Lighthouse Agentic Browsing is a new category in Google Lighthouse that, per publicly available documentation, showed up in version 13.3 in May 2026. It covers four areas that shape how tools read a page by machine: whether llms.txt is present, support for the WebMCP protocol, the quality of the accessibility tree, and layout stability (CLS). The score isn’t weighted from 0 to 100 — it shows the share of checks passed, and each one clearly reads as pass or fail. You can run the audit in Chrome DevTools, in PageSpeed Insights, or through the Lighthouse CLI (command line). This guide explains what the audit tests, how to run it, and what order to fix failed checks in.
What Lighthouse Agentic Browsing is
Per publicly available documentation and release tracking, the Agentic Browsing category showed up in Lighthouse 13.3 in May 2026. It started appearing as its own category in public builds of the interface, in PageSpeed Insights over the following weeks, with the Chrome 150 DevTools integration following on. The category measures how well a site is set up for AI agents — tools that can read a page by machine and, in some cases, take action on it.
Unlike the classic Lighthouse categories, it doesn’t use a weighted 0 to 100 score. It shows the share of checks passed instead (“3 of 4 checks passed,” for example). The category is new, and the audit still works with simpler pass/fail scoring. In practice that means every check tells you what’s missing and where to start fixing it.
The four areas the audit checks
Lighthouse Agentic Browsing tests four specific areas:
Checks in the Agentic Browsing category
- llms.txt A machine-readable summary of your content at the root of the domain (/llms.txt, for example). Lighthouse checks that it's present and has a basic structure.
- WebMCP A frontend protocol that lets an agent take specific actions on the site (add to cart, submit a form). Still experimental.
- Accessibility tree A specific subset of accessibility audits that matter to AI agents (programmatically detectable names, ARIA attributes).
- Layout stability (CLS) Cumulative Layout Shift — if elements move around, an agent and a visitor alike can target the wrong spot. The recommended value is CLS < 0.1.
1. llms.txt
A machine-readable summary of your content at the root of the domain — https://yoursite.com/llms.txt, for example. Lighthouse checks whether the file exists and has a basic structure. If it’s missing, that check fails. It’s usually the cheapest fix in the whole category — in most cases, a static llms.txt text file at the root of the site is enough. There’s a separate guide to llms.txt for the details.
2. WebMCP
WebMCP (Web Model Context Protocol) is described in the available documentation as an experimental protocol sites use to offer agentic tools specific actions. It differs from the usual MCP servers running on the backend — WebMCP runs in the visitor’s browser and describes what a tool can do on a given page (click, form, search).
As of 2026 the standard is still experimental and only a handful of agents actively support it. Implementing it mainly makes sense for online stores, booking sites, and SaaS apps, where an agent can realistically take action.
3. Accessibility tree
Per the available documentation, the accessibility tree is the key data model for agentic reading of a page. Agentic tools reading a page rely on the same structure screen readers use for blind users. Lighthouse checks a specific subset of accessibility audits, mainly:
- Names and labels — every interactive button, link, and input needs a readable description (text in the button,
aria-labelon an icon,<label>on an input). - ARIA attributes — correct
role,aria-label,aria-describedby.
Things like tab order and focus management relate to usability, but per the available documentation Lighthouse concentrates here mainly on names and labels and ARIA. Optimizing for AI agents often improves your classic accessibility score too and makes the site easier to use for people on assistive technology.
4. Layout stability (CLS)
Cumulative Layout Shift measures how much page elements move around after load. For a human visitor that means frustration (“I went to click the button and it moved”). An unstable layout can also make an action harder to complete — an agent and a user alike can target an element that has since shifted or moved.
Lighthouse targets CLS < 0.1, the same as classic Core Web Vitals. The optimization work is shared with the Performance category.
How to run a Lighthouse audit
There are three ways to run the audit. Chrome DevTools is the recommended one.
| Aspekt | Chrome DevTools (recommended) | PageSpeed Insights / CLI |
|---|---|---|
| Where it runs | Locally in the Chrome browser | PageSpeed Insights: the pagespeed.web.dev web interface. CLI: a local Node environment. |
| What it can audit | Deployed and local URLs, logged-in pages | PageSpeed: publicly reachable URLs only. CLI: local and public alike, good for CI/CD. |
| Category selection | Pick categories, device (mobile/desktop), and mode | PageSpeed: default settings. CLI: fully configurable through parameters. |
| Output | An interactive report right in DevTools, with a history of previous runs | PageSpeed: a web page with the report. CLI: an HTML / JSON file for archiving and CI/CD. |
| Good for | Developers, one-off audits, diagnosing a specific page | PageSpeed: a quick audit of a public page with nothing to install. CLI: automation in a CI/CD pipeline. |
-
Load the URL you want to audit
Open the page in Chrome. We recommend closing other tabs and extensions that could skew the results (ad blocker, password manager, screen reader). The audit should run in as clean an environment as possible.
