Lighthouse's Agentic Browsing score is mostly the accessibility work you already skipped

Two of the four areas Lighthouse looks at were already on the backlog. Accessible names, real `<button>` elements and labelled inputs are what screen readers have needed for years ^[6], and CLS is a Core Web Vital that the dev.to walkthrough concedes already affects mobile usability ^[10]. Those two carry the high priority ratings; llms.txt and WebMCP are marked low ^[12][13]. Half the list is a rescoring of hygiene you already owed, which is why the category is easier to fund than it first looks ^[2].

The scoring rule is where the number goes soft. Points come only from applicable checks ^[5], and a missing llms.txt can show up as Not applicable rather than as a failure ^[11]. A site that has published neither llms.txt nor a WebMCP tool is therefore graded on semantics and layout stability alone ^[1], and can post a clean sheet without doing a single agent-specific thing. The sample report in the article reads three of three checks passed ^[14]. A three-item denominator will not survive comparison between two pages, and nobody should quote it as evidence that an agent can drive a checkout.

The framing worth keeping is the negative one. dev.to is explicit that this is a readiness signal rather than a ranking system, and that a failed check is not an SEO penalty ^[3]; Google has separately said llms.txt neither improves nor reduces rankings ^[11]. That removes the usual manoeuvre in which accessibility work has to be smuggled through as an SEO line item. What the checks reward is what screen readers reward: a control that states what it is, on a page that does not move under the pointer after load ^[8]. The article's console snippet for listing controls with no accessible name is the cheapest way to see how far off you are ^[7].

WebMCP is the part that changes the shape of a front end rather than its markup. The example registers a tool on `document.modelContext` with a name, a description, an input schema, an async execute function and an abort signal, behind a `'modelContext' in document` feature check ^[15][16]. The candidate workflows listed are product search, booking, quote generation, support tickets and account operations ^[17]. An input schema types the arguments; it does not establish that the caller is entitled to make the call. Pointing registerTool at an existing read-only search is a small job. Pointing it at booking publishes a write path whose only described guard is that schema, and the source offers a priority rating, not a security review ^[13].

One caveat over all of it: this is a single walkthrough on dev.to, last updated 21 August 2026 ^[19], describing a category that is still experimental ^[1]. The score is a regression test on markup, not a certification.