The State of PDF Accessibility
We validated 84 real documents from government, higher education, healthcare and procurement against the PDF/UA standard. Four passed.
of the documents we tested failed PDF/UA before any remediation.
77 of 81 documents · median 98 distinct failures each
A PDF that looks perfectly normal on screen is often unreadable to someone using a screen reader. Underneath the visual layer, a tagged PDF carries a structure that tells assistive technology what is a heading, what is a table, what order to read the page in, and what an image shows. Most PDFs carry none of it.
That claim gets repeated often in accessibility circles, usually without numbers behind it. We had the numbers sitting in our own validation logs, so we cleaned them up and published them.
This note reports what we found when 84 documents that real organizations actually publish, including municipal notices, construction bid packages, course syllabi, clinical handouts and annual reports, were run through an independent PDF/UA validator exactly as received.
Eighty-four documents, no synthetic samples
Every document in this corpus was a real file uploaded to AccessiblePDF by someone who wanted it remediated. We removed everything that was not a genuine third-party document: our own marketing PDFs, engineering test builds, pipeline output that had been fed back in, editor iterations, and synthetic fixtures. Where the same document was uploaded more than once, we kept a single copy.
Corpus composition by sector
| State and local government | 22 | |
|---|---|---|
| Construction and procurement | 19 | |
| Commercial | 19 | |
| Higher education | 11 | |
| Healthcare and nonprofit | 8 | |
| Nonprofit and association | 5 |
Failure is close to universal
Of the 81 documents that produced a baseline validation report, 77 failed PDF/UA. The four that passed were documents whose publishers had clearly already done accessibility work.
Failure was not marginal. The median document carried 98 distinct violations, the average was 108, and the worst single document produced 530. Across the corpus the validator recorded 8,731 individual failures.
The failures are basic, not exotic
The most common problems are not subtle edge cases in the specification. They are the foundations. Two thirds of documents contain page content that is neither tagged as real content nor marked as decorative, which means a screen reader simply does not know what it is looking at. More than half never declare a document title for the viewer to announce. A quarter contain images with no alternative text at all.
Most common PDF/UA failures, by share of documents affected
| Untagged page content | 67.9% | |
|---|---|---|
| No PDF/UA identifier | 64.2% | |
| Title not shown in viewer | 53.1% | |
| Not marked as tagged at all | 33.3% | |
| Link has no description | 28.4% | |
| Link missing alt text | 28.4% | |
| Font subset incomplete | 27.2% | |
| Figure missing alt text | 25.9% | |
| No structure tree | 25.9% | |
| Metadata missing title | 24.7% | |
| Text language undefined | 22.2% | |
| Metadata stream invalid | 22.2% |
Untagged content and a missing structure tree are not cosmetic defects. They are the difference between a document a blind student can navigate by heading and a document that reads as one undifferentiated wall of text, or as nothing at all. A missing document title means the viewer announces a filename instead of a name.
Long-form documents are the worst offenders
The burden is not spread evenly. Course materials, clinical handouts and marketing collateral averaged well over a hundred violations each. Standardized construction contracts were the cleanest category by a wide margin, averaging 17, most likely because they are produced from controlled templates by publishers who generate them at volume.
The pattern that emerges is unsurprising in hindsight: documents assembled by hand, with mixed images and multi-column layout, fail hardest. Documents generated from a system fail least.
Average violations per document, by document type
| Course material | 160.7 | n=7 | |
|---|---|---|---|
| Clinical or patient material | 151.4 | n=7 | |
| Marketing collateral | 140.8 | n=5 | |
| Public information guide | 139.0 | n=6 | |
| Technical reference | 131.3 | n=9 | |
| Government form | 67.7 | n=9 | |
| Bid or RFP package | 64.9 | n=11 | |
| Government notice or plan | 64.0 | n=7 | |
| Annual report | 40.5 | n=4 | |
| Construction contract | 16.8 | n=5 |
Most of it is machine-fixable
47 of these documents were subsequently processed through automated remediation and re-validated, one automated pass per document with no human edits. Total violations fell from 5,879 to 202, a reduction of 96.6 percent, and the median document finished at zero. 38 of the 47 passed PDF/UA outright with no human intervention.
The nine documents veraPDF still flags deserve a closer look, because the validator is not the last word. Two of them pass PAC, the other major PDF/UA checker, with no failures at all, including a 33-page hand-assembled exhibit package that is the most structurally complex document in the corpus. One is flagged for a single finding that targets a blank region of the page, content a human reader and a screen reader would both correctly ignore. The genuinely resistant cases share a profile: complex tables whose rows disagree about how many columns they have, and scanned pages whose text layer cannot be reliably recovered. Those still need a person.
