AccessLumens Research · 2 September 2026
The State of US County PDF Accessibility 2026
58.1% of the county PDFs we scanned have no tag tree at all.
A PDF without a tag tree is not a document to a screen reader. It has no headings to navigate by, no reading order, no table structure, and no way to tell a caption from a paragraph. Under the Department of Justice’s ADA Title II rule, PDFs are explicitly “conventional electronic documents” and therefore regulated web content. We drew a random sample of 100 US counties, harvested up to 20 documents from each one’s public portal, and validated 941 of them.
58.1%
Have no tag tree
547 of 941 documents
54/100
Counties we could measure at all
46 yielded nothing
33.4%
Are pictures of text
need OCR, not tagging
0
Median score
out of 100
The deadline has not passed, and that is the point. On 20 April 2026 the DOJ extended ADA Title II compliance to 26 April 2027 for entities serving 50,000 or more people, and 26 April 2028 for smaller ones. There is still time. But the rule’s exception for documents posted before the compliance date does not cover documents people currently use to apply for, access, or take part in a service, and that describes most of what a county publishes: permit applications, tax forms, benefit paperwork, meeting agendas.
Retagging a back catalogue takes months, not weeks. A county that starts the week before the deadline has already missed it.
“An untagged permit form is not a hard-to-read document. It is a blank page with a mouse-only signature box.”
Key findings
1. Untagged is the norm, not the exception.
58.1% of the 941 documents we could assess have no document structure tree. Weighted by county rather than by document, the average county’s untagged rate is 57%. Both numbers are reported because they answer different questions: the first is what a resident encounters, the second is what a typical county publishes.
2. Size does not predict this.
Counties in the smallest population fifth had an untagged rate of 59.3%, against 49.7% in the largest fifth. That looks like a gradient until you read the middle: the second fifth is the worst in the study at 72.1%, above both ends, on 7 counties. With so few counties behind each band there is no reliable trend here, and we are not going to claim one. What the bands do show is that no size of county is in good shape, including the largest, which face the earlier 26 April 2027 deadline.
3. The failures cluster on structure, not on decoration.
WCAG 1.3.1 failed in 98.9% of documents. These are not cosmetic complaints: they are the criteria that decide whether assistive technology can determine what a piece of content is. A document can look immaculate on screen and still fail every one of them.
4. Some documents cannot be fixed by tagging at all.
33.4% of the documents are pictures of text: scanned paper with no characters in the file. Tagging cannot help them, because there is nothing to tag. They need OCR first, which is a different budget line and a different vendor conversation, and we report them separately for that reason rather than folding them into the untagged number.
How counties are spread
Share of each county's sampled documents that have no tag tree, across the 54 counties that yielded documents.
By population band
Counties sorted by population and split into five equal-count fifths. Scores exclude colour contrast, so they are not comparable with a customer report.
| Population band | Counties | Documents | Untagged | Mean score |
|---|---|---|---|---|
| Smallest fifth | 13 | 214 | 59.3% | 23.1 |
| Second fifth | 7 | 136 | 72.1% | 13.9 |
| Middle fifth | 11 | 205 | 58.5% | 21.6 |
| Fourth fifth | 11 | 201 | 54.7% | 18.8 |
| Largest fifth | 12 | 185 | 49.7% | 23.7 |
Most-failed WCAG criteria
Counted once per document, not per occurrence, so a single enormous document cannot dominate.
What we could not measure
Reachability is a finding of this study, not a footnote. Every county we could not measure is listed here with the reason.
| Reason | Counties |
|---|---|
| Blocked by bot protection we did not bypass | 19 |
| Portal did not respond | 9 |
| No public county portal exists | 7 |
| No PDFs found within the crawl's limits | 4 |
| Documents found, but none would download | 2 |
| Rate-limited too heavily to measure | 2 |
| robots.txt refused us | 2 |
| Portal returned an error | 1 |
21 counties sat behind bot protection or rate limiting we chose not to defeat. Blocking is not spread evenly across the sample: 0 of the 20 smallest counties were blocked, against 21 of the other 80. We report that split without drawing a conclusion from it. It is tempting to argue that bot protection costs money and therefore hides better-resourced counties, which would mean these figures flatter the sector, but the pattern across the bands is not orderly enough to support the claim in either direction, and a guess about the counties we could not see is exactly the kind of thing this study should not be making.
