SourceCited / Agent-ready
Guide · The agentic web
Browsers now act on pages, not just read them. Agent-readiness is mostly disciplined fundamentals — plus Chrome's Lighthouse audit and the emerging WebMCP protocol.
Agentic browsers (ChatGPT Atlas, Perplexity Comet, Copilot Mode in Edge) parse your accessibility tree and act through your forms — so agent-readiness means semantic HTML, labeled inputs, stable layout, and facts as text. Score yourself with Chrome Lighthouse's Agentic Browsing audit (default since May 2026), and pilot WebMCP rather than betting on it.
Agentic browsers do not just render your pages — they act on them. ChatGPT Atlas (launched Oct 21, 2025, macOS-first, now being folded into a unified OpenAI desktop app), Perplexity Comet (July 2025, now free on every platform), Copilot Mode in Edge, and Gemini in Chrome all browse, click, fill forms, and increasingly buy on the user's behalf. That makes agents a second audience, with different needs from both human visitors and classic crawlers.
Agents read your page largely through the accessibility tree — the same structure screen readers use — and act through your interactive elements. What breaks a screen reader breaks an agent: unlabeled inputs, divs styled as buttons, critical facts trapped in images, layouts that shift mid-click.
| Agents need | Why | Do this |
|---|---|---|
| Semantic HTML & real headings | Agents navigate structure, not pixels | Proper landmarks, one h1, ordered h2/h3 |
| Labeled, accessible forms | Agents fill what they can identify | A label for every input; native controls over div-buttons |
| Layout stability (CLS) | An agent clicking a moving target misses | Reserve space for ads/embeds; watch CLS in Lighthouse |
| Facts as text, not images | Prices or policies inside images are invisible | Visible price, availability, and returns in HTML text |
| Short, low-friction flows | Every step is a failure point | Fewer fields, clear states, no surprise modals |
Accessibility now pays double: compliance for humans, parseability for agents. The same fix serves both.
Act as a browsing agent trying to complete this task on the page I'm pasting: [task, e.g. "find the price and start checkout"]. Navigate by the accessibility tree, not the visuals. List every point where you would fail or guess — unlabeled inputs, non-semantic buttons, facts only in images, ambiguous links, layout-shift risks — with the exact HTML fix for each. End with a pass/fail verdict for the task.
Chrome's Lighthouse added an Agentic Browsing audit category — default since Lighthouse 13.3 (May 2026). It checks llms.txt presence, WebMCP wiring, accessibility-tree integrity, and layout stability: a free, concrete score for how usable your site is to agents. Run it from Chrome DevTools, and add it to CI so regressions get caught before agents hit them.
WebMCP is a W3C draft (co-authored by Google and Microsoft engineers, announced Feb 10, 2026, in a Chrome origin trial) that lets a page expose callable tools to in-browser agents via navigator.modelContext — the agent calls your search or add-to-cart function instead of guessing at your DOM. The declarative path (attributes on existing forms) is the low-effort pilot. MCP is the server-side sibling — created by Anthropic and donated to the Linux Foundation's Agentic AI Foundation in December 2025 — worth an endpoint if you run APIs or data products. Status for both: emerging. Pilot them; do not rebuild around them yet.
Agents can be manipulated by page content (prompt injection). Never publish text you would not want an agent to execute as an instruction, and treat agent-triggered actions on your site as untrusted input.
A browser whose built-in AI agent can act — click, fill forms, complete tasks — on the user's behalf rather than just displaying pages. ChatGPT Atlas, Perplexity Comet, and Copilot Mode in Edge lead; they read pages through the accessibility tree and act through your interactive elements.
Not yet. It is a W3C draft in a Chrome origin trial, not a shipped standard. The declarative form-attribute path is cheap enough to pilot — and Lighthouse already checks for it — but semantic HTML, labeled forms, and stable layout deliver value today regardless.
For agents, directly: they parse the accessibility tree, so an unlabeled control or an image-only fact is invisible to them. It does not change classic rankings by itself, but it decides whether an agent can complete tasks on your site.