OCR reads the text in a scanned document and adds it back to the PDF as a real, invisible text layer sitting behind the page image. The page looks exactly as it did; the difference is that it becomes searchable, its text can be selected and copied, and a screen reader can read it aloud. That last point makes OCR an accessibility measure as much as a convenience — an un-OCR'd scan is completely opaque to anyone using assistive technology. This runs on our server, because the OCR engine and its language models are far too large to ship to a browser: the file streams through the pipeline and the searchable PDF streams back, with nothing stored, though the file does leave your device. Accuracy tracks scan quality closely, so a clean 300 dpi scan is worth far more than any setting.
What OCR adds and what it does not change
The page image stays exactly as it was. The engine recognises the characters, works out where each one sits, and writes an invisible text layer positioned to match. Visually nothing has changed; structurally the document now contains its own text.
That is what makes search, copy and screen readers work. It is also why an OCR'd scan is still a scan: the file size stays roughly the same, the image is unimproved, and any smudge or skew in the original is still there behind the text you can now select.
Scan quality decides accuracy
300 dpi is the working standard for text; below about 200 dpi, small type starts to be guessed rather than read. Straight pages matter as much as resolution — skew is one of the most reliable ways to degrade recognition — and even lighting matters more than sharpness.
Clean printed text in a common typeface reaches very high accuracy. Handwriting is largely out of reach for this kind of engine. Unusual fonts, heavy background patterns, faint carbon copies and text over images are the other predictable trouble, and no amount of processing recovers what the scan did not capture.
Expect to check, not to trust
Even at high accuracy, a long document contains errors, and the characteristic ones are quiet: a 0 read as an O, a 1 as an l, a comma as a full stop. For prose that is harmless; in a reference number or an amount it is not.
So treat the text layer as a search index rather than a transcript. Search it, copy from it, let a screen reader use it — but verify anything numeric against the image, which is still right there behind it.