Release notes

last updated 2026-06-18

2026-05-16 — First case study at `/research/`: an increased a14y score halved an agent’s token use

New /research/ hub on a14y.dev publishes empirical case studies on what a14y actually buys an AI agent. First case study: same Astro docs, same eval task, same agent. Raising the site’s a14y score from 37 to 89 cut Claude’s token use by 49%, tool calls by 52%, and wall-clock by 30%. Answer quality was statistically indistinguishable (judge 84 ± 3 vs 83 ± 5; both pass 3/3). The first measured answer to “does shipping for agents actually save them work?“

Every per-site scorecard under /leaderboard/[site]/ ships its own Open Graph card, so links shared to social or chat preview the score, the site, and its category at a glance.

Rebuilt landing page at a14y.dev: clearer “what it is, who it’s for, what running it looks like.” The old /research/ leaderboard moved to /leaderboard/.
New press kit at /press/: one-pager, logo lockups, and ready-to-paste product screenshots.
Three new draft scorecard checks under markdown.* shipped spec-first, then implementation: a complete end-to-end example for community contributors to model from.
Site crawler memory bounded so large-catalog audits no longer OOM.

Landing page relaunched around the embeddable readability badge: show your score, link to your scorecard, share it anywhere.
Draft scorecards now render an inline diff against the latest published version, so reviewers see exactly what would change before it lands.

One shareable score summary across the CLI, the Chrome extension, and the agent skill. Every surface emits the same paste-ready output and badge URL.
Embeddable a14y badge: any site can drop in a live score badge sourced from a14y.dev.
Per-site scorecard pages: each leaderboard entry now has its own deep-link URL with the full check breakdown.
Draft scorecard workflow opens contributions: anyone can propose a check change with spec, implementation, and a reviewable diff against the published scorecard.

The a14y Chrome extension is live on the Chrome Web Store. Install once, audit any page from the toolbar.
New /research/ page: an agent-readability leaderboard ranking 234 sites, with score histograms, category breakdowns, a hardest-checks ranking, and a deep-link scorecard for every entry. The first public answer to “how readable is the web for agents?“

Pinned the agent skill at metadata.version 0.2.0 so installs are versionable and indexable.
Itemized-checklist plan: every fix proposed by the skill is its own row the user can keep or skip — no monolithic “approve all”.
Design-matching guardrails: the skill must read the project’s existing layout and styles before adding any user-facing content.

Installable into any spec-compliant agent — Claude Code, Codex, Cursor, OpenCode — with npx skills add timothyjordan/a14y.
Published the a14y agent skill under /.well-known/agent-skills/ per the Cloudflare Agent Skills Discovery RFC.
Every page on a14y.dev advertises the skill via <link rel="agent-skills"> so agents can pick it up automatically.

Crawler stops chasing phantom URLs harvested from .md mirrors, so site-mode audits no longer surface non-routes as failing pages.

Anonymous GA4 events from the CLI, the Chrome extension, and the docs site so we can see which checks people actually rely on.
Opt-out is one click on the docs at /privacy/, one toggle in the extension options, and --no-telemetry on the CLI.

Project name and npm package are now a14y — the homophone of “a11y” makes the agent-readability framing legible at a glance. The legacy package names remain installed as aliases.
a14y check <url> is the default CLI invocation; the bare a14y <url> form also works for quick audits.
Manual light/dark toggle on a14y.dev with the user’s choice winning over OS preference.

--output agent-prompt produces a markdown plan an LLM can execute end-to-end: ranked failing checks, affected URLs, and re-run instructions in one document.
Now the audit isn’t just a score — it’s a hand-off to the coding agent that’s going to fix the site.

First public scorecard pinning 38 checks across the three layers of agent readability — discoverability, parsing, and comprehension.
CLI, Chrome MV3 extension, and Astro docs site shipped together so the same engine evaluates a site whether you run it from the terminal, the toolbar, or your CI.