The short answer
Our scans started at 147.7 KB to 571.4 KB a page, at 300 dpi. At 200 KB a page, all 7 stayed readable. At 100 KB, 6 of 7 did. At 50 KB, only 3, and at 25 KB, 2: the forms and receipts with large, sparse type.
What decides it is not the file size but how many pixels are left for each inch of the page. Every page we tested read cleanly at 87 dpi or more. The highest resolution at which any page failed was 82 dpi. So the rule of thumb: shrink a scan as far as you like, as long as its picture keeps about 87 pixels per inch, and a bit more for small print.
Lossless compression does nothing to a scan
The lossless rewrite saved nothing on 8 of the 8 scans. That is expected: a scan is a JPEG picture already, and a lossless tool cannot touch the picture without decoding it. Any real saving on a scan comes from making the picture smaller, which is a loss, so the only question is how much loss the page can take.
Every page, every size
| Page | Target | Size | Resolution | Error rate | Readable |
|---|---|---|---|---|---|
| Book page | Lossless | 571.4 KB | 300 dpi | 0.0% | Yes |
| Book page | 200 KB | 172.1 KB | 117 dpi | 0.0% | Yes |
| Book page | 100 KB | 82.2 KB | 74 dpi | 5.8% | No |
| Book page | 50 KB | 41.9 KB | 47 dpi | 57.3% | No |
| Book page | 25 KB | 24 KB | 37 dpi | 68.4% | No |
| Letter | Lossless | 440.5 KB | 300 dpi | 0.0% | Yes |
| Letter | 200 KB | 162.2 KB | 142 dpi | 0.0% | Yes |
| Letter | 100 KB | 83.1 KB | 87 dpi | 0.4% | Yes |
| Letter | 50 KB | 43.5 KB | 56 dpi | 36.9% | No |
| Letter | 25 KB | 23.1 KB | 37 dpi | 70.5% | No |
| Two-column page | Lossless | 541.8 KB | 300 dpi | 69.3% | Not scored |
| Two-column page | 200 KB | 170 KB | 122 dpi | 69.3% | Not scored |
| Two-column page | 100 KB | 85 KB | 75 dpi | 71.7% | Not scored |
| Two-column page | 50 KB | 39.7 KB | 48 dpi | 71.7% | Not scored |
| Two-column page | 25 KB | 20.8 KB | 37 dpi | 73.1% | Not scored |
| Bank-style statement | Lossless | 261.7 KB | 300 dpi | 0.0% | Yes |
| Bank-style statement | 200 KB | 151.5 KB | 215 dpi | 0.0% | Yes |
| Bank-style statement | 100 KB | 83.3 KB | 133 dpi | 0.0% | Yes |
| Bank-style statement | 50 KB | 42.5 KB | 82 dpi | 4.2% | No |
| Bank-style statement | 25 KB | 23.2 KB | 51 dpi | 66.5% | No |
| Receipt | Lossless | 147.7 KB | 300 dpi | 0.0% | Yes |
| Receipt | 200 KB | 147.7 KB | 300 dpi | 0.0% | Yes |
| Receipt | 100 KB | 86.2 KB | 243 dpi | 0.0% | Yes |
| Receipt | 50 KB | 37.5 KB | 146 dpi | 0.0% | Yes |
| Receipt | 25 KB | 20.4 KB | 92 dpi | 0.8% | Yes |
| Small print | Lossless | 346.5 KB | 300 dpi | 0.0% | Yes |
| Small print | 200 KB | 159.6 KB | 171 dpi | 0.0% | Yes |
| Small print | 100 KB | 88.5 KB | 107 dpi | 0.4% | Yes |
| Small print | 50 KB | 39.4 KB | 68 dpi | 48.7% | No |
| Small print | 25 KB | 19 KB | 45 dpi | 74.3% | No |
| Form | Lossless | 151.7 KB | 300 dpi | 0.0% | Yes |
| Form | 200 KB | 151.7 KB | 300 dpi | 0.0% | Yes |
| Form | 100 KB | 78.9 KB | 240 dpi | 0.0% | Yes |
| Form | 50 KB | 37.9 KB | 152 dpi | 0.0% | Yes |
| Form | 25 KB | 20.3 KB | 95 dpi | 0.0% | Yes |
| Invoice | Lossless | 210.7 KB | 300 dpi | 0.0% | Yes |
| Invoice | 200 KB | 192.7 KB | 288 dpi | 0.6% | Yes |
| Invoice | 100 KB | 78.9 KB | 161 dpi | 0.8% | Yes |
| Invoice | 50 KB | 43.2 KB | 101 dpi | 0.4% | Yes |
| Invoice | 25 KB | 21.7 KB | 62 dpi | 30.1% | No |
Error rate is the share of characters the text reader got wrong, against the exact text we drew on the page. At or under 1% we call a page readable: at most one slip in a hundred characters.
