it is what WebMCP plus a good knowledge base give you.
If you would rather see the outcome than the report, point us at your site and watch your mate answer real questions about it in a couple of minutes — then run Lighthouse again and compare. Plans are on pricing.
Questions
Is agentic browsing part of my Lighthouse performance score?
No. It is a separate category and it does not produce a 0-100 number at all. Chrome describes it as informational and unbenchmarked, reporting a fraction of checks passed plus a pass or fail per audit rather than a ranking.
What do I need to run it?
Chrome 150 or later. The quickest route is the Lighthouse command line tool with the agentic-browsing category; the WebMCP audits additionally need your site registered for the WebMCP origin trial.
Does a green result mean my site is agent-ready?
Not on its own. A plain page with no AI features at all can pass the checks it is eligible for, so treat a green as the floor, not the finish line — the useful signal is which individual audits are failing.
Does this affect my Google Search rankings?
There is no evidence that it does, and Google has not said it does. Treat the category as a readiness diagnostic for AI agents, which is a different audience from the search index.
Why does layout shift matter to a robot?
Because agents click by coordinates and screenshot what they see. A layout that moves under them causes misclicks, which is why cumulative layout shift is one of the checks.
What is the single highest-value fix?
Give every interactive element a real, programmatic name. Buttons labelled only with an icon are invisible to the accessibility tree that agents read, and that one fix improves the experience for screen-reader users at the same time.