Accessibility issues repeat.

What the study found

Automated accessibility testing normally produces a queue: one row per failing element, worked through in order. That framing treats every row as an independent defect. This dataset suggests many of them are not.

• 60.3% of confirmed findings (69,224) belonged to a high-confidence recurring cross-page pattern.
• 90.9% of eligible sites (80 of 88) had at least one recurring pattern.
• 27.7% of recurring patterns — those spanning 20 or more pages — accounted for 82.8% of all recurring findings.
• The median affected site had 47.6% of its recurring findings concentrated in just three patterns.
• 72.2% of eligible serious and critical findings were recurring, against 36.2% of moderate ones.

Patterns clustered where you would expect: navigation, headers, footers, modals, shared forms and repeated cards. Navigation patterns alone reached 5,676 pages. This is consistent with defects living in shared components rather than being made page by page — though it identifies a probable shared implementation worth investigating, not a proven cause.

Why the method is conservative

A finding counted as recurring only where the same rule failed against the same distinctive component fingerprint on at least two different pages of the same site. Generic markup — a bare button, a framework utility class — was never accepted as evidence of a shared component.

Where the evidence was too weak to judge, findings were markedunclassified rather than assumed isolated. That share is 29.0% of all findings, and it is preserved in the published figures. It means 60.3% is a floor, not an estimate.

The report also declines to rank issue types by raw instance volume, because different rules cap how many elements they record per page at different thresholds. Ranking censored counts against each other would produce a confident and wrong league table.

The counter-example

A classifier that labelled everything a component problem would be telling you nothing. Links that aren't visually distinguishable within a paragraph (link-in-text-block) appeared on 35.9% of sites but only 9.0% of pages, and just 23.3% of eligible findings were recurring. It behaves like the editorial problem it is, and needs content standards rather than a template change.

Broken skip-link targets are the clean contrast: 96.7% of eligible findings recurring, 100% median site spread. One skip link, defined once in a shared layout.

What it does not establish

The 92 websites were a purposive sample from one Australian business network directory, anonymised, and are not representative of Australian businesses or of the web. Every finding is automatically detectable, which is a real but partial slice of WCAG.

The study does not establish complete keyboard journeys, logical focus order, keyboard traps, screen-reader usability, whether alternative text is useful rather than merely present, whether error messages aid recovery, cognitive usability, or WCAG conformance for any site. It covers public pages only — no authenticated areas, no downloadable documents. WCAG 2.5.3 Label in Name was not evaluated, and the report says so plainly.

Method at a glance

Scan date

8 August 2026

Sample

92 anonymised business websites, purposive not random, 0 excluded

Pages

4,135 recorded, capped at 100 per site; 18 sites reached the cap

Engine

axe-core 4.11.2 plus RedFlag-native automated checks

Rule scope

WCAG 2.0 A, 2.0 AA, 2.1 AA and 2.2 AA additions, plus six explicitly enabled axe rules. Level AAA not systematically evaluated

Population

114,843 confirmed violations. 27,011 advisory findings excluded

Verification

Main findings independently re-derived with a second implementation; pattern grouping differences below 0.6 percentage points

Every issue in the report maps to a documented rule in theRedFlag rules library.