We Compressed 24 Real PDFs: the Median Saving Was 12%, Not 70%

Short answer: we compressed 24 public documents and the median saving was 12%, not the 40–70% that gets quoted everywhere. Two files came back bigger. The full result table is at the bottom, and the numbers below are the ones we then used to correct our own guide.

Every "compress your PDF" page, including ours until today, quotes a confident range. Almost none of them say where the number came from. So we measured it.

What we measured, and how

The corpus: 24 public documents, chosen to be real rather than convenient — synthetic files don't compress like real ones, because real documents carry scanner noise, embedded fonts and re-saved images that a generated page doesn't have.

Source	What it is	Files
IRS	Tax forms and instruction booklets (2–142 pages)	6
arXiv	Open-access research papers, text with figures	8
Internet Archive	Scanned US federal court filings and digitised newspapers	10

The method. Every file went through the same engine our own compress tool uses, with the same settings, one at a time. For each we recorded the byte size before and after, the HTTP status, the page count before and after, and the characters extracted per page. A result only counted if the response was a valid PDF and the page count matched the original — a check that turned out to matter (see "what nearly went wrong").

Classification was measured, not assumed: a document counted as scanned if fewer than 100 characters per page could be extracted from it. That is the difference that decides everything else about compression, and you can't tell it from the filename.

The result

	Default settings
Median saving	11.6%
Mean saving	16.4%
Best	81.2%
Worst	−2.8% (the file grew)
Files saving under 10%	11 of 24
Files that got bigger	2

Nearly half the documents saved less than a tenth of their size. The headline range you see everywhere — 40 to 70 percent — described three of our 24 files.

Scanned documents compress worst, not best

This is the finding that contradicts the common advice most directly, including the advice we had on our own site:

Document type	Median saving
Digital (text, forms, papers)	12.6%
Scanned	2.2%

The usual claim is the opposite — that scans, being image-heavy, are where compression shines. The reason it's wrong is simple once you see it: the scanner already compressed the file. A scanned page is a JPEG, and JPEG is lossy. Re-encoding an already-lossy image a second time gives up quality without buying much space, because the redundancy was removed the first time.

The corollary is more useful than the number: if you want a smaller scan, the setting that matters is on the scanner — scan at 200–300 DPI instead of 600 — not on the compressor afterwards.

Two files got bigger

irs-p17.pdf (142 pages) came back 2.8% larger. A court filing grew by 0.8%. Both are long text documents with almost no images.

There's nothing pathological here. Compression re-writes a PDF's internal structure; when there are no images to re-encode, the rewrite can cost more bytes than the savings it finds. It's a reminder that "compress" is not a guaranteed-smaller operation, and any tool promising otherwise hasn't measured.

The setting nobody changes

The engine has a compression level. Our tool — like most — was using the default. So we ran all 24 files again at the most aggressive level:

	Default	Aggressive
Median saving	11.6%	50.4%
Files under 10%	11 of 24	4 of 24
Files that got bigger	2	0
Scanned documents (median)	2.2%	38.7%

A 142-page tax booklet that grew at default settings shrank by 2.6% aggressively; a 3-page payroll form went from 7.6% to 93.5% — 841 KB down to 54 KB.

Does it damage the document? We checked, because a 93% saving invites suspicion:

• Page count: preserved on every file.
• Extracted text: preserved. irs-f1040 went from 5,207 characters per page to 5,206; the papers and the court filings were identical or within a rounding of it.
• Visual: we rendered the same page from the 841 KB original and the 54 KB output at 150 DPI and compared them. The form's rules, boxes, black-filled headers and 8-point footnotes are indistinguishable.

The real cost is time, not quality — and it is wildly uneven. 22 of the 24 finished the aggressive pass in under 9 seconds, including a 75-page, 6.8 MB paper. The other two — 126- and 142-page tax booklets, dense with embedded fonts — took 4 and 6 minutes, against 45 seconds at default.

So we changed our tool, with a guard

Publishing "there's a setting that quadruples the saving" and not using it would be indefensible. Our compress tool now uses the aggressive level, except on documents over 100 pages, which keep the default. That threshold isn't a guess: it's the line between the 22 files that finished in seconds and the two that took minutes. Nobody should wait five minutes for a download to save a few percent.

