Our test: the same PDF appearance, different structure

The test file contains invented monthly figures, three headings and a simple table. The before-file has no tag structure. The after-file was produced using our PDF engine and an explicit structure plan. The content is synthetic; the reported measurements are real.

This tests a bounded technical repair. It does not prove an entire customer workflow, every document type or an automatic assessment process.

A PDF can look finished while lacking information about what its text elements mean. We built a small document whose content is easy to verify: monthly figures, headings and a table. This limits variables and makes the distinction between appearance and structure visible.

The before-file contains the visible information without a tag tree. The after-file gives that content roles and relationships: a main heading, two subheadings, column headers associated with data cells and a document language of Swedish as used in Sweden.

The technical question is narrow: can those structural properties be added while preserving text and the rendered page? That differs from establishing usability in every reading situation. We therefore distinguish measured machine results from outstanding practical checks.

Validation used veraPDF 1.30.2 with its PDF/UA-1 profile for both files. The same tool and profile make the runs comparable, but do not create a percentage in which each check has equal significance for readers.

The text comparison examined preservation of text content. It supports the conclusion that the transformation retained words and figures under the method used. It does not establish correct semantics for every tag or a useful reading sequence.

The image comparison used PDFium at scale 1. Pixels were identical in that rendering. This is a specific appearance result, not a promise about every PDF reader, zoom level or print output.

Together, the comparisons show that structure could be added without changing this test page’s visible content. They complement one another because each examines different properties; none replaces the others.

The after-file passed 2,004 individual automated checks in the specified profile. Check counts differ between the files because their structure differs. Do not turn 55 repaired checks into an overall accessibility percentage.

The useful finding is that the visible page could be preserved while adding machine-readable structure. Zero failures in this profile still does not establish that all information works for all readers.

Sources: [1]

PAC assessment, practical testing with Adobe Reader and NVDA, independent AI review and human delivery approval are not included. This example must therefore not be described as a fully assessed customer case or a certified PDF.

The next practical step would be heading navigation and table reading in a recorded PDF-reader and screen-reader combination. A suitable task is locating February’s 105 hours and comparing January. That behaviour must be observed before describing how the file works in that environment.

A small synthetic document cannot establish handling of complex notes, forms, mathematics or large reports. Each additional content type needs checks appropriate to its requirements. This example is useful because its question and limits are explicit.

• The generator is scripts/create-seo-pdf-demo.py in the project checkout.
• Raw reports, the structure plan and results were retained with the files.
• The findings apply to these hashes. A modified file needs new assessment evidence.

fore.pdf
e8daf04f9e5e0e8dd2e33b1e9dcd1bd90e8503fd118bfb6f6128050bd17534a6

efter.pdf
3cf2192f183a96b8baf66748563c98c8fcc6ada1165581cc62bd65f44c988bf6