Comparison runs page by page: the text of both documents is matched up and the result comes back as a separate PDF report highlighting what changed, what was added, and what was removed. The report is the thing to read – it shows differences in the context of the page rather than as a list, so you can see immediately which clause acquired new wording.
The method has limits. It compares the text layer, so two un-recognised scans have nothing to compare – run OCR on them first. If the document was heavily reflowed between versions (different font, margins, or pagination), the report gets noisy: formally almost everything changed even though the meaning did not. It works best where edits are surgical – which is exactly the case with contracts and policies.
Two practical details. Mind the file order: load the older revision first and the newer one second, otherwise 'added' and 'removed' swap places. And compare documents of common origin – a version reassembled by a different program differs from the original more than you would expect.