Lighthouse 13.3 shipped a new report category in May: Agentic Browsing. It sits next to Performance, Accessibility, Best Practices and SEO, and it checks something most websites have never been tested for — whether a machine can read, navigate and act on your site.
The category is experimental and deliberately doesn’t produce a 0–100 score. It emits pass/fail signals across four areas. WebMCP: actions and forms your site exposes for agents to call, declared in markup or registered in JavaScript (still an origin trial). The accessibility tree: agents parse the same roles, names and labels screen readers depend on, which quietly turns accessibility debt into conversion debt. llms.txt: a machine-readable summary of your site at the domain root. And layout stability: agents screenshot the page, pick coordinates, then click — if your layout shifts in the gap, they press where the button used to be.
We ran gosmartr.co.uk through it: 2/2 on the category, alongside 100/100/100/100 on desktop. Then we did the honest thing and noted that 2/2 is the baseline — example.com passes 2/2. The checks that differentiate are the ones most sites haven’t shipped: a real llms.txt (ours is live at /llms.txt as of this week) and, later, WebMCP tools once the origin trial stabilises.
Our position is the same one we wrote in July about agents: they don’t look at your website, they read it. For twenty years web design optimised for human eyes — layout, colour, hierarchy. The sites that win the next few years will be legible twice: to the person deciding, and to the agent acting on their behalf. Semantic HTML, programmatic names on every interactive element, stable layouts, machine-readable summaries. None of it is exotic. All of it is now being measured.