What OCR can and cannot read
Character recognition turns a picture of a page into searchable text. How well it does that is decided almost entirely by the scan rather than by the software.
What does OCR add to a PDF?
A layer of invisible text positioned behind the picture of the page. The page still looks exactly as it did, and the words are now selectable, searchable and readable by a screen reader, because they exist as text underneath the image.
Nothing is replaced. The scan remains the visible document and the recognized text sits behind it, which is why a recognition mistake changes what a search finds without changing what a reader sees.
Which scans does it read well?
Flat, straight, high-contrast pages of ordinary printed type at around 300 DPI. Under those conditions modern recognition is accurate enough that a page of prose comes back with a handful of errors at most, usually in the punctuation.
It reads printed tables and columns reasonably too, though it may return them in an order that reads oddly. Reading order is a guess about layout rather than a fact stored in the image, and a two-column page is where that guess most often goes wrong.
Where does it struggle?
Handwriting, photographs taken at an angle, faint or unevenly lit scans, decorative or very small type, and pages that are mostly stamps and signatures. Below roughly 150 DPI accuracy falls away quickly however clean the page is.
Language matters as well. Recognition is trained per language, and a German page read with an English model loses the umlauts and the long compound words first. Choose the language of the document rather than the language of your interface.
How should I treat the result?
As a searchable index rather than a transcript. It is exactly right for finding which of two hundred scanned invoices mentions a supplier, and it is the wrong thing to paste into a contract without reading it.
Where a document has to be accurate word for word, check the extracted text against the page. That is a few minutes of work and it is the only way to know, because recognition does not know when it is wrong: a confident mistake looks the same as a correct reading.
Where to go from here
OCR a PDF adds the text layer and lets you pick the language first, PDF to text shows you what the recognition actually produced, and exporting the pages as images is how to see what the engine was given to read.
What DPI and quality actually change covers the resolution floor recognition needs, and why editing PDF text is limited explains why a recognized layer still cannot be edited like a document.