-
Notifications
You must be signed in to change notification settings - Fork 12
Evidence and Research
fecarrico edited this page Jul 25, 2026
·
2 revisions
Why this page exists: A11Y.md doesn't start from zero. Every hard rule in the standard responds to something measured in the field. This page collects that evidence — what each source is, where to read it, what it found, and what this project does about it. It grows as new data appears.
- What it is: The largest recurring accessibility survey on the web — an automated evaluation of the top 1,000,000 home pages, run annually by WebAIM since 2019.
- Link: webaim.org/projects/million
-
Main findings (2026):
- 95.9% of home pages have detectable WCAG failures — the first regression in six years.
- 56.1 errors per page on average (+10.1% over 2025); one error every 26 page elements.
- 133+ ARIA attributes per page on average — more than 6× the 2019 number — and pages with more ARIA correlate with more errors, not fewer ("ARIA soup").
- The same six error categories (low contrast, missing alt text, unlabeled inputs, empty links…) have accounted for 96% of all errors for seven straight years.
- What A11Y.md does about it: The core file's anti-patterns target exactly those six recurring categories — including a named ARIA Soup anti-pattern (no ARIA is better than bad ARIA: no redundant roles, no labels duplicating visible text, no static never-updated states). The AI Behavior Contract prefers native HTML over ARIA: the agent reaches for semantics first, and ARIA only where the platform has no native answer. More generated markup should not mean more broken markup.
- What it is: Atlassian's engineering team adopted DESIGN.md (portable markdown design context for coding agents) and published what happened when they measured it against their MCP server in real production work — the closest thing the field has to a benchmark of markdown-as-context itself.
- Link: atlassian.com/blog/how-we-build — testing portable design context in practice
-
Main findings:
- +92% token consumption (7.21M vs 3.75M) and slower runs (6m46s vs 5m01s).
- Context coverage of ~30% vs ~80% via MCP, with 2.7× higher token variance.
- The critical failure mode: the markdown context induced agents to recreate components instead of reusing the existing library.
-
What A11Y.md does about it: This is the failure mode the standard's Reinventing the Complex Wheel anti-pattern always pointed at — and since v1.1.0, two contract rules close the gap directly: Component Reuse (the AI must extend what the project already has) and Decision Memory (
A11Y-DECISIONS.mdrecords choices between equally conformant alternatives, so twenty modals don't ship with twenty different focus patterns). Lazy Context Loading keeps the persistent core small precisely because token cost is real.
- What it is: A Search Engine Journal analysis (2026) connecting two curves that crossed: web accessibility regressing for the first time in years, exactly as AI agents became the majority of HTML traffic — and agents read pages through the accessibility tree, the same interface screen readers use.
- Link: searchenginejournal.com — the accessibility tree is how AI agents read your site
-
Main findings:
- Semantic accessibility is no longer only a compliance topic: it is now a prerequisite for a site being readable by machines at all.
- The US DOJ acknowledged having overestimated overlays and AI remediation — automated detection catches only ~30% of real barriers.
- What A11Y.md does about it: Remediation-after-the-fact is exactly the model the project rejects. Principle Zero moves accessibility to generation time: if the agent writes semantic, operable markup from the first prompt, the accessibility tree stays healthy for humans and machines.
- Anthropic brings persistent memory to Claude Managed Agents — mounted context files became a first-class industry paradigm; A11Y.md is that paradigm applied to accessibility.
- AI-generated code is inaccessible by default — independent confirmation of the statistical bias this project exists to counter.
- Early signs of EAA enforcement across Europe (Deque) — the regulatory pressure the Certifiable pillar and the EN 301 549 mapping prepare adopters for.
Third-party voices talking about A11Y.md — kept here both as a thank-you and as a map of how the ecosystem reads the project. Quotes are verbatim from each source.
- Claude for Open Source (Anthropic) — in July 2026, A11Y.md was selected for Anthropic's program supporting open-source maintainers worldwide.
- Smashing Magazine — Smashing Newsletter #564 (June 2026) — dedicated an item, "Accessibility Context System", to the project: "Felipe A. Carriço created A11Y.md, a context system for building accessible software by default."
- Jury Vetrov's Product Design Digest — April 2026 (in Russian) — listed under Standards and best practices as a context instruction file you load into an AI session so generated code stays accessible.
- Design Fragments #131 (in Chinese) — recommended as a tool for following WCAG 2.2 AA and ADA "from the first line of code".
- en<able> — design & accessibility channel (in Russian, via Desigram) — "a ready-made context you load into Claude (or any other model) and it starts checking your layouts against WCAG 2.2."
- Roger Wong — "A11Y.md: accessibility for AI agents" (June 2026) — dedicated post describing the project as a "persistent context architecture" that makes any agent (Claude, Cursor, Copilot) follow WCAG 2.2 AA from the first line of UI code, and framing it as a validation protocol rather than a passive guideline.
- Vitor David — "IA não conserta sistemas inacessíveis" (May 2026, in Portuguese) — senior product designer and member of ABNT's CB-040 accessibility committee (the body behind Brazil's NBR 17225 web accessibility norm), arguing that AI scales existing patterns — good or bad — and recommending A11Y.md as a way to "teach, document and build repertoire" for the AI.
- Frontend Dogma (Jens Oliver Meiert) — listed in the accessibility archive on April 26, 2026 — the day the repository went public — and later shared on Bluesky.
- Radar — Diyor Khakimov — catalogued under Accessibility × AI as a "DESIGN.md-style context file that installs WCAG rules into AI coding agents so accessible HTML is the default output."
- ACCESSIBILITY.md — ecosystem comparison (Mike Gifford / CivicActions) — lists A11Y.md in its multi-project matrix as a "persistent context protocol for AI-assisted accessibility", crediting its "strong protocol framing and workflow/checklist orientation". The comparison frames the ecosystem's projects as complementary layers — a reading we share.
- INCLUSION.md (Branon Eusebio) (May 2026) — proposes a companion file for contextual and representational inclusion, citing A11Y.md as an established repository convention "alongside CONTRIBUTING.md and AGENTS.md" for shaping implementation correctness and agent behavior.
📬 Know a study we should list, a mention we missed — or a number we got wrong? Open an issue or PR. Entries need a public source with a link; that's the whole point of this page.
Project A11Y.md | The Persistent Context System for Accessibility Created by Felipe A. Carriço | MIT License
♿ Accessibility is not a feature; it's a precondition for use.