0.00 GB / 1.00 GB monthly quota
0.00 GB / 1.00 GB additional quota
0 / 5 daily conversions
/month
Email with pasword reset link sent.
Enter your email address and we'll send you a link to reset your password.
Published on September 3, 2026
You scan a contract, and the scanner hands you a PDF. It looks like a document. It opens like a document. But search it for a word you can plainly see on the page and you get nothing at all.
That's because a scan is a picture of a page, wrapped in a PDF. Every letter you're reading is a pattern of dark pixels, and nothing in the file says what those letters are. To your computer, the page is one flat image — as searchable as a photo of a wall.
OCR — optical character recognition — is what closes that gap. It looks at the picture, works out which shapes are which letters, and writes the answer back into the file.
Here's the part that surprises people: turning a scan into a searchable PDF doesn't change how the page looks. Not by a pixel.
The recognised words are added as an invisible text layer laid exactly over the image, each word sitting on top of the picture of itself. You still see your original scan — the paper texture, the signature, the coffee ring. But Ctrl+F now finds words, you can select a paragraph and copy it, and other software can read the contents.
The page looks identical and behaves completely differently. That is the whole point of a searchable PDF: nothing is redrawn, something is added underneath.
If you'd rather have just the words with no page at all, the same pass can hand you a plain text file instead.
This is the setting that decides whether OCR is useful or useless, and it's the one most people skip.
Character recognition doesn't read letters in isolation — it weighs what it sees against what words are plausible in a given language. Point it at the wrong language and it will still produce text, confidently, and that text will be wrong.
We measured it. A page of Cyrillic run without naming its language — so English was assumed — came back around 40% correct. That's the worst possible outcome: bad enough to be useless, good enough to look like it worked.
Naming the language is worth several points even when automatic detection is close:
On a page of a few hundred words, the difference between 91% and 99% is dozens of mangled words instead of a couple.
Because that makes it worse, and we have the numbers. Every additional candidate language is another way for a short word to be misread as something plausible in a language your document isn't written in.
When we tested a set that included Russian and Serbian but dropped German and French, German fell from 91% to 85% — by adding languages, not removing them. The pairing matters more than the count. So automatic mode runs a deliberately small, tested set, and every other language stays one click away in the picker.
The rule of thumb: if your document is in English, German, French, Russian or Serbian, automatic is fine. Anything else — name it.
Plenty of PDFs are a mix: pages exported from Word, plus a few scanned pages someone photographed and inserted.
Handled carelessly, OCR on such a file is destructive — the naive approach re-renders every page as an image and recognises it afresh, which means the pages that had perfect text now have guessed text instead. Recognition is very good, but it is never as good as the real characters that were already there.
So the pages that already carry text are left exactly alone, and only the image pages are recognised. Nothing you already had gets downgraded.
A page fed slightly askew — three or four degrees, the sort of thing you barely notice — reads correctly. The words come out right.
The damage shows up when you copy. The invisible layer follows the tilted baselines, so selecting a paragraph and pasting it returns fragments in the wrong order — 01, then INVOICE No. 2026-0, then EUR. The text was recognised perfectly and is useless anyway.
01
INVOICE No. 2026-0
EUR
That's why the page is straightened before recognition rather than after. It costs a fraction of a second per page and it's the difference between text you can paste and text you have to retype.
Less than people expect. A twenty-page scan is a matter of seconds rather than minutes, so page count is rarely the thing to worry about — a hundred-page document is a normal request, not a heroic one. Very dense pages and busy periods push that up, but not into a different order of magnitude.
Images are quicker still: a photograph goes straight to recognition without the extra step a PDF needs.
A scanned archive that nobody can search is a filing cabinet with the labels removed — everything is in there, and finding any of it means opening things one at a time. OCR puts the labels back without touching the papers. The one thing worth doing deliberately is telling it what language it's reading; that single choice is the difference between a document you can search and a document full of words that almost say what the page says.