PDF to Word text versus scan: a controlled editability test

We converted the same one-page design twice: once with real PDF text and once as a full-page image. The visible sources matched, but the Word documents did not contain the same kind of content.

Controlled test result

Only the PDF with a text layer produced editable words

Searchable source50 extracted words; no page image
Searchable DOCX312 text characters, seven paragraphs, and one table
Scanned sourceZero extracted words; one full-page image
Scanned DOCXZero text characters; one inline image

The sentinel phrase cobalt river 4729 was present in the searchable DOCX XML and absent from the scanned DOCX. This is a direct content check, not a judgment made from screenshots.

Fixture and reproducible method

The controlled Letter-size page contains a heading, four prose lines, a three-column ruled table, names, numbers, and the sentinel phrase. The searchable source stores those elements as PDF text and vector rules. The scan source was made by rendering that same page at twice its base scale and placing the resulting raster image over a new 612 x 792 point PDF page.

  1. Generate both sources with the repository script at scripts/evidence/generate_fixtures.py.
  2. Submit each PDF to the current Converttwo /api/convert/pdf-to-docx route.
  3. Inspect DOCX ZIP integrity, OOXML text, paragraphs, tables, drawings, and sentinel content.
  4. Render each DOCX with LibreOffice and compare it with the corresponding PDF at 108 pixels per inch.

The September 14 run used current application code at commit 3caf8a7, Python 3.11, pdf2docx 0.5.13, PyMuPDF 1.28.0, and LibreOffice 26.2.5.2 on Windows. The repository deployment pin is pdf2docx 0.5.8, so exact reconstruction details can vary on another environment.

What the rendered Word files looked like

Searchable PDF source: characters and vector rules
Searchable result: editable text, with “Item Alpha” clipped to “Item Alph”
Scanned PDF source: one full-page raster image
Scanned result: appearance retained as an image, not editable words

Measured source and output properties

Exact values from the published files
PropertySearchable pathScanned path
Source PDF4,055 bytes; 50 words; zero images56,166 bytes; zero words; one page image
Converted DOCX37,554 bytes; seven paragraphs; one table; no inline image81,856 bytes; one paragraph; no table; one inline image
Text in DOCX XML312 characters; sentinel foundZero characters; sentinel absent
Rendered outputOne Letter page; 48 extractable words; no page imageOne Letter page; zero extractable words; one page image
Pixel comparison2.96% of pixels differed from the searchable source at 108 ppi9.32% of pixels differed from the scan source at 108 ppi

What this result means in practice

A page can look readable without containing any readable text for the converter. In this test, the scan remained visible because it was carried into Word as a picture, but searching or editing the printed words was impossible. The converter does not perform OCR.

The searchable result was editable and reconstructed the ruled region as a real Word table, but editability did not guarantee perfect transcription or geometry: the final letter in “Item Alpha” was clipped in the rendered DOCX and line positions shifted slightly. A converted DOCX should therefore be proofread even when its source has a clean text layer.

Limit of this test

These are simple, generated one-page sources. They do not predict performance for handwriting, skewed camera scans, unusual fonts, multilingual OCR, complex forms, or damaged PDFs. No OCR engine was tested because Converttwo does not include one.

Download the exact test files

Test your own text-based PDF

Keep the source PDF beside the result and verify editable words, tables, names, and numbers.

Open PDF to Word

Related evidence and guidance