Skip to main content

How we audit — and what we will not claim

WCAG 2.1 AA has 50 success criteria. Our automated checks can prove failures on 22 of them — only those can count in the score. Another 21 are examined by advisory checks or AI review and reported without scoring, and 7 are left to a guided manual checklist. Every report states exactly which criteria were scored on that run.

Three evidence classes, never mixed

Every finding in a report carries one of three labels, and only the first one affects the score:

  • Machine-verified

    A deterministic check reproduced the failure in a real browser, and its evidence meets the check's registered proof: a positive observation, or exhaustive absence over a closed set. Every machine-verified check is also tested against real-browser pass and fail cases. Only these findings reduce the headline score.

  • AI-assessed

    A model examined screenshots and page structure and judged something likely wrong — alt-text quality, link purpose, visual structure. Useful signals, clearly labeled, never counted as proof and never scored.

  • Needs verification (candidate)

    An observation whose evidence cannot prove the failure on its own — a missing caption track on a video whose audio may contain no speech, for example. Reported with the exact verification step a human should take, excluded from the score.

Scoring: no credit for what we did not test

The score counts only criteria where a machine-verifiable check actually ran on the audited pages — at most 22 of the 50. A criterion examined only by AI review, or with nothing on the page for its checks to examine, is reported but never scored, and a tool that fails removes its criteria from the count. Within the scored set, a criterion without confirmed findings scores 100%; one with confirmed findings keeps 75% (minor), 50% (moderate), 25% (serious) or 0% (critical) of its credit, by its worst finding. The four WCAG principles are weighted 30% perceivable, 30% operable, 25% understandable and 15% robust, redistributed across the principles that have scored criteria.

Each report also shows a sensitivity scenario: what the score would be if every unconfirmed risk signal turned out to be a real failure. You see the optimistic and pessimistic bounds, not just one flattering number.

Sampling is disclosed, never hidden

Some checks examine a sample — focus visibility is judged by comparing screenshots with and without keyboard focus on up to 40 focused elements per page, and up to 5 linked PDF documents are checked per page. When a criterion's evidence comes from a capped sample, the report marks it as sampled with the exact numbers, and a clean result means no issues in the sample — not proof for every element.

Real browser interaction, not just DOM scanning

Beyond axe-core scanning, the audit drives a real browser: it tabs through your interface and verifies suspected keyboard traps by actually trying to escape them, submits forms empty to observe error handling, tests reflow at 320 pixels and text resize, and records how every page was actually acquired — final URL, HTTP status and load outcome.

Known limitations

We would rather tell you what an automated audit cannot do than let a number imply it did:

  • Human-judgment aspects — whether alt text is accurate, whether an error message helps, whether the focus order is meaningful — are left to AI review or the manual checklist and never counted as proven.
  • Pages behind login and flows requiring credentials are recorded as unreachable, not silently skipped.
  • Multi-step journeys (checkout, registration) are tested in their initial states; full journey completion needs a human tester and the report includes step-by-step instructions for it.
  • Closed shadow DOM, canvas/WebGL applications and cross-origin embeds are only partially inspectable — the report discloses these boundaries where they occur.

The reports practice what they audit

Every generated PDF report — including the public samples — is automatically validated against PDF/UA-1, the accessibility standard for PDF documents. Reports carry their scoring-model version and generation date, and the narrative is cryptographically bound to the finalized evidence it describes.

Scores from different scoring-model versions are never comparable; every report states its version. Methodology last updated September 2026.