Browser agent interaction semantics

Browser agents act on the rendered page, not just its text. Google's web.dev guide (last updated 1 April 2026) says agents use screenshots, the raw HTML/DOM and the accessibility tree. It recommends native button and a elements over styled div and span, role and tabindex where native semantics are missing, labels bound to inputs with for, interactive elements larger than 8 square pixels, a stable layout and no transparent overlays over controls. OpenAI said its ChatGPT Atlas browser (deprecated, with a shutdown date of 9 August 2026) used ARIA roles, labels and states to interpret interactive elements, and says ChatGPT agent works from screenshots of its virtual browser. Chrome's experimental Lighthouse Agentic Browsing category (Chrome 150+) audits a subset of accessibility checks (programmatic names, tree integrity) and layout shift, because agents read the accessibility tree to find interactive elements. This measures whether an agent can complete an action, which is separate from whether it can read the content.