– AI agents primarily perceive websites through three methods: analyzing screenshots, parsing the DOM, or reading the accessibility tree, with the accessibility tree becoming the most reliable interface. – The accessibility tree, originally built for screen readers, is now the key interface for AI agents, meaning that improving web accessibility directly improves agent compatibility. – Using semantic HTML elements like `
The web is no longer a human-only space. For the first time, automated traffic now surpasses human visits, a trend confirmed by the 2025 Imperva Bad Bot Report. This shift means the most critical audience for your website may not be a person at all, but an AI agent performing tasks on a user’s behalf. To function effectively, these agents need to understand your site’s structure, and the key to that understanding is the same framework built for web accessibility. The accessibility tree, long championed for screen readers, is rapidly becoming the primary interface between AI and your content.
Major platforms employ different methods to interpret websites. Some, like Anthropic’s Computer Use and Google’s Project Mariner, rely on a vision-based approach, analyzing screenshots to identify elements. This method can be computationally heavy and fragile. Others, notably OpenAI’s ChatGPT Atlas and Microsoft’s Playwright MCP, prioritize the accessibility tree, querying structured data like ARIA roles and labels. The most capable systems, including Perplexity’s Comet, use a hybrid model that combines visual analysis with this structured data. The clear industry direction is toward reliability, and that means leading with accessibility information.
This makes your site’s accessibility tree its functional agent interface. It is a simplified version of the DOM that strips away stylistic clutter, exposing only interactive elements and their roles. For an AI operating within a limited context window, this efficient data structure is essential. Research underscores its importance. A rigorous UC Berkeley and University of Michigan study for CHI 2026 tested an AI agent on real-world web tasks. Under standard conditions, success was nearly 80%. When restricted to keyboard-only navigation, mimicking screen reader use, the success rate plummeted to 42%. The study identified perception gaps, cognitive gaps, and action gaps that hinder agents on poorly structured sites.
The foundation for a robust accessibility tree is semantic HTML. Using native elements like `
