How it works

An accessible PDF can be read by everyone, including people who use screen readers, magnifiers or other assistive technology. That needs real text rather than pictures of text, a structure that says what is a heading, a paragraph, a list or a table, a title and a language, and a description for every meaningful image. PDF Accessibility takes care of each of these.

What happens to your PDF

Text recognition. Every page is checked first. Pages that are scanned images, with no real text, get text recognition (OCR) so their words can be read aloud, searched and copied. Recognised text can contain errors.
Tagging. If the PDF has no structure, it is tagged automatically: headings, paragraphs, lists, tables and figures. Skipped heading levels are corrected and bookmarks are created from the headings. A PDF that is already tagged keeps its structure untouched.
Document settings. The title, language and viewer settings are set so the title is announced instead of the file name, and any document security is removed so assistive technology is not blocked.
Image descriptions. Each figure is cropped and sent to the AI service with its caption and the text around it, which drafts alternative text. Identical images, such as a logo on every page, are described once.
Review. Nothing is written to the PDF until you approve it. Every figure must be approved or marked decorative.
Building. Approved text is written to each figure, and decorative figures are removed from the reading order.
Checking. The result is checked against the original and you get an accessibility report.

Reviewing a PDF that is already tagged

Choose “Review alt text in a tagged PDF” when the document already has a good structure and you only want to check its image descriptions. Its tags are kept exactly as they are: no text recognition and no re-tagging. Every image is listed with the alternative text it already has, marked “Kept as is”. You can keep it, edit it, regenerate it with the AI, or mark the image decorative, and “Restore original” always brings back the text from your file. Images that have no alternative text get an AI draft to review.

Writing good alternative text

Say why the image is there. Describe what a reader needs from it in this document, not everything in it.
Keep it short. One or two sentences, around 125 characters, is usually right. Use the long description for charts, diagrams and tables.
Do not start with “image of” or “picture of”. Screen readers already announce a figure.
Do not repeat the caption. Add what the caption leaves out.
For charts, give the finding. Name the chart type, what is measured, and the trend or the notable values.
Mark purely ornamental images decorative. Borders, flourishes and repeated logos only slow a listener down.
Check names and facts. The AI does not identify people and can misread small text. You know the document; it does not.

Further reading: the W3C “alt decision tree” and the WebAIM article “Alternative Text”.

What the report can and cannot tell you

The report covers things a program can verify: tags, alternative text on every figure, heading levels, language, title, bookmarks and security. It cannot judge whether a description is accurate, whether the reading order suits a complex layout, or colour contrast. It is not a certificate of PDF/UA or WCAG conformance.

Your data

Text recognition and tagging run on this server. The AI service receives the cropped figures and up to a few hundred characters of nearby text, never the whole file. Guest uploads and results are deleted after 24 hours. If you sign in, your documents are kept in My documents until you delete them. Either way, “Delete now” removes a document immediately.

AI service: Anthropic (claude-haiku-5-5). PDF engine: PDFix 8.7.10 (unlicensed). Text recognition for scans: available.