Why a scan has so much to give up
A scanner has no idea which parts of the page matter. It samples every square inch at high resolution in full color, and stores the result losslessly, so a page that’s 95% blank paper still costs the same per pixel as a page dense with text.
Re-encoding the page as a JPEG is lossy, and the thing it discards is mostly detail your eye can’t resolve at reading size. That’s why the reduction is dramatic here and negligible on a document the bank generated as text: on a text document there’s nothing redundant to remove, so re-encoding only adds.
What each resolution costs you in readability
The setting you choose is a straight trade between file size and how well small marks survive. This is what each one looks like on a scanned page.
- Faithful: 200 dpi
- Small print, stamps and signatures keep their edges. Use this if the file will be printed or inspected. The default for anything going to a lender or a lawyer.
- Balanced: 150 dpi
- Where most scans land. Reads fine on screen, prints acceptably, and often cuts the file by more than half again versus 200 dpi.
- Strong: 110 dpi
- Screen reading only. Numbers and small type start to blur at this point, so check a figure before you send one.
- Extreme: 80 dpi
- The largest reduction, and the one most likely to make a document unusable. Fine for a signature page, wrong for a table of figures.
- Grayscale
- Worth trying on a document scanned in color that has no meaningful color in it. A color-in-color-out scan of black ink on white paper carries two channels of pure noise.
The ceiling that quietly reduces your setting
Very large pages are scaled down on the way out to stay under the renderer’s pixel limit. In practice this means a 300 dpi setting on a big page may not come out as 300 dpi.
It isn’t a bug and not something you can switch off. The alternative is a canvas the browser refuses to allocate, which fails outright rather than degrading. It’s worth knowing because the figure in the result panel is what actually happened, and it’s the number to trust.
A scan usually has no text layer to lose
The usual argument against re-encoding is that it destroys the text layer, form fields, links and bookmarks. On a scanned page most of that was never there. The page is a picture, so there’s no text layer to destroy.
Form fields are the exception worth checking. A scanned form with fillable fields drawn on top of it is a real and common thing, and re-encoding flattens them into the image. The compressor reports the field count it removed so you can see it happened.
What we can and can’t tell you about your file
There’s no table of real-world results here, and the reason is the same one that makes the site worth using: nothing is uploaded, so there’s no collection of customer scans to measure. A site showing averages from “thousands of real documents” either kept them or invented the number.
What you get instead is your own file’s two numbers, computed on your machine, in front of you. For a document this variable, where scan resolution, color depth and how much of the page is ink all change the result, that’s the only measurement that means anything.