The reason to compare two PDFs is almost never curiosity. A contract came back from the other side and you need to know what they touched. A report went through three people and you want to see the third round’s changes without rereading all of it. In both cases the question is narrow — what is different — and the risk is a comparison that quietly answers a different question.
The commonest way to get this wrong is alignment. If one version has a line the other does not, and both columns are simply listed top to bottom, then everything below that point is offset by one row. Read across and you are told the third line of the old version became the third line of the new one, when they are unrelated. The comparison looks thorough and is actively misleading.
So here, both sides keep one row per pair, and where a line exists on only one side the other shows an empty gutter. Reading across is always comparing the same place in the document. Within a changed line, the words that differ are marked, so a corrected figure and a rewritten clause do not look alike.
Three cases get an explicit answer rather than a number. Two files that read identically but were saved at different moments are reported as identical in every way a reader would notice, with the byte difference explained rather than counted. Two scanned documents get a specific refusal: with no text in either file, the honest answer is that the pictures differ and nothing more, which is less useful than a diff and more useful than a wrong one. And a large layout shift is reported as movement, with a warning that a small edit inside a moved paragraph is the easiest thing in a comparison to miss.
Nothing is uploaded to compare, both files are left exactly as they arrived, and there is no limit on how long either document is.