Methodology

Every criterion deserves a verdict — or a protocol, not a blindly checked box

Our method starts from a simple observation: most audit tools stop where automation stops. Grown out of an R&D lab focused on automated compliance detection, we built ours to go further, while working hard not to mistake a tooling artifact for a real defect.

WCAG and RGAA, two independent frameworks

WCAG 2.1/2.2 levels A and AA (56 criteria) and RGAA 4.1.2 (106 criteria) are tested each on its own terms, not one derived from the other. A simple RGAA↔WCAG mapping table produces both false positives and false negatives: a single WCAG criterion can point to about twenty thematically unrelated RGAA criteria.

For every criterion covered by our approach, we produce a verdict — automated with no human check, or quickly confirmed by hand on an analysis already done (never a re-audit). Criteria that require direct human judgment, with no automated method at all, aren't integrated into our approach: as a research lab, we aim to push the limits of automation as far as possible — combining traditional methods with LLM judgment for human-style analysis — rather than falling back on a classic manual audit.

Regulatory scope covered: EAA (European Directive 2019/882), WCAG 2.1/2.2, RGAA 4.1.2. See also where RGAA version 5 stands, currently being drafted.

Our approach, step by step

Scoping

Site mapping, grouping by page template (not raw URL count), explicit scope presentation for a conscious trade-off, documented exclusions in the final report rather than silently dropped.
Large-scale automated detection

axe-core 4.9.1 scanning in a real browser, rich DOM inspection (accessible names, heading structure, ARIA roles), automated keyboard navigation, contrast, zoom and reflow, text spacing, pointer-gesture cancellation, CSS/JS-disabled comparison, office documents. An advanced battery rounds this out: interface-component contrast (luminance computed on the element, its containers, and its children), CSS color declarations with fallbacks, information carried by pseudo-elements, hidden content, hover-revealed content with no keyboard equivalent, complex gestures caught by instrumentation set up before the page loads (non-applicability proven when no gesture exists, with a systematic search for a button-based alternative), context changes on input, status messages, the geometric position of form labels, engaging forms, flash animations, and focus return when modal dialogs close (real open/close sequences simulated, not assumed).
Real evidence instead of assumption

A bridge to a real screen reader (NVDA), a real session-timeout test (actual wait, no simulation), triage assisted by a multimodal vision/language model via API for unlabeled icons, touch-gesture and screen-rotation simulation, and a step-by-step audit of a full transactional flow (results, pricing, form) that never goes through to a real payment.
Scaling to the whole site

Parallel automated crawl to audit hundreds of pages in cascade, aggregating defect patterns to spot shared components rather than isolated page-by-page symptoms.
Cross-checking

Every potential defect is checked against one question: is this a real site defect, or a tooling artifact? See our methodological rigor principle below.
Deliverables and prioritization

CSV, Excel workbook, Word report, and interactive chat — in French and English — with severity-based prioritization and systemic defects highlighted for maximum leverage.

RGAA aggregation at the unit-test level

Our methodology relies on the 258 official RGAA 4.1.2 unit tests rather than a single overall evaluation per criterion. A criterion is only declared compliant if all its applicable tests pass across every page in the audited sample. The verdict for each criterion is calculated automatically from the test results, not entered by hand; the official compliance rate excludes criteria that don't apply to the audited site. This is what makes our central promise possible: every criterion deserves a verdict — or a protocol, not a blindly checked box.

Local W3C validation

The rendered source code of every page — the one assistive technologies actually consume — is validated by a local instance of the W3C validator, with no page-size limit. Messages are filtered to the scope of the official test (tags, nesting, duplicate attributes) to avoid off-topic noise.

Multimodal-model judgment, kept in check

A multimodal vision/language model, called via API, judges images declared decorative, text-as-image, information carried by color alone and by shape/position alone — on the full page capture — as well as the relevance of labels and the site map. For transparency, we display a dual rate: the official rate, which includes the model's judgments, and the mechanically proven rate without them. An exhaustive appendix lists every piece of text judged; that material is kept for sampling-based review.

Consistency across screens

Each page's navigation signature (menu order, placement of the site map and search) is compared against the others, along with the consistency of field labels serving the same function across different screens — a frequent defect that page-by-page auditing never surfaces.

Real evidence: at the core of our difference

Instead of leaving a criterion "not tested" for lack of automation, we built protocols to get real evidence at a reasonable implementation cost.

Real screen-reader bridge (NVDA ↔ Playwright)

We capture the speech NVDA actually says — not an approximation of what it should say. This confirms defects with exact quotes of what a blind user hears (for example, an English date pronounced in French for lack of a lang attribute).

Real session-timeout test

An actual 30-minute wait, with spaced checks, to verify there's no silent warning or logout on a real transactional flow — not an accelerated simulation.

Vision-model-assisted triage

For every icon button flagged by the DOM as lacking an accessible name, a vision model proposes a relevant label and judges icon/function correspondence. Framed as triage, not a final verdict — a dedicated column flags cases needing human review.

Touch gestures and screen rotation

Simulated via Playwright emulation, without dedicated physical hardware — useful for verifying pointer-gesture cancellation (WCAG 2.5.2) and landscape-orientation behavior.

Methodological rigor: avoid mistaking a tooling artifact for a real defect

A principle applied strictly throughout our audits: avoid documenting as a real defect a behavior that is actually a malfunction of the audit tooling itself.

A "15-30s rendering freeze" initially logged as a site defect turned out to be a screenshot-tool timeout — removed after investigation.
Two accidentally duplicated defects from one audit pass were caught and merged during a regression run.
The vision-triage tool was itself audited: an early version produced false positives on buttons whose accessible name actually came from an associated <label> — fixed before going into production, revalidated at zero false positives.

What remains fundamentally manual

Some of the 162 criteria (19% RGAA, 29% WCAG) require direct human judgment that no automated method can cover yet — screen-reader parity, the payment step, alternate-document fidelity, authentication flows. We don't integrate these criteria into our approach: as a research lab, we aim to push the limits of automation as far as possible — combining traditional methods with LLM judgment for human-style analysis — rather than falling back on a classic manual audit. We state this limit openly rather than hiding it — that's a trust argument, not a weakness to conceal.

Screen-reader parity

What NVDA announces can differ from JAWS or VoiceOver on some complex components.

The payment step

The transactional flow is audited step by step (results, pricing, form), but the payment step itself is never gone through, by principle.

Sampling-based review of the model's judgments

The multimodal model's judgments (alt text, site map, labels) are checked by human sampling against the exhaustive appendix kept for every audit.

Alternate-version fidelity

Judging whether an alternate office document truly contains the same information as its web counterpart isn't automatable, even though its detection and technical tagging are.

A transparent method, verifiable results

Apply for a free audit