What PDF compression actually does to your file

"Compress PDF" sounds like one operation. It is really three or four different ones, and which of them helps depends entirely on what is inside your file.

A PDF is a container, not an image

Inside a PDF are objects: text with font references, vector drawings, embedded images, metadata, annotations, form fields. Compressing the file means making those objects smaller or throwing some of them away. It does not mean re-encoding the page as a picture.

That is why the three levels do quite different amounts of work.

• Light (PDF Compress) repacks the file structure and nothing else. Many small objects get packed into compressed object streams, which removes overhead the original generator left behind. Page content comes out byte-identical.
• Balanced repacks and clears document metadata — title, author, subject, keywords, producer. Small, free, and it strips the author name your word processor quietly attached.
• Strong also drops annotations and form data. In a document that has been through a review cycle, comment threads and form entries are frequently the largest removable objects in the file.

Why your text PDF barely moved

A PDF exported from a word processor is mostly text plus font subsets, already compressed by whatever produced it. There is very little slack left. Ten to twenty percent from repacking is a normal result, and almost nothing is not a bug.

Why your scan barely moved either

This is the case worth understanding. A scanned page is a JPEG wrapped in a PDF, and repacking the container does not touch the JPEG — the image is copied across untouched, at whatever resolution the scanner produced.

The way to genuinely shrink a scan is to re-render it: convert the PDF to JPG at a moderate scale, then convert those images back into a PDF. You are choosing a resolution and a quality level, which is a lossy decision — but it is the only one that moves a 40 MB scan. Work on a copy, and read the result at full size before sending it anywhere.

Merge and split are lossless

Merge and split copy whole page objects from one document into another rather than re-encoding anything, so a merged file is the sum of its inputs and an extracted range is a faithful copy of those pages. Splitting a 200-page report into ten files will not degrade a single glyph.

Two practical notes

Password-protected files have to be unlocked in a reader first — an encrypted document cannot be parsed without its password, and no amount of clever handling changes that.

And all of this happens in the browser. A contract, a payslip or a medical scan is read by the page on your own machine; there is no upload, so there is no copy sitting in someone's storage bucket afterwards.