A scanned PDF is really just a photo of a page. You can't search it or copy a sentence out of it. OCR reads the text in each page image and embeds it as an invisible, selectable text layer.
The visible page stays exactly the same. Only an invisible text layer gets added underneath, which is what lets you search, select, and copy text from something that used to just be a picture.
You've got a scanned contract or old report and need to search it for a specific clause or figure. Run OCR once, and Ctrl+F starts working on it like any normal document.
The language dropdown matters more than it might seem — recognition accuracy drops sharply if it's set wrong, since each language model is trained on that script's specific letterforms and common letter combinations. If a scanned page mixes two languages (an English form with a Hindi stamp, for instance), pick whichever one makes up the bulk of the text you actually need to search for; OCR will still pass over the other script's characters without errors, it just won't recognize them as meaningful text.
No, only a hidden, selectable text layer is added on top of the existing page images.
It's built for printed text and is generally unreliable on handwriting.
No, recognition runs entirely on-device using Tesseract.js, an OCR engine that runs right in the browser.
The OCR engine downloads its language model once, the first time you use it. Later pages process faster.
Recognition accuracy drops noticeably, since each language model is tuned to that script. Re-run OCR with the correct language selected for better results.
It'll still try, but accuracy depends heavily on scan clarity. Sharper, higher-contrast scans consistently produce more accurate, usable text.