How to Make a Scanned PDF Searchable

You have a PDF. You press Ctrl+F, type a word you can plainly see on the screen, and get nothing. The document is a scan: every page is a photograph, and the words in it are pixels rather than text. Optical character recognition is the step that changes that.

What OCR actually produces

OCR does not rewrite your PDF into a word processor document. It reads the image, works out where the characters are, and writes an invisible text layerpositioned on top of the picture. The page still looks exactly like the scan — same coffee stain, same slight rotation — but now there are real characters sitting behind it. Search finds them. Copy and paste retrieves them. A screen reader can read them aloud.

That two-layer arrangement is why a searchable scan is sometimes called a “sandwich” PDF, and it is the reason OCR is safe to run on documents you care about: the original image is never replaced, only annotated.

How accurate it really is

Nobody gets 100%. What you get depends almost entirely on the input, in roughly this order of importance:

• Resolution. 300 dpi is the practical floor for body text. A 150 dpi scan of 9pt type will produce nonsense no engine can fix, because the information simply is not in the pixels.
• Contrast and straightness. Clean black-on-white, square to the page. A photo of a page taken at an angle under a desk lamp is the hardest case.
• Typeface. Ordinary serif and sans body text is well handled. Script fonts, heavy ligatures and handwriting are not — handwriting in particular is a different problem class and you should not expect usable results.
• Language. The engine needs the right language model. Turkishı and ş, German umlauts and Polish diacritics come out as noise if you run the page through an English-only model.

Expect near-perfect results on a clean 300 dpi office scan, and expect to proofread anything worse. Tables and multi-column layouts are the usual sources of surprise: the characters are right but the reading order is not.

Doing it without uploading the document

This matters more than it sounds. The documents people OCR are exactly the documents they should be careful with — signed contracts, medical letters, bank statements, ID scans, old case files. Every conventional online OCR service works by receiving that file onto a server it controls.

The OCR tool here runs Tesseract compiled to WebAssembly inside your own tab. The page images are rendered locally, recognised locally, and the text layer is written into the PDF locally. Open your browser's network panel while it works: the language model comes down, your document does not go up. The same reasoning behind in-browser CAD conversion applies here, and applies harder, because a scanned document is usually more sensitive than a bracket.

The cost is that recognition happens on your CPU. A twelve-page scan takes a noticeable minute or two on a laptop rather than seconds on a server farm. That is the trade you are making, and for a confidential document it is an easy one.

Once the text is there

A searchable scan unlocks the rest of the workflow. You can now find and replace across the document in the PDF editor, because there is finally text to find. You can redact by selecting words rather than by drawing boxes from memory. And you can compare two versions of a scanned contract, which is impossible while both are just pictures.

A short checklist

• Rescan at 300 dpi rather than fighting a 150 dpi file.
• Pick the correct language — including a combined model for bilingual documents.
• Straighten badly skewed pages before recognising them.
• Proofread anything you will rely on legally. OCR is an aid, not a witness.