ADA Title II · What a checker can prove
Can an automated checker prove a PDF is accessible?
No. A checker decides part of the standard, never all of it. On the one real document library we have run end to end — 2,691 documents — zero passed automatically.
What "passed" actually means, and what it does not
Recordstake tests PDFs with veraPDF against the machine-checkable subset of PDF/UA (ISO 14289-1). The honest reading of a pass is "no machine-checkable rule failed" — never that the document is accessible. WCAG 2.1 AA includes requirements no validator can decide: whether alt text actually describes the image it sits beside, whether a table's headers mean what the layout implies, whether the reading order makes sense read aloud rather than merely present.
A document can clear every machine-checkable rule and still fail a person's read of it. That is not a gap in the tooling. It is the boundary of what software can decide, stated as what it is rather than papered over.
Roughly a third of the standard cannot be decided by any machine
The Matterhorn Protocol is the PDF Association's own conformance model for PDF/UA — the body that wrote the standard, not a vendor's summary of it. Extracted directly from the protocol document itself: of its failure conditions, roughly a third require human judgement and cannot be resolved by any checker, however well built.
A rule can pass while the real problem is still there, and this is a named case in the standard itself, not a hypothetical: marking real content as a decorative artifact removes it from the accessibility tree entirely. The automated check finds nothing wrong — there is nothing left in scope to check — while a screen reader user loses the content outright. A validator reports a clean file. A person using the file has nothing read to them at all.
Two of the standard's most common real-world defects — a heading that was never tagged as a heading, and a list that was never tagged as a list — are on the same list, for the same reason: a machine can confirm a tag is present or absent, not whether the document was structured the way a person reading it would expect.
Why zero out of 2,691 is not an unusual result
On the one real municipal document library this project has run end to end — a full account is here — not one of 2,691 documents passed an automated PDF/UA check outright. That is not a verdict on that city in particular. Published estimates put automated tooling's coverage at roughly 30–40% of accessibility problems industry-wide, and this corpus sits inside that range: real defects exist that a checker can find, and a real share of the standard sits entirely outside what any checker can decide.
Any vendor claiming a fully automated PDF/UA check is contradicting the standard body's own conformance model. That is not our reading of the situation; it is what the Matterhorn Protocol itself says about which of its own conditions software can resolve.
We are not attorneys and this is not legal advice. We produce evidence and a recommendation you can hand to your counsel, who decides.
See where your own documents land
The free scan reports what an automated check actually found, in the same three buckets as everything else we produce: what was measured, what a person judged, and what nobody assessed — never blurred into a single pass or fail.
Get the free scan
Three fields — your agency, its website, and where to send the result. Nothing to
install, nothing to sign, and no meeting.
Related questions
- When is the ADA Title II deadline for your agency? — both dates, and why a lot of published advice still has the old ones.
- Does the rule cover your old PDFs? — the two exceptions, and why age currently discriminates nothing.
- What does PDF accessibility remediation cost? — $2,000 for the Inventory, published because almost nobody else in this field publishes a price.
- How do I get a free ADA accessibility scan for my agency? — three fields, a confirmation from your own inbox, no login required.
- What one city's document library actually looks like — every document and every sheet, tested, with the parts nobody assessed marked as such.