How to OCR a Scanned PDF Into Searchable, Selectable Text
A scanned PDF is a photo of a document - you cannot search it, copy from it, or have software read it. OCR is how you turn those pixels back into words.
Scan a paper document and you get a PDF that looks like text but behaves like a picture. Try to search it and nothing is found. Try to copy a line and you get nothing. That is because the file holds an image of the page, not the letters. OCR - optical character recognition - is the process that reads those images and reconstructs the actual text.
Getting good OCR is partly the tool and partly the input. A crisp, straight scan recognizes almost perfectly; a faint, skewed photo of a crumpled page produces garbage. Most OCR complaints trace back to bad source images, which is good news because you can control that.
What OCR actually produces
Good OCR does not replace your scan with plain text. It adds an invisible text layer beneath the image, so the page still looks exactly like the original but is now searchable and selectable. This is called a searchable PDF, and it is what you almost always want - the visual fidelity of the scan plus the utility of real text.
The alternative, extracting text only, throws away the layout and is useful mainly when you need the words in a spreadsheet or database rather than a readable document.
Get accurate recognition
Accuracy is mostly decided before OCR runs. A few habits raise the hit rate substantially.
- Scan at 300 DPI. Lower and the letters lack detail; much higher wastes space without helping recognition.
- Keep the page straight - deskew crooked scans, because tilted text confuses the recognizer.
- Prefer high contrast: dark text on a clean white background. Faint photocopies and colored backgrounds hurt accuracy.
- Tell the tool the right language, and turn on any dictionary or spell-check pass so common words self-correct.
- For forms and tables, use OCR that preserves structure rather than dumping everything into one text stream.
Check and fix the results
OCR is never perfectly certain. After running it, search the document for a few words you know are in it - if they are found, the text layer took. Then spot-check the recognition-prone characters: the digit 0 versus the letter O, 1 versus l versus I, and rn misread as m. These are the classic errors.
For a handful of documents, correcting a few mistakes by hand is quick. For high-volume, high-stakes text (legal discovery, financial records), plan a verification step rather than trusting raw OCR output, because a single flipped digit in an amount matters.
Make OCR part of intake, not an afterthought
The best time to OCR a document is the moment it enters your system, so everything filed is searchable from day one. Teams that OCR at intake can later find any contract clause or invoice line by searching; teams that pile up raw scans build a folder no one can search.
Atlas can OCR scanned PDFs so uploaded documents become searchable and their text is available to the built-in AI assistant - meaning you can not only find a scanned contract but ask questions about it. Wherever you work, the principle is the same: recognize the text at intake and your archive stays useful instead of turning into a pile of images.