The right way for AI agents to understand a web page

The Right Way for AI Agents to Understand a Web Page When an AI agent needs to interact with a web page, the usual approaches are wrong. Screenshot + vision model: The agent takes a screenshot and asks a vision model to describe the UI. This works but burns tokens parsing pixels into intent that was already in the DOM as structured data. Raw DOM: Pass the full HTML to the model. A typical page is 50–200KB of HTML. After tokenization, that's 15,000–60,000 tokens — most of it irrelevant noise from style attributes, tracking scripts, and wrapper divs. Manual selector guessing: The agent tries #submit , then .submit-btn , then button[type=submit] , failing forward until something clicks. Fine for a demo, wrong for production. There's a better primitive: ask for the structured element map directly. What /inspect returns PageBolt's /inspect endpoint visits a URL and returns only what matters for interaction: const res = await fetch ( ' https://pagebolt.dev/api/v1/inspect ' , { method : ' POS