How do you compress a PDF without losing searchable text?
In PrivConvert, choose Optimize images or Lossless when text selection or fillable fields matter. Optimize images can reduce eligible image data while retaining document structure. Lossless repacks PDF data without lowering image quality. Balanced and Strong render whole pages as images; if that candidate is returned, the original text layer and form fields are gone.
Keep “Return the original when the result is not smaller” enabled for everyday use. Download the result and check both its size and the features you need before replacing your original.
What happened to our two test documents?
Both generated documents contain two pages of selectable text and one fillable field. The illustrated document adds a noisy color test image. We disabled the keep-original option for the six mode comparisons so a larger candidate would remain visible. Sizes below are bytes; each result links to the actual file.
Scroll the table sideways to see all columns.
| Document / mode | Bytes | Extracted characters | Form fields |
|---|---|---|---|
| Text / Original | 6,392 | 3,106 | 1 |
| Text / Lossless | 3,461 | 3,106 | 1 |
| Text / Optimize images | 3,461 | 3,106 | 1 |
| Text / Balanced (flattened) | 397,368 | 0 | 0 |
| Illustrated / Original | 1,804,610 | 1,118 | 1 |
| Illustrated / Lossless | 1,795,479 | 1,118 | 1 |
| Illustrated / Optimize images | 398,058 | 1,118 | 1 |
| Illustrated / Balanced (flattened) | 425,492 | 0 | 0 |
For the illustrated sample, Optimize images was 77.94% smaller and kept the measured text and form field. Balanced was 76.42% smaller but lost both. Lossless reduced this image-heavy sample by only 0.51%. These are different tradeoffs, not a ranking that applies to every PDF.
The small text document tells a different story: both structure-preserving modes reduced it from 6,392 to 3,461 bytes, while Balanced grew it to 397,368 bytes. Turning on the default keep-original option returned a 6,392-byte file identical to the source by SHA-256. In that case the original text remains because the flattened candidate was not returned.
Repeat the PDF test
- Download the text sample and illustrated sample. These contain generated test data, not customer documents.
- Open Compress PDF, select one sample, and choose Lossless, Optimize images, or Balanced. Keep metadata removal off.
- For the comparison only, turn off “Return the original when the result is not smaller.” Run each mode and save the output.
- Check file size, select and copy a sentence, and try the field on page two. Use Extract PDF Text to inspect the text layer.
- Repeat Balanced on the text sample with the keep-original option on. Confirm the unchanged-result message.
The tested Balanced preset renders at 150 DPI with JPEG quality 0.8. The Strong preset uses 96 DPI and quality 0.55; it was not included in this measured table. The structure-preserving paths use qpdf in the browser. The qpdf command-line documentation explains the distinction between stream compression and optional image optimization.
What this test does not prove
We checked extracted page text, marker presence, page count, and form-field count. We did not audit every link, annotation, digital signature, reading order, or accessibility tag. Equal text counts alone do not establish complete fidelity. The samples are not scanned contracts, signed documents, or a representative set of real-world PDFs.
A scan that has no text layer will not become searchable just by compressing it. Use OCR PDF for text recognition, then check its output and recognition accuracy separately. Preserve the original if signatures, accessibility, or exact archival fidelity matter.
Test record and corrections
These are first-party observations, not independent certification. Tests used Chrome 150 on macOS 15.3 and the live PrivConvert tools at revision 8f7181b. File-size and structure checks used the downloaded outputs, not estimated encoder settings.
- Full results, exact byte counts, settings, and SHA-256 hashes
- Input sample manifest and reproduction notes
- Input generator source and output inspection script
See what local processing covers. To report a different result or an error in this guide, contact PrivConvert with the tool, settings, and a non-sensitive example.