At a glance
Compare the exact source PDF with the exact HTML revision. Check text presence, order and meaning separately, with special attention to numbers, units, negation, dates and link targets. An extraction-similarity score can help locate omissions but cannot establish correct semantics or accessible reading order.
Small textual changes can matter more than a high overall similarity score.
| Source | Converted text | Review finding |
|---|---|---|
| Do not submit originals | Submit originals | Negation lost |
| Fee: $1,250 | Fee: $1250 | Verify formatting and value |
| 0.5 mg | 5 mg | Decimal lost; high consequence |
| Apply by 12 October | Apply by 21 October | Date changed |
Scroll sideways to see all columns. Use arrow keys when the table is focused.
Establish the comparison pair
Record the source byte digest and output revision or digest before review. Keep a stable copy of both while logging differences. A page served today may have changed since a report or screenshot was created.
Review the rendered source as well as extracted text. PDF text can be positioned fragments, scanned images or an OCR layer that disagrees with what the page visibly shows. Treat the source image as evidence for recognition checks, while obtaining owner clarification when the authored source itself is ambiguous.
- Fix the source and output revisions.
- Check all important sections for presence.
- Compare meaningful reading sequence.
- Verify consequential text and links in context.
- Log differences before approving the revision.
Choose a test set by consequence
For every document, check headings, opening and closing sections, page transitions and any text near figures or tables. Add identifiers, contact details, instructions, dates and conditions that drive a reader’s action. Long documents need a stated sampling policy, with important content reviewed beyond the sample.
Check whether repeated page headers became body text, whether hyphenated line wraps changed words and whether footnotes still refer to the right passage. These are different defects; one global text percentage cannot explain all of them.
- Numbers and units: compare signs, decimals, ranges and totals.
- Conditions: verify not, unless, only, before and after.
- Links: compare destination and descriptive text.
- Sequence: read paragraphs around column and page boundaries.
- OCR: compare questionable characters directly with the image.
Record what passed and what remains unknown
Use the conversion QA log to name the exact source location, output element, defect, correction and verification result. After correction, repeat the affected comparison and any checks that depend on it. A changed heading can affect both meaning and navigation.
Text fidelity is one dimension. Inspect tables, figures, forms, keyboard interaction and the assistive-technology experience separately. A readable visual page and complete extracted text can still conceal a confusing document structure.
Sources and scope
Methods, worksheets and illustrative examples are DocAccessible editorial contributions. These references support the technical or product details described above. W3C techniques are informative guidance; they are not a certification scheme.
- pypdf: text extraction
Why text extraction differs from OCR and cannot establish all visual or semantic meaning.
- W3C PDF3: reading and tab order
The relationship between PDF tag order, reading sequence and interactive elements.
- W3C PDF7: OCR for scanned documents
Recovering actual text from scanned pages and checking the result.
Written by the DocAccessible team under our editorial policy. Suggest a correction if a source or product detail needs updating.