What PDF compression can and cannot do

Compression is the most oversold operation in the PDF world. Sites advertise percentages they cannot deliver on the file you are holding, and the reason is that "PDF" describes two completely different kinds of document that happen to share an extension.

Two documents, one format

A PDF exported from a word processor is a set of drawing instructions: place this glyph here, in this font, at this size. It is text and vectors, and it is already extraordinarily compact. A forty-page report might be 300 KB.

A PDF from a scanner is a stack of photographs. Each page is a JPEG, already compressed by the scanner, wrapped in just enough PDF to hold it. The same forty pages might be 40 MB.

Almost everything sold as PDF compression only ever mattered to the second kind.

What the usual techniques achieve

We measured a six-page scanned document, 2.5 MB, against each technique in turn:

• Re-saving with object streams: 0%
• Stripping metadata: 0%
• Structural garbage collection with a full PDF toolkit: 0%

Not "a few percent". Nothing measurable. Every one of those techniques tidies the *container*, and the container was never the problem — the images inside it were, and none of those methods touch them.

What actually works

Re-encoding the page images. Lower the resolution, lower the JPEG quality, put them back:

• 150 DPI at 75% quality: −5%
• 120 DPI at 60% quality: −53%
• 96 DPI at 50% quality: −70%

Real reductions, and they come entirely from throwing away image data. That is the whole mechanism. Anyone promising you a large reduction with no quality loss on a scan is describing something that does not exist.

The part that catches people out

Run that same treatment on the text document — the 300 KB report — and it comes back fifty times larger, with the text layer destroyed. Every searchable word becomes a picture of a word. The file is bigger, worse, and no longer works with a screen reader.

This is not a hypothetical. It is what happens when a compression tool applies one strategy to every input, which most of them do, because checking is more work than not checking.

A practical rule

Before compressing, ask what you are holding. If you can select text on the page with your cursor, the document is already about as small as it can honestly get, and compression has nothing to offer you but harm. If you cannot — if it is a scan — then compression is a genuine trade, and you should choose where on that trade you want to sit rather than accepting whatever a preset picked.

If a tool cannot tell you which kind of file you gave it, it is not in a position to tell you what it is about to do to it.