Lighthouse now measures agentic browsing — what Google is checking, and why our AI-Ready stacks already score green without further implementation

TL;DR — the 90-second summary

What is new?

Lighthouse has gained an experimental category Agentic Browsing — four topic areas: WebMCP integration, discoverability, accessibility tree, layout stability.

How is it scored?

No 0-to-100 score — instead a pass ratio plus pass/fail status per audit.

Am I covered as a Moselwal customer?

Yes, for the most part. moselwal/webmcp delivers the WebMCP tools, the A11y discipline is part of our editor/template convention, CLS is controlled through the site-package templates, and llms.txt is active on every platform.

What Lighthouse now measures

Google has positioned Lighthouse as the de-facto quality gate for web delivery since Performance, Accessibility, Best Practices, SEO and the PWA era. With the new category Agentic Browsing — currently flagged experimental — another axis is added that targets a very specific question: How well can an AI agent understand this page and interact with it?

The category is not just documentary. Lighthouse typically runs as part of CI/CD pipelines, so Agentic Browsing will become a measurable requirement on front-end teams over the coming months — comparable to what Core Web Vitals have been since 2020. Whoever has structurally built for agents today sees a green score tomorrow without the build system going red.

The five audit areas in detail

1. Registered WebMCP tools

Lighthouse talks to the page over the Chrome DevTools Protocol (CDP) WebMCP domain and observes which tools are registered. Both declarative tools (marked up in the HTML via special tags) and imperative tools (registered in JavaScript via navigator.modelContext.registerTool()) are scored.

2. Forms without declarative WebMCP

Forms that are not declared through WebMCP are a black box for agents — the agent has to guess what the input fields mean and which validations apply. The audit finds forms without WebMCP annotation and flags them.

3. WebMCP schema validity

If a tool declares an input schema (a JSON-Schema-like structure with types and required fields), Lighthouse checks whether the schema is syntactically and semantically valid. Invalid schemas produce audit failures.

4. llms.txt as a discoverability signal

Lighthouse checks whether the domain root carries an llms.txt — a machine-readable summary of the site that LLMs can use as an entry document. Important note: Google’s own AI Optimization Guide from May 2026 states that llms.txt is not mandatory for SERP visibility — Lighthouse measures it anyway, because it remains useful for agent discovery.

5. Accessibility audits from an agent perspective

Agents use the accessibility tree as their primary data model. Lighthouse picks a subset that is particularly relevant for machines:

• Names and labels — every interactive element needs a programmatic name.
• Tree integrity — roles and parent/child relationships must be valid.
• Visibility — no content is hidden in the A11y tree while interactive.

Plus cumulative layout shift (CLS) as a stability metric.

Why there is no 0-to-100 score

Google says so explicitly in the documentation: because the standards for the agentic web are still emerging, the goal is to collect data and deliver pass/fail signals first, instead of cementing a ranking. That is methodologically sound — a 0-to-100 scale that gets refined later would make the score historically unstable.

In practice that means: you get a pass ratio in the category header (e.g. „6 out of 8 audits passed“) plus pass/fail/warning per audit. CI/CD integration is possible (pipeline break on critical failures), but the evaluation has to be a little richer than a simple score threshold.

What our AI-Ready stacks already cover — audit by audit

Here is the direct mapping per audit area:

Concretely: anyone running a TYPO3 site on our AI-Ready platform should see all mandatory audits green on the first Lighthouse run with the new category. The pass ratio should be 8/8, with occasional warnings for edge cases.

AI-Ready Commerce (Sylius) — same setup, one additional tool

For shop sites, the tool family for product search and product detail comes on top. It is part of our Sylius site-package extension and follows the same declarative pattern as the TYPO3 tools.

If you are not with us — minimum implementation per audit

The audits are not exclusive to our platform — any CMS or e-commerce system can satisfy them if the discipline is there. Per audit area, the minimum implementation:

WebMCP tools: The simplest path is declarative WebMCP markup in the HTML — no JavaScript complexity, no timing issues. Three tools are a good start: search, navigation, page-content. Shops add product-search.

Form annotation: Each form gets a name property on every input element plus a WebMCP tool annotation that declares the form as an invocable action.

Schema validation: Define a JSON Schema per tool, run it against a validator before deploy. A CI pipeline step.

llms.txt: A static Markdown document at the web root that summarises the site structure and the most important content. Generation can be manual or a build step from page metadata.

Accessibility: Semantic HTML instead of <div> soup. Every interactive element gets a programmatic name. ARIA only as supplement. CLS below 0.1.

If you have a TYPO3 or Sylius estate and want to set this up yourself: our open-source package moselwal/webmcp is built precisely for this and is MIT-licensed. The tool registry, the schema validation and the CDP hook are included out of the box.

Bottom line

Lighthouse Agentic Browsing is the next axis along which CMS and commerce platforms get measured. The category is experimental, but the five audit areas (WebMCP tools, form annotation, schema validity, llms.txt, A11y plus CLS) are structurally stable and won’t drift in other directions.

Anyone running on our AI-Ready CMS or AI-Ready Commerce platform sits largely green on the audits without further implementation — because the layers Lighthouse measures are the mandatory building blocks of our stack. Whoever wants to retrofit it themselves has the right toolbox in moselwal/webmcp (open source, MIT) plus semantic-HTML discipline. Both paths lead to the same outcome: a Lighthouse pass rate that doesn’t need to be patched together in a maintenance sprint, but is a by-product of a clean architecture.