Full-page photos, scanned pages, no structure: why the first automatic run stopped at 84 points, which new rule solved the rest – and what a person should look at despite the passed check.
Anonymised case study from a real processing run; the name of the organisation and the document title have been changed.
An urban development plan with 344 pages could be brought from 0 to 100 points fully automatically and now passes the PDF/UA-1 check – but only at the second attempt. The first run stopped at 84 points, and not because of tables or headings, but because of full-page photos and the jumble of letters that text recognition had read from these photos. This case study shows what caused it, which rule solved it and what remains for a person to do despite the passed check.
Starting point
An integrated urban development plan (in Germany: integriertes Stadtentwicklungskonzept) describes where a town or city wants to go in the coming years: an analysis of the current situation, a guiding vision, fields of action, measures, plus specialist contributions. Such plans are adopted politically, discussed in public participation procedures and consulted again and again over the years. They are among the documents that many people actually want to read.
The plan of a large city discussed here came with the typical difficulties:
- 344 pages, divided into chapters and specialist contributions.
- Many full-page photos, usually with a short caption.
- Some scanned pages that had already been through text recognition (OCR).
- No usable tag structure – the check resulted in 0 out of 100 points.
What the check found
0 points does not mean that the document was badly made; visually, it was a carefully designed plan. It means that assistive technologies cannot do anything with it. Without tags, there are no headings for jumping from one field of action to the next, no figures with alt text and no separation between content and page decoration. At best, a screen reader reads the text out in the order in which it is stored in the file – practically unusable for 344 pages. On the scanned pages, there was the added problem that nobody had checked the existing text layer, including everything that text recognition had read from images.
What DokAudit did automatically
First run. The fully automatic mode built the structure: headings, paragraphs, lists and tables, running headers and footers as artefacts, plus title, language and bookmarks. DokAudit then checked again with veraPDF and its own checks of the logical structure. Result: 84 out of 100 points. The PDF/UA identifier could not be set yet – DokAudit only sets it if the check is actually passed. Two things stood in the way:
- Photos as decoration: the full-page photos were marked as artefacts, i.e. as meaningless decoration. For screen reader users they did not exist, even though they are part of the content.
- OCR letter jumble as text: on the photo pages, text recognition had found supposed letters in image structures. These strings of characters without real words were tagged as paragraphs and would have been read aloud.
New ‘photo pages’ rule. A rule of its own came out of this case. It detects pages that consist almost entirely of an image with little text and treats them the way a person would:
- The photo becomes a figure (Figure). The page’s caption serves as alt text, for example ‘Photo: neighbourhood festival in the pedestrian zone’. If there is no caption, the alt text is left open for an AI suggestion – provided AI is permitted for the organisation.
- Text on the page that hardly consists of real words is recognised as OCR noise and becomes an artefact.
In this document, 18 full-page photos thus became figures with alt text, and 53 elements with OCR noise disappeared from the reading order as artefacts. The rule has been part of the fully automatic mode ever since.
Result
Key figures of the urban development plan case study| Feature | Original file | First run | With the ‘photo pages’ rule |
|---|
| Score | 0 out of 100 | 84 out of 100 | 100 out of 100 |
|---|
| PDF/UA-1 | failed | not yet passed, no identifier | passed, identifier set |
|---|
| Full-page photos | no structure | as decoration (artefact) | 18 figures with the caption as alt text |
|---|
| OCR noise | no structure | tagged as paragraph | 53 elements as artefacts |
|---|
100 out of 100 points and a passed standards check mean that all machine-checkable requirements are met. It does not mean that every question of content has been answered.
What a person should still do
- Two chapter divider pages: here, the chapter title was adopted as alt text, following the pattern ‘Photo: SPECIALIST CONTRIBUTION …’. Formally this is fine, but thin in substance. Better is a sentence that says what the photo shows – or, if it only creates atmosphere, the decision to treat it as decoration and mark up the chapter title as a heading.
- Four pages without searchable text: someone should look at whether they contain content that is missing as text – such as a map with a legend – or whether they are pure image pages for which alt text is sufficient.
- Colour contrast: the contrast measurement reported a warning for several thousand characters. Contrast is not a PDF/UA rule, but it is a WCAG requirement (1.4.3). The ‘Improve colour contrast’ switch would automatically darken text that is too light; because this changes the appearance, it is optional. Alternatively, the design can be adjusted in the next edition (Colour contrast in PDFs).
- Proofread the captions: they come from the document and are therefore correct in substance. The editorial team should check on a few examples whether they are sufficient for people who cannot see the photo.
- Listening test: listen to a few chapters with a screen reader, especially the formerly scanned pages (Test it yourself).
Lessons learned
- Large plan documents can be solved automatically. 344 pages alone are no reason to push a document into the accessibility statement as a ‘disproportionate burden’.
- Photo pages and OCR noise need rules of their own. General automation easily mistakes a full-page photo for a background, and text recognition finds letters where there are none. Only the targeted rule brought the step from 84 to 100.
- Captions are a good start, not the last word. They immediately provide an accurate statement, but not always a helpful one (Writing good alt text).
- Text recognition over photos is a risk. Anyone who makes documents searchable after the fact should look specifically at photo pages afterwards (Scanned archive documents).
- 100 points is a technical result. It shows that the checkable rules are met, not that everything is understandable. The open points in the test report are part of the acceptance.
Sources:
As of: 10/2026. This article gives a general overview and is not legal advice. The legal and standards texts in force are authoritative; in individual cases, the law of the German states (Länder) may differ.