What the resolution means in pixels
Our scans are US Letter pages at 300 dpi, which is 2,550 × 3,300 pixels. To reach a size, the engine first lowers the JPEG quality and then scales the picture down. The letter at 100 KB came out at 741 × 959 pixels, which is 87 dpi across the page, and every character was still read. At 50 KB it was 478 × 619, or 56 dpi, and the reader lost 36.9% of the characters.
So 87 dpi, less than a third of what the scanner captured in each direction, was enough for every typed page we tried. A clean, typed page needs surprisingly little, and it goes quickly once it drops below that.
Readability falls off a cliff
The drop is not gradual. The small-print page went from 0.4% wrong at 100 KB to 48.7% wrong at 50 KB. The statement went from 0.0% to 4.2% between the same two sizes. In between there is no setting that is a little bit worse: a page reads, or it starts to turn into guesses.
That is a reason to leave a margin. If a form allows 200 KB and your page reads fine at 100 KB, send it at the larger size. Nobody will thank you for the smaller file, and someone may have to squint at it.
Dense pages need more bytes than sparse ones
The book page and the letter are full of body text. The smallest readable files we reached for them were 83.1 KB to 172.1 KB. The form and the receipt have less text in larger type, and they still read cleanly at 20.3 KB to 20.4 KB. Small print needed the most resolution: it passed at 107 dpi and failed at 68.
The invoice sat between them. It has a table of short lines and figures, and it still read cleanly at 43.2 KB, but at the next step down the reader got 30.1% of its characters wrong. On a page of figures that matters most: a misread word can often be guessed from the words around it, a misread digit cannot.
So before you type a size, look at the page. A certificate with a few lines and a seal can go very small. A dense statement or a page of fine print needs room.
Step by step
- Scan in grayscale unless colour matters. A colour picture has three channels where grayscale has one, and a black-on-white page gains nothing from the other two.
- Scan at 200 or 300 dpi, well above the 87 dpi our pages needed. You can always shrink later; you cannot add back detail the scanner did not capture.
- Compress to the size the form asks for. For a one-page text document, a target of 100 KB or more keeps it readable in our tests.
- Open the result and read the smallest text on the page at 100% zoom. If you can read it, so can the person at the other end.
- If the form's limit is too small for a dense page, check whether it takes several files, and split the scan page by page.
How we measured
The pages come from our own test set: made-up text drawn at 300 dpi in common fonts, each saved as one grayscale JPEG on a Letter page, the way a scanner writes it. Because we drew the text, we know exactly what each page says. We ran the site's own compress code in Node 22.23.4 on November 11, 2026, at each target, then drew each result at 300 dpi and read it with Tesseract, the same text reader and English model our OCR tool ships. A two-column page is left out of the figures, because the reader takes its lines across both columns even before compression.
A text reader is stricter than a person in some ways and more forgiving in others, so treat these as a guide, not a guarantee. The bench script is in our repository.