We report this because it bears on policy, not only on products. If the majority of the backlog can be cleared mechanically, then the compliance problem facing schools and agencies is smaller and cheaper than the raw violation counts suggest. It is the residue that needs expertise, and the residue is a minority.
What this corpus cannot tell you
These figures describe a convenience sample, and the sampling bias runs in an obvious direction.
- Self-selection. Every document here was uploaded to a remediation service by someone who already suspected it had a problem. Documents that are already accessible are underrepresented, because nobody sends those to be fixed. The true population failure rate is almost certainly lower than 95 percent.
- Size. Eighty-four documents is enough to describe which failure types dominate. It is not enough to publish precise rates by sector, which is why the document-type chart omits categories with fewer than four documents.
- Validators are the floor, not the verdict. Headline results come from veraPDF against the PDF/UA-1 profile, and this corpus shows the limits of any single machine check in both directions. Among the nine documents veraPDF still flags, two pass PAC, the other major PDF/UA validator, with zero failures, and one is flagged only on a blank region of the page. Machine validation catches structural violations; it cannot judge whether alternative text is useful, whether reading order makes sense to a human listener, or whether a flagged region even contains content. The best verification of accessibility remains a cross-check with PAC and, above all, listening to the document through a screen reader. A document can pass every automated check and still be a poor experience, and it can fail a check and read perfectly.
- We are not disinterested. Palanisi Labs sells PDF remediation. That is a reason to read the fourth finding critically, and a reason we have published the method, the exclusions and the limitations in full rather than only the headline.
How this was produced
Corpus construction
All jobs processed by AccessiblePDF between 27 April and 19 August 2026 were extracted, 562 in total. Files were excluded when they were pipeline output re-uploaded for testing, bracketed engineering builds, truncated slices of a document also present in full, editor iteration chains, vendor sample fixtures, unnamed scratch uploads, coursework, or documents authored by Palanisi Labs.
Remaining files were normalized to collapse re-upload variants, then deduplicated so each real document appears once. That left 84 distinct documents, of which 81 produced a baseline validation report.
Classification and anonymization
Each document was assigned a type and sector from its filename, then the filename was discarded. No filename, publisher, organization, uploader, or email address appears in this report or in any dataset derived from it. Documents are identified only by a sequential number assigned at extraction.
Validation
Every document was validated with veraPDF against the PDF/UA-1 profile (ISO 14289-1:2014), in the state it was received. Baseline figures are from that first validation. A violation is one entry in the validator’s issue list; a document failing the same rule in twelve places counts twelve violations but one affected document. The two are reported separately throughout.
Remediation outcomes and the engine-version correction
The remediation pipeline changed materially across the collection window, and an earlier internal draft of this note scored each document by whatever engine version happened to run it last, which both undercounted the current system and double-counted a handful of documents uploaded under multiple filenames. For the figures published here, re-upload variants were collapsed to one document each (47 unique remediated documents). Where a document has a result on the engine version currently in production, including the verification runs that gated engine releases, that result is used: 12 documents were scored this way, one automated pass, as received, no human edits. All other documents keep their most recent recorded outcome. Finding four reports these corrected figures.
Outputs that veraPDF still flagged were additionally cross-checked with PAC and by human inspection of the flagged regions; those cross-checks are reported in finding four and in the limitations, but the headline pass count remains veraPDF-strict — a document only counts as passing when veraPDF reports zero failures.
Reuse and citation
These figures may be quoted freely with attribution. Suggested form: Gibson, T. (2026). The State of PDF Accessibility. Palanisi Labs.
Journalists and researchers who want the underlying anonymized dataset, the full 44-rule failure table, or the exclusion log can request them; we will provide them.
Why the deadline moved
On 20 April 2026 the Department of Justice issued an interim final rule postponing the ADA Title II web accessibility compliance dates by roughly a year. Entities serving populations of 50,000 or more now have until 26 April 2027. Smaller entities have until 26 April 2028.
What did not change is the standard. WCAG 2.1 Level AA remains the requirement, the rule still covers documents rather than web pages alone, existing archives are still in scope, and the underlying Section 504 and Section 508 obligations were never paused at all. The Department stated it fully anticipates implementing the regulation at the new dates.
