Compressing a PDF without losing quality, what actually changes

Published on September 20, 2026. 4 min read

Every PDF compressor shows a percentage. None of them says what was removed to get it. So we opened the files.

Three real documents, compressed with our engine and with a widely used service, then compared object by object. Measurements taken in September 2026.

What a PDF actually holds

A PDF is not an image. It is a set of objects: text with its fonts, vector drawings, images, and invisible markers that describe the structure of the document.

Compressing a PDF means pulling on three levers, and they do not cost the same.

Rewriting the structure. Grouping objects, dropping what is no longer referenced, recompressing streams. Modest gain, no loss.

Reducing the images. Lowering the number of dots that make up each image. This is the main lever, and the only one that degrades.

Removing objects. Dropping metadata, thumbnails, sometimes structure markers. Variable gain, and this is where the bad surprises hide.

Case 1, a 482 page manual, 19 MB

Our engine returns 17.7 MB, 7% smaller, without touching a single image. The compared service returns 19 MB, 1% smaller.

Why so little on both sides? Because the 1,035 images in this manual are already at screen resolution. There is nothing to take without spoiling it. Our 7% comes only from rewriting the structure, so with no loss at all.

This case is the most common and the least talked about: most professional documents are already optimised, and the only honest gain there is modest.

Case 2, a 68 kB invoice set, with no images at all

Here we lose. Our engine returns 38.5 kB, 43% smaller. The compared service returns 31.8 kB, 53% smaller.

We opened both files to understand. Text is intact on both sides, so are the fonts. The difference lies elsewhere: this document holds 233 structure markers, headings, rows and table headers. They are what lets a screen reader announce which column each figure belongs to.

They are intact with us. They are gone with them.

In other words, part of their advantage comes from a document that became less accessible. On an invoice, that means a blind person hears a series of numbers without knowing which are amounts and which are dates.

Case 3, a scanned medical atlas, 186 MB

Our engine returns 97.5 MB, or 51.5 MB at the strong level. The compared service returns 70.8 MB, or 27.3 MB at its strongest setting.

They go lower, and we counted why. We added up the dots that make up the images in both files: theirs keeps half as many. The rest is identical to within four percent, including the way each dot is encoded.

So their whole advantage comes from resolution removed from the photographs. In a dermatology atlas, that resolution is what lets you recognise a lesion when you zoom in. The document still looks fine as a thumbnail and becomes useless in practice.

One detail measured along the way: in this document the background photographs were already at low resolution, 72 dots per inch. They were reduced once more anyway. An engine that measures resolution before acting would never have touched them.

What we took from it

Three rules, applied by our engine.

Measure before acting. The useful resolution of an image depends on its display size, not on its number of dots. An image already below the target is never touched.

Never remove what weighs nothing. The structure markers of a document weigh a few kilobytes and carry its accessibility. Removing them to gain a percentage is a bad trade.

Return the original when there is nothing to gain. A file two percent lighter but degraded has no value. We would rather say so.

And speed

The 482 page manual goes through in nine seconds. The 186 MB atlas in thirty six. Time follows the number of dots to decode, never the size of the file.

Reproducing these measurements

File sizes can be checked in any file explorer. The number of dots in an image and the presence of structure markers can be read with free tools, qpdf and pikepdf, which we use ourselves. We keep the date of these measurements up to date, because a compression engine changes, and the figures with it.