How accessible are small-government PDFs? Monthly tracker
Wave of 2026-10-16 · next update around 2026-11-16 · measured by Hieu Tran with AI agents, from public websites only
Every month we check the same 300 randomly chosen US city, county and special-district websites and the PDFs they link from their home page and main sections. The checks are the first ones any accessibility review runs: is the file tagged, is it a scan with no text, does it have a title and a language. Under the ADA Title II rule, entities of 50,000 people or more must meet WCAG 2.1 AA by 26 April 2027, smaller ones and special districts by 26 April 2028.
83%of PDFs fail at least one basic machine check (534 of 647)
52%have no tags, so a screen reader gets no structure (334 of 647)
41%of PDFs created since the rule change on 20 Apr 2026 are still untagged (75 of 185)
83%of sites have at least one untagged PDF (87 of 105)
Of 647 PDFs from 105 websites in 46 states, 534 of 647 (83%); 95% interval 77–88% fail at least one check. They hold 7,353 of the 8,207 pages we checked. On 24 of 105 (23%) of the sites, every PDF we checked was untagged.
Newer files are better, but 41% are still untagged
The rule's exception for older documents covers files posted before the compliance date, unless people still need them to apply for or take part in a service; agendas, packets, budgets and forms posted from then on are covered. So the files that matter most are the new ones. Of the PDFs created since the Justice Department moved the dates (20 April 2026), 75 of 185 (41%); 95% interval 31–51% are untagged, against 236 of 434 (54%) of older files.
Table view
Year created
PDFs
Untagged
Scanned, no text
2021 or earlier
117
86 (74%)
47 (40%)
2022
34
22 (65%)
8 (24%)
2023
34
13 (38%)
5 (15%)
2024
59
20 (34%)
6 (10%)
2025
105
56 (53%)
31 (30%)
2026
270
114 (42%)
49 (18%)
28 PDFs record no creation date and are left out of this chart.
How the file was made decides most of it
The producer field inside each PDF says which program wrote it (grouped here; the raw field is left out of the data because some name the office). That is the cheapest thing an office can change. Files from Word, Save as PDF were untagged 5% of the time (11 of 212); files from Microsoft Print to PDF 100% (35 of 35).
Made with
PDFs
Untagged
Scanned, no text
Word, Save as PDF
212
11 (5%)
0 (0%)
Adobe (Acrobat, InDesign)
134
61 (46%)
3 (2%)
Not recorded
104
99 (95%)
67 (64%)
Scanner or copier
97
96 (99%)
86 (89%)
Other software
41
22 (54%)
1 (2%)
Microsoft Print to PDF
35
35 (100%)
3 (9%)
Browser or Google Docs print
15
4 (27%)
1 (7%)
Mac (Quartz)
9
6 (67%)
1 (11%)
Word's own Save as PDF keeps tags when "Document structure tags for accessibility" is on (the default); Microsoft Print to PDF never adds them, and copier scans almost never do. Groups with fewer than 3 PDFs are left out.
By website platform and type of government
Website platform
PDFs
Untagged
Scanned, no text
WordPress
247
128 (52%)
69 (28%)
CivicPlus
213
123 (58%)
51 (24%)
Not identified
99
38 (38%)
20 (20%)
Wix
40
21 (52%)
15 (38%)
Drupal
31
17 (55%)
6 (19%)
Revize
9
4 (44%)
0 (0%)
Granicus (govAccess)
6
1 (17%)
0 (0%)
Platform is detected from the home page; groups with fewer than 5 PDFs are left out. A platform's own document tools only help if the office uses them.
Type
PDFs
Untagged
Scanned, no text
City
476
254 (53%)
131 (28%)
County
85
32 (38%)
9 (11%)
Special district
86
48 (56%)
22 (26%)
Check your own site. The free Title II PDF check runs the same crawl and the same checks on any website and lists each file, with a page count and a cost estimate. Nothing to sign up for.
Method
Panel. 300 domains drawn at random (fixed seed) from the 12,991 city, county and special-district domains in CISA's public .gov list, October 2026. The same panel is re-measured every month, so changes are changes on the same sites. Entities that use .org, .com or .us addresses aren't in the list, so they aren't in the panel.
Reach. 238 domains answered with a website; 9 of those ask bots not to read them (robots.txt) and are left out; 46 refused automated reads (they may differ from the rest, for example by sitting behind a commercial firewall); 183 could be read, and 110 of those link PDFs from the home page or up to 6 sections one click away (agendas and minutes first). 18 post agendas on an outside platform (Granicus, CivicClerk, iQM2 and others), whose files we don't check.
Files. Up to 8 on-site document links per website, picked at random; 54 turned out not to be PDFs (Word files, web pages) and 20 timed out or were over 15 MB. One request at a time per site, identified as TitleIIPDFCheck with a link back to this tool, and only pages and files the site's robots.txt allows.
Checks. Tagged (a structure tree and the Marked flag), scanned (images and no fonts), document title (info dictionary or XMP), language, page count, producer and creation date. Our parser agreed with pypdf on tags, title and language on all 79 of 80 randomly drawn files from this wave that pypdf could open (the 80th is AES-encrypted), and on page count for 78 of them; the 77 files of the pilot sample (12 October 2026) agreed on all four. Where our parser can't read a page count (3 files) the file counts as 0 pages, so page totals are, if anything, low.
How it's made. AI agents built and ran the crawl; every number is computed by a script from the CSVs below, and the parser was cross-checked against pypdf.
What it doesn't say. Passing these checks doesn't make a file accessible: reading order, alt text, tables and contrast need a person. Failing them almost always means the file doesn't conform. A creation date can be missing or reset by an editing program. Eight files is a sample of each site, not its whole library. Intervals on PDF shares resample whole sites, because files from one office tend to be made the same way.
Data
CC BY 4.0. One row per website and one row per PDF, with type of government, state, platform and the check results. Names and URLs are left out on purpose: this is about how small governments publish, not about any one town.