Lighthouse Agentic Browsing: how to read the results

In short

• Agentic Browsing is a new Lighthouse category from version 13.3. Google describes it as under development, so its contents may change.
• Three audits count towards the score: a well-formed accessibility tree, layout stability (CLS), and a valid llms.txt.
• The WebMCP audits and ai-catalog.json are informative for now. They appear in the report but do not lower the score.
• Most of what you fix here also helps people. Good labels on buttons and form fields, and a page that does not jump while it loads, are ordinary good practice.
• A high score does not mean an agent gets the job done. Lighthouse looks at how the page is built, but does not try to complete a task.

How to run the test

In recent versions of Chrome, the category is in the Lighthouse panel in DevTools. Tick Agentic Browsing and run the analysis. From the command line, you can run just this category:

npx lighthouse https://www.yourcompany.com/ --only-categories=agentic-browsing --view

Run the test on the pages that matter most: the home page, a service or product page, and the page with your contact form, order, or booking. The result applies to one page at a time.

What counts towards the score

We ran Lighthouse 13.5 against our own website and looked at how the audits are weighted in the report:

Audit	Counts towards the score?	What it looks for
Accessibility tree is well-formed	Yes	Buttons, links, and form fields have names and correct roles
Cumulative Layout Shift (CLS)	Yes	Content does not move while the page loads
llms.txt follows recommendations	Yes	/llms.txt exists and is valid Markdown with at least one H1
WebMCP form coverage	No, informative	Forms missing WebMCP annotations
WebMCP tools registered	No, informative	Which WebMCP tools the page offers
WebMCP schemas are valid	No, informative	That the tools are described correctly
Valid ai-catalog.json	No, informative	A manifest that lets agents and registries discover your services

The weighting may change in later versions, so it is worth looking at each audit rather than just the number.

The accessibility tree

Agents that use the browser often read the page through the accessibility tree, the same structure screen readers use. That is where they find buttons, links, and form fields. An icon-only button without a name, or a form field without a label, is as hard for an agent as for someone who cannot see the screen.

How to fix it:

• Give every form field a visible <label>, not just placeholder text.
• Give icon buttons a name with aria-label.
• Use real <button> and <a> elements instead of clickable <div>s.
• Check that ARIA attributes point to elements that exist.

This audit builds on the same rules as the accessibility category. If you have fixed your accessibility issues, you are usually most of the way there.

Layout stability (CLS)

An agent looks at the page, decides where to click, and clicks. If the content moves in the meantime, for example because an image, an ad slot, or a consent banner appears, it hits the wrong element. It is the same problem as when you tap the wrong link because the page jumped.

How to fix it:

• Set width and height on images and video.
• Reserve space for content that loads later.
• Avoid banners that push content down after the page is shown.
• Use font-display and preloading so fonts do not cause large shifts.

llms.txt

The audit is simple: the file must exist at /llms.txt, be Markdown, and contain at least one level-1 heading. If the file is missing entirely, the audit is “not applicable”. If the server returns an error, it counts against you. We have written a separate guide on what llms.txt is and how to create one.

WebMCP

WebMCP is a proposed web standard in which the website itself describes what an agent can do, such as “search products” or “request a quote”, as tools with a name, a description, and fields. The agent then does not have to guess from buttons and layout. Tools can be defined with annotations on ordinary HTML forms or with JavaScript.

The standard is new. It is available in Chrome as an origin trial from version 149 and behind a flag for developers. That is why it does not count towards the score yet. Lighthouse still lists forms that lack annotations, which is a good overview of where to start.

ai-catalog.json

The newest audit looks for an ai-catalog.json manifest following the Agentic Resource Discovery specification. The idea is that agents and registries can discover and verify the services you offer. The audit is informative, and most websites get “not applicable”. We would not prioritise it until the standard has settled.

Our own result

Page	Agentic Browsing	Notes
Home page	99	CLS 0.069. The contact form lacked WebMCP annotations.
The llms.txt guide	100	No forms, no layout shifts

After the first run, we added WebMCP annotations to our contact forms: a name and description on the form and a description on each field. We did not enable automatic submission, so a person always presses “Send”.

When we looked at the tool Chrome built from the form, we found a trap: the hidden spam honeypot, a field called website, was included as a parameter. An agent filling in the form would likely have completed that field too, and the enquiry would have been discarded as spam. Fields marked readonly are kept out of the tool but are still submitted with the form, so we made the honeypot read-only. If your form has a spam field, it is worth checking.

What a high score does not tell you

Lighthouse examines how the page is built. It does not open an agent and ask it to book an appointment, fill in a form, or find a price. A page can score 100 and still stop an agent: a step requires signing in, a calendar only works with a mouse, a button says “Next” without saying what happens, or the payment is handled by a third party that blocks bots.

Our own home page scores 99, but that number says nothing about whether an agent can actually send our contact form. The only way to find out is to try. We have written a checklist of what most often stops agents.

Want to know whether an agent actually gets the job done?

The Lighthouse result is a good starting point, but only a test with an agent shows what happens in practice. An agent test is part of Foretrukket (in Norwegian), our AI-visibility service for Norwegian businesses. It helps them get recommended when customers ask ChatGPT and Google.