Does this compare text or images?
Images — each page is rendered to a raster image and compared pixel by pixel. This catches any visual change, including ones that wouldn't show up in a pure text diff, but it can't tell you which words changed.
What happens if the PDFs have different page counts?
Only the pages present in both PDFs (by index) are compared; the tool flags the page-count mismatch so you know some pages weren't checked.
Why does a small formatting tweak show a large changed percentage?
If text reflows even slightly, every pixel from that point down can shift, which shows up as 'changed' even though the content itself barely differs — this is a visual diff, not a semantic one.
Is anything uploaded?
No. Both PDFs are processed entirely in your browser; nothing is sent anywhere until you choose to download a diff image.