The reasons DOJ gave for the delay are worth reading closely, because they are not the reasons most people assume.
- Automation was not good enough on its own. The Department pointed to the limits of the tooling, noting that automated remediation cannot reliably handle complex content, and educational and STEM material in particular, without human review.
- Institutions do not have the people. Public entities were found to lack the staffing, specialized expertise and budget the original timeline assumed.
- Litigation exposure. Entities that could not meet the date faced private enforcement risk regardless of good-faith effort.
- The standard itself was ambiguously cited. The 2024 rule linked to a WCAG page carrying several versions rather than fixing one.
The first rationale is an empirical claim about technology, and this corpus is a test of it. The result is that DOJ is right, and also incomplete.
Right, because automation did not finish the job. 9 of 47 documents still failed after an automated pass, and the ones that resisted were exactly the category the Department named: complex tables whose rows disagree on column count, and scanned pages whose text layer cannot be recovered. Those need a person.
Incomplete, because the same run cleared 96.6 percent of all violations and took 38 of 47 documents to a clean pass with no human touch at all. Automation is not a substitute for expertise. It is the thing that decides whether the scarce expert spends the afternoon on ten documents or on one.
Read together with the rest of this note, the picture is not that the deadline was unrealistic. It is that the arithmetic was. Institutions holding thousands of documents, averaging a hundred violations each, were asked to clear them with staff they do not employ. The extension bought a year. It did not change the arithmetic, and a year of manual remediation does not close a backlog of that size.
What we built, and what it does not do
This note was produced by Palanisi Labs, which operates AccessiblePDF, a commercial PDF remediation service. The corpus above is our own validation log. We have described the method, the exclusions and the sampling bias in full so that the findings can be judged independently of the product. This closing section is the product part, kept separate on purpose.
Where the automation does best
The measured strengths, from the data above rather than from marketing copy: documents whose structure is recoverable clear reliably, with 81 percent passing PDF/UA with no human edit and a median of zero remaining violations. The failures that dominate the corpus, meaning untagged content, missing document titles, undeclared language, absent PDF/UA identifiers and missing image descriptions, are precisely the categories automation handles well. Together those account for the majority of everything we found.
Automated remediation runs at $0.50 per page with no minimums and returns an independent veraPDF report with every file.
Where a person is still required
Complex tables and unrecoverable scans. That is the honest boundary, and it is why the second half of the product exists rather than a claim that the first half is sufficient.
The Tag Editor
The editor is the answer to the expertise gap DOJ identified. It is a browser-based visual editor that works against the live page, so remediation stops being a specialist skill performed in a desktop tool and becomes something a subject matter expert can do on their own documents. The hardest documents in this corpus, the 33-page exhibit package among them, were finished in it.
- A visual tag tree beside the live page. Click content to find its tag, click a tag to see its content. The full PDF/UA vocabulary: headings, paragraphs, lists, tables, figures, links, captions, artifacts, with reading order reordered by keyboard.
- Smart splits for real-world typography. Split by line, at a line, into words, or one click turns a bold lead-in into a heading by splitting exactly where the font changes.
- Region tools. Drag to tag a figure or mark decoration as artifact instantly. Untagged documents open with draft tags generated from the text layer, so nobody starts from zero.
- Forms and tables handled properly. Create seven field types from a tag’s region, link existing fields, write tooltips; set header scope, column and row spans; alt text, actual text and language per element.
- The validator in the room, in plain English. Real veraPDF results translated out of clause numbers, checks that run live as you work, and a searchable rules guide, so the issue count falls while you edit instead of after an export-validate-repeat loop.
- A screen-reader preview, because that is the real test. An NVDA-style simulation reads the document the way assistive technology announces it, navigable by heading, list, table and link. The question it answers is whether the document sounds right, not only whether it passes.
- Built for hard documents. OCR for scans on the way in, CAD and diagram handling, and Apply writes the edits into the real PDF with bookmarks and the PDF/UA identifier stamped, with version history, an exportable change log, and cloud-saved sessions that resume where you left off.
- Accessible itself. Keyboard-first navigation throughout, on the principle that an accessibility tool that fails its own standard has no business selling one.
The editor is $50 per month for a single user or $225 per month for five seats with a shared workspace, with no per-page charges on editor runs, and either plan cancels at any time.
Any PDF can be run through a free compliance check that returns the complete list of PDF/UA failures at no charge, with no contract and no sales call. That report is useful whether or not you ever remediate the file with us, and it is the fastest way to find out where your own backlog actually stands.