Measured again through the live tool, after the change:

	Before	Now
Median saving	11.6%	50.4%
Files saving under 10%	11 of 24	4 of 24
Files that got bigger	2	1
Slowest file	46s	50s

The one document that still comes back bigger is the 142-page booklet that stays on the default path — the guard protects its speed, and its 2.8% growth is the price. We'd rather publish that than pretend the rule is free.

What nearly went wrong

Our first aggressive run reported that irs-p17.pdf had shrunk by 98.2%. It hadn't. The script had written whatever came back to a file and measured its size without checking the HTTP status, so a failed request looked like a spectacular result.

We caught it because the number was too good, re-ran the whole set with two extra gates — status must be 200, page count must match — and the same file's real saving turned out to be 2.6%. Every figure above comes from the validated run; the raw table is published below so anyone can check.

We're including this because a benchmark that doesn't say how it could have been wrong isn't worth much.

What we changed on this site

Two things, both because of these numbers:

1. The guide's claims are gone. It used to say text-heavy documents shrink 20–40% and scanned PDFs 50–80%. Our own measurements contradict both, so the guide now states the measured medians and the range, including the part where a file can grow.
2. The tool compresses harder, under the 100-page guard described above. Median saving on this corpus went from 11.6% to 50.4%.

The full table

24 documents, default settings versus aggressive, all validated. Sizes in bytes.

Document	Type	Pages	Before	Default	Aggressive
irs-f941	digital	3	841,648	7.6%	93.5%
court msnb.180080	digital	2	1,098,795	81.2%	91.8%
arxiv-1505.04597	digital	8	1,648,684	68.9%	91.3%
arxiv-2005.14165	digital	75	6,768,044	1.1%	84.8%
arxiv-2010.11929	digital	22	3,743,814	39.4%	78.0%
arxiv-1512.03385	digital	12	819,383	13.4%	75.5%
irs-f1040	digital	2	220,237	10.6%	71.3%
irs-f1120	digital	6	340,034	12.6%	69.6%
irs-f4562	digital	3	206,401	9.5%	68.8%
court pamd.86362	scanned	5	275,221	0.6%	68.6%
arxiv-1810.04805	digital	16	775,166	15.1%	59.0%
court mnd.138153	digital	5	324,710	15.2%	51.1%
court mnd.221307	digital	16	446,554	6.2%	49.7%
arxiv-1706.03762	digital	15	2,215,244	15.5%	44.2%
arxiv-1412.6980	digital	15	584,641	12.6%	42.0%
court paed.478939	scanned	10	312,714	2.6%	38.7%
arxiv-1409.1556	digital	14	200,010	29.9%	35.4%
irs-i1040gi	digital	126	4,434,643	17.9%	32.5%
court mowd.148012	digital	95	1,394,934	−0.8%	30.5%
court moed.158745	scanned	18	511,911	2.2%	12.8%
court mieb.936608	digital	50	787,047	7.1%	3.5%
irs-p17	digital	142	3,088,632	−2.8%	2.6%
newspaper NDIGCZAS049827	digital	2	530,391	27.1%	1.8%
newspaper KM_01052_1924	digital	2	222,983	0.2%	0.1%

Three documents are worth a second look: the last three rows saved less aggressively than at default. Two are digitised newspapers whose images are already at the quality floor, and mieb.936608 is a 50-page filing where the aggressive pass rewrote structure it couldn't improve. Aggressive is a better bet on average, not a guarantee on any single file — which is the same lesson as the two files that grew.

(Raw CSV with timings, HTTP statuses and page-count checks: docs/olcumler/sikistirma-2026-08-15.csv in the repository.)

If you just want a smaller file

• Check what you actually have. Try to select a line of text. If nothing selects, it's a scan — and compression is not your lever, the scanner's DPI setting is.
• Delete pages before compressing. Removing a blank scanner back-page saves 100% of that page, which beats any setting. Organize Pages does it in a few clicks.
• Compress once. Re-compressing an already-compressed file mostly costs quality for nothing — that's the same lossy-twice problem scans have.
• If it's still too big, split it. Splitting two 10 MB halves past a 25 MB mail limit works when compression can't.

Our compress tool is free to use, and the step-by-step version of all this is in the compress PDF guide.