Inside every PDF there is a cross-reference table (xref) – a map of which object sits at which byte offset. A reader parses that first and the pages second. If a download was cut short, an application crashed while saving, or a generator broke the spec, the offsets stop lining up and the viewer refuses the whole document – even though the pages, fonts, and images are physically present.
Repair sidesteps that table: the file is scanned for the objects themselves, the structure is rebuilt from what is found, and the document is reassembled into a valid PDF. That fixes a corrupt or missing xref, truncation in the bookkeeping section, and the non-standard output of exotic generators – the classic 'the file is right there but will not open' cases.
What repair cannot do is invent data that is not in the file. If a transfer died halfway and half the bytes never arrived, only the surviving pages come back. Comparing the file size against what you expected is a quick sanity check. Repair also does not guess the password of an encrypted document: for that you need the password itself and the unlock tool.