Agent Audit

Agents — and assistive tech — understand <button> and <a> natively. A <div onclick="..."> is invisible to the accessibility tree unless you add role and tabindex, and even then it's a worse signal than the real element. This check flags non-semantic elements that look interactive (have onclick, role="button", etc.).

Every form control should have an associated <label> — either via for="input-id", by wrapping the input, or via aria-label/aria-labelledby. Without this link, agents and screen readers don't know what a field is for.

If you must use a <div> or <span> as a button, it needs role="button" (or appropriate role) and tabindex="0" so it lands in the accessibility tree and keyboard focus order.

A non-default cursor is a strong signal that an element is interactive. Vision-based agents pick up on it. We pass any element that uses a deliberate cursor — pointer, grab, move, zoom-in, etc. — and only flag elements still using the inherited auto/default. Known exceptions we silently skip: disabled buttons (which correctly use not-allowed), native text inputs (cursor: text is right for them), and <details> containers (the click target is the <summary> inside, which is checked separately). Requires computed styles, so bookmarklet only.

Tiny or zero-size interactive elements are a hint that something is broken or hidden. The web.dev guidance: keep visible area larger than 8 px². Requires layout measurements, so bookmarklet only.

Transparent overlays or opacity: 0 covers that still capture clicks confuse vision-based agents — they see one thing and click another. We flag elements with computed opacity: 0 or visibility: hidden that still take up layout space. Known exception we silently skip: the styled-input pattern, where a native control (e.g. <input type="color">, file picker, checkbox) is hidden with opacity: 0 and overlaid by a styled element. We detect this by the presence of an aria-label or aria-labelledby — that's the signal the developer made the input invisible on purpose, with accessibility intact.

Lighthouse's agentic audit flags Cumulative Layout Shift because content that jumps around as it loads moves targets out from under an agent mid-task. The biggest culprit is media without reserved space, so we flag <img>, <iframe>, <video>, and embeds that have no width/height attribute and no CSS aspect-ratio. In URL mode we can only see attributes and inline styles, so external CSS may still size them — it's a warning, not a failure. The bookmarklet also reads computed aspect-ratio.

An /llms.txt at your site root is an emerging convention for handing language models a machine-readable, markdown summary of your content and structure. It's optional — not required by Google Search, and its benefit is still unproven — so a missing file is reported as a skip, not a failure. URL mode checks it server-side (the browser can't read it cross-origin); the bookmarklet reads it directly since it's same-origin.

WebMCP is a proposed protocol for declaring, machine-readably, what the interactive elements on your page actually do — so an agent doesn't have to infer it. The spec is still in active development, so this check is purely informational: we sniff the signals proposed so far (a typed <script> block, <link rel="mcp">, a <meta> tag, and in the bookmarklet the runtime navigator.modelContext) and report whether any are present. Absence is a skip, never a failure.