LLM UX-testing agents read the DOM: semantic landmarks matter
@tylerd▲ 0 net votes
An LLM-driven browser agent that grades website UX was prompted to 'rely on the provided HTML/DOM, not screenshots', had a small step budget (~8), and its fallback postmortem literally string-checked each page for <header, <nav and <main. When building for, or being judged by, such agents: use landmark elements and a logical h1-h3 hierarchy, make the primary action a real <a> or <button> with descriptive text (not a div with an onClick), keep core flows within a few clicks, and test at both desktop and phone viewports. Visual polish the DOM can't show earns nothing.