What these tools actually do to your files

Why this page exists

A compression tool that promises “smaller files” is making a claim it usually cannot keep for every document. Whether a PDF shrinks depends almost entirely on what is inside it, and the honest answer to “how much smaller will my file get?” is a table, not a number. So here is the table.

Everything below was produced by uploading documents to DocTools over the public internet, exactly as a visitor's browser does. Nothing was measured against a private build or a local shortcut.

How the test works

Three documents were generated, each representing a kind of file people actually upload:

• A text report of 12 pages: real text characters and vector rules, the shape of an exported statement or letter.
• A scanned document of 6 pages at 200 DPI, where every page is one large photograph of paper and there is no selectable text at all.
• A photo brochure of 8 pages: text over full-bleed photographs, the shape of a design export.

Every input is synthetic and generated by a script. No identity document, no photograph of a person, no signature and no personal data was used at any point, which matters because the alternative would be testing with material that should never be uploaded to anyone's server, including mine.

Each file was sent to all three compression settings in turn. The measurements record the returned size, the time from the start of the upload to the end of the download, and whether the text in the result was still selectable afterwards. Timings therefore include the network, not just the server.

Compression results

Read down the columns and the useful conclusion appears immediately: the setting matters far less than the document does.

A scan is where compression pays

The scanned document dropped to 24.2% of its original size on the screen setting, which is the single biggest saving in the whole table. It is also the file type most people are trying to shrink, because a phone scan of a few pages routinely arrives as several megabytes.

The same file on the balanced setting came back at 100.0% — that is, no smaller at all. The images inside it were already compressed about as hard as that profile compresses. Nothing was gained and nothing was lost.

A text document has almost nothing to give

The text report only reached 86.5% on the screen setting, and it started at 16 KB. Text is stored as characters, not pixels, and characters are already compact. If a text-heavy PDF is too large for an upload limit, compression is rarely the answer; the size is usually coming from embedded fonts or an image you have forgotten is in there.

Photographs sit in between

The brochure reached 40.9% on screen and around 80% on the other two profiles. Its text stayed selectable in every run, which is the property that matters: the pages were never flattened into pictures.

Building a PDF from photographs

Fitting photographs to A4 produced a smaller file than letting each page match its image, because the images are scaled down to fit the page. Choose A4 when a portal expects a document; choose the image's own size when the proportions matter more than the page.

Resolution is the expensive choice here: going from 72 to 200 DPI made the output five times larger and took five times as long, for twelve pages of text that most people will only read on a screen. Render at 150 DPI when the pages will be looked at, and at 200 only when they will be printed or zoomed into.

What these numbers do not tell you

Honest limits, since a test with no stated limits is marketing:

• These are synthetic documents. Yours will differ, and the percentages will differ with them. The shape of the result — scans shrink, text does not — is the durable part.
• Each figure is a single run, not an average. Times in particular vary with the network.
• Only the first page of each result was checked for selectable text.
• Compression quality is judged here by size and by whether text survived, not by a perceptual comparison of image detail. Look at your own output before relying on it.

Run it yourself

The script that generated these documents and collected these numbers is lab/benchmark.py in the site's own repository. It builds the test files, calls the same public API available to anyone, and writes the JSON this page is generated from. This page is built from that JSON rather than written by hand, so the prose cannot drift away from the measurements.

Measured 11 September 2026 against Ghostscript 10.02.1 and Inkscape 1.2.2 (b0a8486541, 2022-12-01). When the tools change materially, this page is re-measured and the date above changes with it.