A scanned PDF is a photograph of a document. It looks like text to you and is nothing but pixels to the computer, which is why searching it finds nothing, selecting it selects nothing, and copying it copies nothing. This puts real text behind the picture, so all three start working.
The important thing is what it does not do: it does not redraw your page. The image stays exactly as it was, and the recognised words sit invisibly behind it. That is why a document processed here looks identical afterwards, and why the failure mode is a wrong word in a search rather than a page that suddenly looks different.
Accuracy depends on what you give it. Clean printed text comes out close to perfect. A third-generation photocopy or handwritten notes will have mistakes, and when the page is too unclear to read reliably the tool says so instead of writing guessed words behind your scan — search for a name you are certain is in the document, and you will know within seconds how well it went.
Two things about honesty here. First, if a page already has real text, it is not read — recognising over correct text replaces something certain with a guess, and the panel counts only the pages that are just a picture before anything starts. Second, in a document that is partly typed and partly photocopied — a contract with a scanned signature page, say — only the scanned pages are read. The typed ones are already right.
And it all happens on your device. The page is rendered, read and rebuilt inside your browser; nothing is uploaded, and the language models live here too — that is why only English and Spanish are offered today rather than a long list we could not keep on your machine. Stopping the read leaves the file exactly as it was, byte for byte.