Does a good Lighthouse Agentic Browsing result improve AI visibility?

There is no evidence that a good Lighthouse Agentic Browsing result improves AI visibility: the category measures how well AI agents can read and operate a page, not how often AI systems mention or cite a website, and in June 2026, Google described it as informational and unbenchmarked. What speaks for the audited points is that they make a website easier for agents to use, and some of them help people as well. Whether an offering appears in AI answers can only be shown by measuring the answers themselves.

Why can the Lighthouse Agentic Browsing result vary between two runs?

According to Google, the Lighthouse Agentic Browsing result can change when a website registers WebMCP tools via script (JavaScript) at different moments, when larger changes to the page structure alter the accessibility tree, or when ads, images without dimensions, and late-loading content shift the layout. In addition, PageSpeed Insights and Chrome DevTools do not always run the same Lighthouse version, so even the number of audits can differ.

Does Agentic Browsing replace Lighthouse’s Accessibility category?

No. Lighthouse Agentic Browsing checks only a subset of accessibility rules, namely those that are critical for machine interaction. Lighthouse’s Accessibility category checks more, such as color contrast, and remains the more comprehensive automated check of a page’s accessibility.

Lighthouse Agentic Browsing: what it checks

Lighthouse Agentic Browsing has been a category of Lighthouse since May 2026. Lighthouse is Google’s open-source tool for auditing web pages; it is also built into Chrome DevTools, the web development tools included in Chrome, and into PageSpeed Insights, Google’s free web-based testing page. The category focuses on a different kind of visitor: AI agents that complete tasks on websites on people’s behalf, such as booking an appointment or placing an order. To get this done, Google says, they need predictable, machine-readable signals: they have to recognize buttons and form fields and operate them reliably. Google describes the category as experimental and based on proposed standards.

What the category checks

Accessibility for agents

According to Google, agents use the accessibility tree as their main source of information: the simplified version of a page that the browser derives from the HTML and that screen readers (software that reads pages aloud for blind and visually impaired people) use as well. Lighthouse therefore checks a subset of accessibility rules that are critical for machine interaction. These include whether every button, link, and form field has a name that software can read, whether ARIA attributes (extra information in the HTML for assistive technologies) are used correctly, and whether the page has a title. Buttons and form fields usually get their name from their visible text or from an associated label in the HTML, the label element, which is why semantic HTML often provides it automatically. This is where the audit overlaps with web accessibility.

Layout stability (CLS)

Agents often work with screenshots and click on a specific position on the screen. If the page shifts after an agent has located a button, Google says, the agent can miscalculate its position and miss. Lighthouse measures this with Cumulative Layout Shift (CLS), a metric for how much the visible elements of a page move unexpectedly. Google considers a value of 0.1 or less good and a value above 0.25 poor. Typical causes are images and videos without dimensions; ads, embeds, and late-loading elements without reserved space; and web fonts that render larger or smaller than the fallback font shown while they load. An image with width and height attributes, for example 640 and 360 pixels, tells the browser in advance how much space to keep free.

llms.txt

Lighthouse requests the file /llms.txt from the root of the domain. An llms.txt file summarizes a website’s most important content for language models and AI agents. If the file is missing, the audit counts as not applicable; in May 2026, Google described the file as optional for the time being. The audit fails if the file cannot be fetched, for example because of a server error, or if its content misses one of three minimum requirements. Lighthouse expects Markdown, a simple way of writing formatted text, with at least one level-one heading such as “# Example Company,” at least one link with the link text in square brackets followed directly by the address in parentheses, and at least 50 characters.

WebMCP

WebMCP is a proposed web standard that lets a website offer its functions to AI agents as tools with a name and description instead of leaving agents to work out from buttons and fields what each function does; Chrome has been testing it since Chrome 149 in an origin trial, a time-limited test phase that websites register for. In the simplest case, an existing HTML form gets two attributes for this: toolname, for example with the value book_appointment, and tooldescription with a sentence such as “Books a consultation.” Lighthouse lists the registered tools and flags forms that have no WebMCP attributes, without counting either as an error; since Lighthouse 13.5.0, it also warns when more than 40 tools are registered. A third audit checks whether the attributes are complete: if a form has only one of the two attributes, or a required field has no name attribute, the audit fails. If only optional fields lack details, such as a description in the toolparamdescription attribute, Lighthouse issues warnings, and the audit still does not count as passed.

Catalog for AI agents

Lighthouse 13.5.0 added another audit in September 2026. It validates a catalog under the Agentic Resource Discovery (ARD) proposal, which lets providers describe tools and services for AI agents in a single file. By default, Lighthouse 13.5.0 looks for this file under the name ai-catalog.json; the August 2026 version of the proposal names it ard.json and lists ai-catalog.json as its earlier name. If Lighthouse finds no catalog, the audit counts as not applicable.

How to run it

PageSpeed Insights: enter the address of a page and run the analysis. The category appears in the report next to the other Lighthouse categories; nothing needs to be installed.
Chrome DevTools: from Chrome 150, open the Lighthouse panel and, under “Categories,” check the “Agentic browsing” box, which is unchecked by default. DevTools can also audit pages that require a login.
Lighthouse command line: Lighthouse as a program without a browser interface that developers can run automatically, for example with every new release of the website. Here, too, the audit requires Chrome 150 or later. The category has been part of the default configuration since Lighthouse 13.3.0; Google describes the audits as reproducible and suitable for automated workflows.

How to read the result

Unlike the other Lighthouse categories, Agentic Browsing has no score from 0 to 100. The report shows a fraction such as 3/4, meaning three of four applicable audits passed; individual audits can raise warnings. Audits that do not apply to a page are listed under “Not applicable” and are not counted. That includes WebMCP: a page without WebMCP tools loses nothing, because those audits are then listed as not applicable or as informational only. Google attributes this approach to the early stage of the field: the standards for AI agents on the web are still emerging, so for now Google wants to collect data and give actionable pointers rather than issue a final verdict on websites, and the report labels the category as still under development.

Two of the audited areas benefit not only agents but every visitor: labeled controls help screen reader users, and a stable layout keeps everyone from clicking the wrong link or button. Because the audits follow fixed rules, the result per audit is also useful for recording progress, for example before and after a change to the site’s templates; it can still vary, mainly when the page itself changes while it loads. The building blocks that make a website accessible to agents as a whole are described under agent-friendly website.

Lighthouse Agentic Browsing

Definition