A PDF is a container, not an image format, and almost every confusing thing about PDF tools follows from that. Inside one file you can have selectable text with embedded font subsets, vector paths, raster images, form fields, annotations and digital signatures — all with their own compression. This is why "compress PDF" produces wildly different results depending on what you feed it. A 40-page text document is already close to optimal and might shrink 5%; a 40-page scan is really forty JPEGs in a wrapper and will often drop by 80% or more.
Structural operations — merging, splitting, reordering, rotating, adding page numbers — are lossless. They rewrite the document's page tree and object references without touching the content streams, so nothing is re-encoded and nothing degrades. You can merge and split the same file repeatedly with no quality cost. Compression is the opposite: it works by downsampling and re-encoding the embedded images, so it is a one-way operation and worth doing last, after all your structural edits are final.
PDF encryption deserves a clear explanation because the format supports two different passwords and they are not equally meaningful. A user password is required to open the document at all, and with AES-256 that is genuine cryptography. An owner password only sets permission flags — "no printing", "no copying" — which any conforming reader is free to ignore, and many do. If you need a PDF to actually stay private, set a user password; treat owner permissions as a polite request rather than a control.
Similarly, placing a drawn or typed signature onto a page is a visual signature. It is what the large e-signature services produce in the common case, and it is sufficient for most agreements in most jurisdictions, but it is not a cryptographic PKI signature bound to a certificate. If a counterparty specifically requires a qualified or advanced electronic signature, you need a certificate-based signing service, not an image placed on a page.
What each PDF operation actually changes
Lossless operations can be repeated freely. Lossy ones should be done once, at the end:
| Operation | Changes | Lossless? | Typical size effect |
|---|---|---|---|
| Merge / Split | Page tree only | Yes | Sum of the parts |
| Add page numbers / watermark | Adds a content layer | Yes | Negligible increase |
| Resize pages | Page boxes and scaling | Yes | Roughly unchanged |
| Protect (AES) | Encrypts streams | Yes | Slight increase |
| Compress | Downsamples embedded images | No | 5% on text, 60–90% on scans |
| PDF to JPG | Rasterises whole pages | No | Text becomes unselectable pixels |