How much smaller does PDF compression actually make your file?
We put 12 documents across 6 kinds through all three compression presets on our own tool and recorded what came back — 36 measurements in all. The short answer is that the preset matters far less than what is on the pages.
Compression works almost entirely by reducing the resolution of images. A page of photographs can lose 84.8% of its weight; a page of type gives up almost nothing at any setting. On 4 of 12 documents the gentlest preset returned the file byte for byte unchanged.
How this was measured
Twelve documents were generated covering six categories that behave differently under compression: text only, ruled tables, photographic pages, scanned pages with realistic speckle, line-art charts, and mixed pages combining text with one image. Two documents of each kind, so that a single odd file cannot carry a category.
Each document went through the live site at all three presets — the same High Quality, Balanced and Maximum buttons a visitor presses — through the same endpoint a browser uses. The figure recorded is the size of the file the server actually returned, not an estimate.
The three presets are the only three outcomes that exist: the backend buckets the quality value, so every setting between them produces one of these three results.
You do not have to take any of this on trust. Download the raw dataset (CSV) — all 36 rows, with before and after byte counts for every document.
Reduction by document type
Average percentage smaller, by what is on the pages. Sorted by what the strongest preset achieved.
| Document type | Docs | High Quality | Balanced | Maximum |
|---|---|---|---|---|
| Text only | 2 | 11.9% | 18.3% | 18.3% |
| Mixed | 2 | 21.8% | 27.6% | 27.6% |
| Ruled table | 2 | 42.4% | 46.5% | 46.5% |
| Line art / chart | 2 | 0.0% | 36.9% | 72.2% |
| Scanned page | 2 | 0.0% | 56.5% | 83.9% |
| Photographic | 2 | 0.9% | 66.1% | 84.3% |
The spread within each preset
| Preset | Smallest saving | Largest saving | Median |
|---|---|---|---|
| High Quality | 0.0% | 51.7% | 5.0% |
| Balanced | 16.1% | 66.4% | 38.0% |
| Maximum | 16.1% | 84.8% | 63.4% |
That spread is the finding. Reading down a column tells you almost nothing about what your file will do; reading across a row of the table above tells you a great deal.
Three things worth knowing before you compress anything
1. The gentlest preset often does nothing at all
High Quality left 4 of the 12 documents byte for byte identical, and changed 6 of them by less than 2%. The documents it did nothing to were the scanned pages and the charts — which are exactly the files people are usually trying to shrink. If that preset appears to have failed, it has not; it does not touch images, and those pages are nothing but images.
2. On a text document, the strongest preset buys you nothing
Balanced and Maximum returned the same result on 6 of the 12 documents. Once there are no images left to downsample, a stronger setting has nothing further to take, so you are accepting a quality risk for a saving that does not exist. Where the two differ, it is because the pages carry photographs.
3. The saving comes out of resolution, which means it is a trade
Nothing here is free. A scanned page at Maximum loses most of its weight because its image is being re-encoded at a resolution a screen can show and a printer cannot. That is the right trade for an email attachment and the wrong one for anything going back onto paper. Compress a copy, and keep the original.
What this study does not show
It measures our compression, not compression in general. Another tool using different settings will produce different numbers, and we have not tested anyone else's.
The corpus is twelve generated documents, not a sample of real-world files. Generated documents are cleaner than scans off an office multifunction device, so a genuine scan with heavier noise may compress further than the figures here, not less.
And it measures file size only. It does not grade how the result looks, which is the thing you actually care about at the strongest setting — that judgement is yours, on your own document, which is why the honest advice is to compress a copy and look at it before you send it.
Run it yourself
The corpus generator and the measurement harness are in our repository. The dataset on this page is its output, unedited.
Download the dataset (CSV) — or put your own file through Compress PDF at each of the three presets and compare the byte counts. On your own document that comparison takes about a minute and is worth more than any average.
Try it on your own document
The only number that matters is the one your file produces.
Compress a PDF How compression works