Check one of your own PDFs
Run the same PDF/UA check we used for this study, free, no signup. You will see whether the document has a tag tree, which WCAG criteria it fails, and what a repair would actually change.
Check a PDF freeMethodology
Sampling
This is a probability sample, not a curated list. We sorted all 3,144 US county and county-equivalent governments by Census population estimate, split them into five equal-count bands, and drew 20 counties per band with a fixed seed, giving 100 counties. Equal counts per band, not equal populations, because the question is whether county size predicts anything, which needs comparable sample sizes rather than comparable populations.
Establishing each county’s portal
One domain per county, and it must be the county’s primary public portal: a sheriff’s office, circuit clerk or chamber of commerce is a county publisher but is not that portal, and mixing them in would make any size comparison partly an artifact of which body happens to hold the county’s domain. Six such wrong-publisher matches were caught and corrected. 7 counties were found to have no public portal at all, recorded as a finding rather than a dropped row. Domains were established by reading the live site and recording the evidence, and the provenance is published per county, distinguishing rows a human confirmed from rows an automated check confirmed. Those are not the same standard of evidence: six rows rested on the weakest evidence in the file, and when they were re-examined, three proved wrong. All six were then re-established, three by a person in a browser and three from evidence rather than a page, because those hosts refuse automated fetches and a real browser alike.
Harvest
From each portal we crawled up to 150 pages at one request per second, obeying robots.txt under our own token, and drew up to 20 of the PDFs found, at random with the same seed, so that a county publishing 4,000 documents does not outvote one publishing 40. A random draw rather than the newest or the first the crawler met, both of which are orderings the county’s CMS chooses.
Validation and scoring
Documents were validated with veraPDF against PDF/UA-1 and mapped to WCAG success criteria, the same engine and scoring the AccessLumens product uses, so study and product numbers cannot drift apart. Scores here exclude colour contrast, so they are not comparable with the score AccessLumens reports for a customer’s document; the untagged rate is unaffected.
What “untagged” means, and what it does not
“Untagged” means the file has no document structure tree (no StructTreeRoot), read directly from the PDF rather than inferred from a conformance verdict. That is the most forgiving of the signals available, and deliberately so: a further 344 documents do have a tag tree that veraPDF nonetheless fails on structure, and counting those as untagged as well would put the headline at 94.5% rather than 58.1%. We report the lower number because “has no tag tree” is a fact about the file that needs no interpretation, while “its tags are inadequate” is a judgement. Read the gap as the study understating the problem rather than overstating it.
Reconciliation, naming, and scope
Every attempted county and document ends in either a result or a logged, typed failure, and the two reconcile against the attempted count; the pipeline refuses to publish a dataset that does not add up. Individual counties are not named: the purpose is to describe the sector, not to single out governments that are mostly running on one IT person. Compliance dates are those set by the DOJ’s interim final rule of 20 April 2026 (26 April 2027 for entities serving 50,000 or more, 26 April 2028 for smaller entities and special districts), which extended the original 24 April 2026 date. This is a diagnostic snapshot, not a legal conformance determination (see our disclaimer).
What we know we missed
Three limits are worth stating plainly, because each of them makes a county look worse than it is. The federal .gov registry contains dead entries. Two counties in this sample were recorded against registry domains that do not work: one resolves to an address its own CDN refuses, the other is a parked advertising page. Both counties were initially counted as publishing nothing. One of them, once pointed at its real website, turned out to publish plenty. If the government’s own index of government domains is partly stale, any study built on it will understate the sector unless it checks every domain by hand. Some county sites cannot be read without running their JavaScript. A county whose portal is a client-rendered app serves an empty page to any crawler that does not execute scripts, ours included. We report those as unreachable rather than as empty, but as small counties move onto app-style vendor platforms this becomes a growing blind spot for every crawl of this kind, not just ours. Nearly half the sample could not be measured at all. That is reported above with a reason for every county rather than quietly dropped, because a sector where the documents cannot be reached from outside is itself the finding.