ADA Title II · April 2027
Two annual reports arrive with 17,557 and 25,099 accessibility failures. They leave with fourteen.
Foliowright re-tags an existing PDF in place — it never rebuilds the page, so the document still looks exactly like the document. What it cannot determine, it refuses to guess and hands you instead, with the page and the image attached.
Measured on the six hardest documents we could find
Real PDFs published by six US universities — four annual financial reports, a 155-page accessibility guide, a financial-aid form. No files we wrote ourselves. Checked in PAC 2024, the tool accessibility offices actually open, and in veraPDF, the ISO reference validator.
- Before
- 46,699
- After
- 624
- Structural defects
- 0
machine failures across 419 pages, as published
331 images awaiting a human sentence, 293 source fonts
tag tree, reading order, tables, lists, language, metadata — and zero warnings
| Document | Pages | Before | After | PAC warnings |
|---|---|---|---|---|
| Stanford financial-aid form | 2 | 64 | 4 | 0 |
| Alabama annual report FY24 | 78 | 17,557 | 14 | 0 |
| Rutgers annual report | 86 | 25,099 | 14 | 0 |
| Michigan annual report | 53 | 1,229 | 35 | 0 |
| California accessibility guide | 155 | 1,428 | 240 | 0 |
| Pennsylvania annual report | 45 | 1,322 | 317 | 0 |
PAC still shows a red banner. That is the honest answer.
Every one of those output files still says that, and will keep saying it until a person writes the alt text. PAC's verdict is binary — any unmet checkpoint turns the banner red — and an image with no description is an unmet checkpoint no matter how good the tag tree underneath it is.
Any vendor showing you a green banner on a chart-heavy annual report either wrote the alt text for you, or guessed. We do neither, on purpose. What we can tell you is exactly how many sentences are left and which pages they are on — for a 78-page annual report, nine.
What it refuses to do
The failure mode in this category is not a tool that misses things. It is a tool that produces confident, plausible, wrong output that still validates — so nobody catches it, least of all the person relying on the description.
It will not describe a chart
Asked to caption a bar chart of WCAG conformance, a general caption model told us about “the percentage of people diagnosed with cancer… 71% experienced an increase.” Fluent, specific, entirely invented, and a screen-reader user has no way to know. Data graphics go to a person.
It will not re-embed your fonts
A font the source never embedded cannot be fixed by re-tagging, whatever a product page claims. We report it as the source-document problem it is, and tell you which font on which page.
It will not send your documents anywhere
The default mode is fully local — no upload, no API call, nothing leaves the building. A cloud model is available per figure, chosen by a person, and never automatically.
We are looking for one accessibility office
Not a customer. A design partner. Everything above was measured against public documents we chose ourselves, which is the weakest kind of evidence there is — what it has never faced is a real backlog and a real coordinator's judgement.
20–50 PDFs from your backlog
Syllabi, policies, board minutes, forms, financial tables, slide exports. De-identified is fine.
Whatever reports you already have on them
PAC, Acrobat, CommonLook — so we are measured against your benchmark, not ours.
One hour, showing us how you remediate a document today
Timed. The number that matters to us is minutes per document, and we do not have yours.
Thirty minutes, later, reacting to our output on your files
Is it correct? Is the review faster than what you do now? Where is it wrong?
In return: early access, your priorities shape what gets built, and a tool designed around your workflow rather than a generic one. Nothing to sign, nothing to buy, no tool to switch.
Start a conversation