The Convergence

Screen readers and AI agents arrived at the same dependency from opposite directions: both need the page to say what it means.

How a screen reader reads a page

A screen reader user moves by structure. Headings are a table of contents. Landmarks are a map. Form labels are contracts about what a field wants. The screen reader gets all of this from the accessibility tree, which the browser builds from semantic HTML and ARIA. Decades of WCAG work exist to make that tree truthful.

How an agent reads a page

Browser agents today consume some mix of the raw DOM, the accessibility tree, simplified text renderings, and screenshots. Across all of these inputs, the same properties that help a screen reader help the agent:

• Named, role-bearing elements. A button with an accessible name of “Submit order” is an unambiguous affordance. A div with a click handler and a CSS class of btn-7f3a is a guess.
• Structural honesty. Headings that nest correctly let both readers skim. Lists that are lists, tables that are tables.
• Labels as contracts. A programmatically associated label tells an agent what a field wants without inference.
• Sane reading order. Both readers consume the page roughly linearly. Meaning that only exists in visual layout is invisible to both.

And one property matters to agents alone: size. Every wrapper div, inline style, and tracking attribute is tokens. The cost of div soup is paid per page view, in money and latency, by every agent that visits. (The size claim is quantified in corpus experiment EXP-001.)

Where the convergence holds and where it breaks

The honest version of this chapter says: the overlap is large but not total, and the edges are exactly what the corpus exists to map.

The core overlap is no longer conjecture. A CHI 2026 study from UC Berkeley and the University of Michigan, “Is the Web Accessible for AI Agents?”, with its A11y-CUA dataset, tested agents across accessible and barrier-laden conditions and found task success near 78 percent on accessible pages, dropping by half or more when accessibility barriers were present. W3C WAI maintains active work on AI and accessibility, the Agent Experience (AX) movement is building the industry-side case, and benchmarks like WebArena already feed agents the accessibility tree as their primary structured input. a13i synthesizes, cites, and extends this body of work. It does not claim to have discovered it.

What the corpus tests, building on that literature:

• Semantic markup measurably improves agent task completion on equivalent content, replicated independently with published materials and pinned models (EXP-001).
• ARIA that fixes a screen reader experience usually helps or is neutral for agents, but redundant or incorrect ARIA can actively mislead both audiences.
• Alt text written for humans is close to optimal for agents, while keyword-stuffed alt text harms both.
• Some agent-only aids, like llms.txt or markdown mirrors, are neutral for human accessibility and therefore permissible under the manifesto’s people-first rule, but they never substitute for fixing the page itself.

The economic argument

Accessibility has fought an unfair budget battle for decades because its beneficiaries were treated as a small market. Agents change the arithmetic. The same remediation that opens a page to a blind customer now opens it to every AI assistant that shops, books, files, and reads on behalf of everyone. a13i’s bet is that when that double dividend is documented with numbers, accessible engineering stops being a compliance cost and becomes the obvious default. The people-first rule keeps the order of beneficiaries straight while the economics do the persuading.