-
Open the developer tools and go to Lighthouse
Right-click the page and choose Inspect, or press F12. In the DevTools panel, switch to the Lighthouse tab. If you don't see it, click the two arrows on the right (More tabs) and pick it from the overflow menu.
-
Check Agentic Browsing plus the other categories
Check every relevant category (Performance, Accessibility, Best Practices, SEO, Agentic Browsing). Pick the device (mobile for a consumer-facing site, desktop for B2B tools). Leave Navigation as the measurement mode.
-
Click the Analyze page load button and wait — typically tens of seconds
Lighthouse opens a fresh instance of the page in an isolated environment, walks it, and scores every category. Don't touch the page while the audit runs. When it finishes, you get a clear report of passed and failed checks.
-
Focus on the failed checks
Agentic Browsing doesn't give you the usual 0 to 100 score — it gives you the share of checks passed (“3 of 4 checks passed,” for example). Every individual check expands with specifics on what's missing and how to fix it.
-
Work from cheapest to heaviest
Start with llms.txt (the cheapest), then the accessibility tree (names and labels), then layout stability work (CLS). Leave WebMCP for last — it's a heavy lift, and it mainly pays off for online stores and apps where an agent can take action.
What order to fix failed checks in
Lighthouse Agentic Browsing hands you an immediate list of what to fix. For the most value per hour of work, we recommend the following order.
Auditing the homepage only
Running Lighthouse on the homepage alone underrates the state of the whole site. Audit at least 3 page types: the homepage, a category or listing page, and a product or detail page. Failed checks often differ between templates.
Optimizing without re-running the audit
Changes to llms.txt or the accessibility tree need to be verified with another audit. Without a re-run, you can’t confirm the fix actually worked and that Lighthouse registered it.
An old version of Lighthouse
If you’re on anything older than Lighthouse 13.3, the Agentic Browsing category is either missing or experimental. For a relevant audit, we recommend an updated Chrome and a current version of Lighthouse.
Relying on PageSpeed Insights for a dev build
PageSpeed Insights can only audit publicly reachable URLs. To test a dev environment or logged-in pages, you need Chrome DevTools (local Lighthouse).
What the Agentic Browsing audit won’t show you
Lighthouse Agentic Browsing is a useful technical audit, but it isn’t a complete AI SEO audit. What it doesn’t measure:
- Content quality — whether your answers are useful, factual, citable.
- AI brand visibility — whether AI models know about you at all and cite you.
- Position in AI answers — you have to measure that analytically in ChatGPT, Perplexity, and Google AI Overviews.
- Deeper structured data optimization — Lighthouse isn’t a full structured data validator and doesn’t address the deeper design of your markup for search or AI systems.
Lighthouse Agentic Browsing is a gateway — if a site fails the basic checks here, that can point to weaker technical readiness for broader AI visibility and is worth digging into. The reverse holds too: a 100% share of checks passed in this category guarantees nothing about whether AI will cite you.
Key takeaways
- Lighthouse 13.3 in May 2026 added a new Agentic Browsing category that checks how ready a site is for AI agents.
- The audit tests four areas: llms.txt, WebMCP, the accessibility tree, and layout stability (CLS).
- The score isn’t weighted 0 to 100 — it’s the share of checks passed, with specific pass and fail signals.
- You can run it in Chrome DevTools (recommended), in PageSpeed Insights, or through the Lighthouse CLI.
- Fix order by impact: llms.txt → accessibility tree (names and labels) → CLS → WebMCP.
- The audit is a gateway, not complete AI SEO. Pair it with measurement in real AI tools and a content strategy.
Want to walk a Lighthouse Agentic Browsing audit step by step and fix the failed checks? An AI SEO audit from Sniper Design covers your key pages, works through the failed checks in priority order, and proposes specific fixes with an estimate of their impact.
For transparency: this article draws on publicly available Chrome for Developers documentation as of June 2026 and on tracking of Lighthouse 13.3+; individual checks and their weight may change in later versions. How the content on this site is produced is described on the author page.
