Why PDF Files Get Large
A PDF that will not fit in an email is a daily annoyance, and the usual response is to run it through a compressor and hope. That works sometimes and does nothing other times, because PDFs get large for four quite different reasons and each needs a different remedy.
What actually occupies the space
A PDF is a collection of objects: page descriptions, fonts, images, and assorted metadata. Text and vector graphics are tiny — a page of prose is a few kilobytes of instructions. Anything surprisingly large is therefore not the text, and compressing the text harder will not help.
In practice the weight is almost always one of four things: embedded raster images, embedded fonts, retained revision history, or accumulated metadata and attachments. They are present in very different proportions depending on how the file was produced.
Before compressing, it is worth guessing which. Divide the file size by the page count. Under 100 KB per page is lean. Around half a megabyte suggests images. Several megabytes per page means full-page scans or photographs at high resolution.
Embedded images, the usual culprit
Raster images dominate almost every large PDF. A photograph pasted from a camera is often several thousand pixels across, and it keeps every one of those pixels even when it is displayed two centimetres wide. The file stores the original, not what you see.
Scanned documents are the extreme case, because every page is a full-page image. A 300 DPI colour scan is roughly 8.7 million pixels per page, which is typically one to three megabytes even after compression, repeated for every sheet with nothing shared between them.
Downsampling is what helps here, and it is the one thing generic compressors do well. Reducing images to the resolution actually needed — 150 DPI for screen reading, 300 for print — commonly cuts a scanned file by three quarters with no visible difference on screen.
Fonts
PDFs embed the fonts they use so the document looks identical everywhere, and that embedding has a cost. A single Latin font subset is usually tens of kilobytes, which is negligible. A full CJK font is a different matter: Japanese and Chinese typefaces contain thousands of glyphs and can run to several megabytes each.
Subsetting is the standard mitigation. Instead of embedding the whole typeface, the exporter includes only the glyphs the document actually uses, which for a short Japanese document might be a few hundred characters instead of twenty thousand. Most modern exporters subset by default, but some older tools and some print-oriented presets do not.
The symptom of a font problem is a small document that is inexplicably heavy: a two-page letter at four megabytes with no images in it. Check the document properties, which list the embedded fonts, before reaching for an image compressor that has nothing to work on.
History, metadata and attachments
PDFs support incremental saving: rather than rewriting the file, an editor can append changes to the end. This makes saving fast and preserves an audit trail, but it means that deleted content may still be present in the file, and a heavily edited document can be several times larger than its visible content.
This has a privacy dimension as well as a size one. Text that was removed before the file was sent may still be recoverable from an earlier revision inside the same file. Redaction that consists of drawing a black rectangle over something is not redaction at all, since the covered text is still in the document.
Saving a clean copy, sometimes called "save as optimised" or "linearise", rewrites the file from scratch and drops the accumulated history. Form fields, annotations, embedded attachments and colour profiles also add weight and can be discarded if the recipient does not need them.
When converting to images helps, and when it does not
Rendering pages to images and packing them into an archive is sometimes smaller than the original and sometimes much larger, and it is worth knowing which to expect. For a text PDF the output is almost always larger, because compact character instructions are replaced by millions of pixels.
For a scanned PDF at a high resolution it can genuinely shrink, if you render below the scan's own resolution. A 600 DPI scan exported at scale 2x — about 144 DPI — holds a fraction of the data, which is a reasonable trade when the pages only need to be read on screen.
Note that converting to images discards the text layer along with the size. The result cannot be searched, copied from, or read by a screen reader. It is the right choice for sharing a visual snapshot of a page, and the wrong one for archiving a document someone will need to work with.