Original benchmark · downloadable dataset

Browser PDF tools accuracy benchmark: 22 workflows tested

We ran adversarial synthetic files through every GoPDFConverter workflow, then inspected the downloaded outputs with tools independent of the browser code. This page publishes the cases, results, limitations, and checksums.

Bottom line

All 22 tested workflows produced valid outputs that matched the requested operation for their named stress case. That does not mean every possible PDF is supported; it means each published result below has direct output evidence rather than a page-load check.

22/22Named browser workflows produced the expected tested result
100One-page PDFs validated inside the largest split ZIP
94.4%Reduction on the 9.2 MB image-heavy maximum-compression fixture
3 sizesDesktop, tablet, and mobile layouts checked separately

Results by workflow

The checksum identifies the exact downloaded artifact used for the recorded result. Word conversion rows were re-run after the v81 reconstruction upgrade; the remaining rows come from the full real-output release audit completed on the same fixture corpus.

WorkflowStress caseIndependently verified result
Merge PDF3 PDFs, 18 pages, reordered on mobile18 pages; exact reordered marker sequence preserved
Split PDF100-page PDF split every page100 valid one-page PDFs; markers 1, 50, and 100 preserved
Compress PDF9.2 MB four-page image PDF, Maximum preset516 KB, four valid pages, visually correct; 94.4% smaller
Rotate PDFPages initially at 0°, 90°, 180°, and 270°Output rotations 90°, 90°, 180°, and 270° as requested
Organize PDFMove page 2 first; delete page 311 pages; marker order 2, 1, 4 through 12
Delete PagesDelete pages 3 and 710 pages; only the requested markers absent
Extract PagesExtract pages 2, 5, and 9Three pages in the exact requested order
Sign PDFTyped signature on page 5, top-leftSignature image appears only on page 5 at the requested position
Watermark PDFRed 50% watermark on pages 2 and 4Text exists only on pages 2 and 4; render is visually correct
Page NumbersPages 2-12, starting at 10, “Page n of N”Page 1 unchanged; pages 2-12 labeled 10 through 20
Crop PDFMixed sizes; crop pages 2-4 by 10 mmCropBox changed by 28.35 points only on pages 2-4
Protect PDF12 pages, strong passwordBlocked without password; every page readable with it
Unlock PDFWrong password, then correct passwordFriendly rejection, followed by lossless unlocked 12-page output
PDF to TextPages 2, 5, and 9 with separators8,608 characters; only requested page markers present
OCR PDFImage-only scanUnique marker and numbered lines recognized in TXT and valid DOCX
PDF to WordThree pages, real table, image, columns, headers, footers, landscapeThree pages; landscape geometry, numbered lists, headers, footers, and left-before-right order preserved
Word to PDFDOCX with headings, styles, lists, merged table, image, headers, footersSearchable Letter PDF; geometry, constrained image, and basic header/footer preserved
JPG to PDF1.4 MB photo plus small JPGTwo correct pages; optimized output 814 KB
PNG to PDFTransparent PNGOne valid page; transparency displays correctly
HEIC to PDFReal HEIC photo containerOne visually correct PDF page; 810 KB
PDF to JPGPages 2 and 5 at 144 DPITwo 1,224 × 1,584 JPEG files in a valid ZIP
PDF to PNGPages 1 and 3 at 200 DPITwo 1,700 × 2,200 opaque-white PNG files in a valid ZIP

How the outputs were checked

  1. 1. Synthetic fixtures

    Files contain unique page and content markers, deliberate rotations, transparency, images, columns, encryption, and known geometry. No private user documents are used.

  2. 2. Real browser workflows

    Files are selected through the interface, controls are changed, processing finishes, and the actual download is captured.

  3. 3. Independent parsers

    Outputs are read with pypdf, pdfinfo, python-docx, ZIP validation, and text extraction rather than trusting the code that created them.

  4. 4. Visual rendering

    Poppler and ImageMagick renders catch clipping, black transparency, missing images, incorrect placement, and page-orientation errors that metadata alone misses.

  5. 5. Checksums

    Every row in the CSV includes a SHA-256 hash for the exact output evaluated.

Important limits

This is a product validation dataset, not a market-wide competitor ranking. A passing row applies to the stated fixture and tested browser workflow. Browser memory, damaged files, unusual encodings, handwriting, certificate-backed signatures, PDF/A requirements, and highly designed Word layouts can require specialized desktop software.

The Word conversion tests are deliberately demanding because PDF and DOCX encode layout differently. The PDF to Word tool offers editable reconstruction and an appearance-preserving visual mode; the Word to PDF tool labels its browser-pagination limits directly.

How to cite or reuse the data

The CSV is licensed under CC BY 4.0. Cite “GoPDFConverter Browser PDF Tools Accuracy Benchmark 2026,” link to this page, and name the workflow and stress case you use. Do not generalize a fixture result into a guarantee for every document.