You press Ctrl+F, type a word you can see on the screen, and get no results. The document is a scan: to your computer it is not text at all, but a photograph of text. There is nothing to search because there are no characters in the file, only pixels arranged in the shape of characters.
Optical Character Recognition is the process that fixes this — and understanding what it actually produces explains both why it is so useful and where it quietly fails.
What OCR produces
OCR examines the image, recognises shapes as characters, and writes those characters into the file as an invisible text layer positioned exactly behind the visible image.
That layering is the clever part. The page still displays as the original scan — same appearance, same imperfections, same coffee ring in the corner — but sitting underneath is real, selectable text. Search matches it. Your cursor selects it. Screen readers read it. Copy and paste extracts it. Nothing looks different, and everything behaves differently.
This is why an OCR'd document is strictly better than the scan it came from, and why running OCR at the moment you scan is worth the extra few seconds.
When you need it
- Anything from a scanner or a phone camera. All of it starts as image-only.
- Faxes and photocopied documents, which are scans with extra degradation.
- Documents flattened to images, sometimes done deliberately for redaction.
- Before archiving anything. An archive you cannot search is a filing cabinet. See PDF document management tips.
- Before converting to Word. Conversion needs text to work with; there is nothing to convert in an image. See how to convert a PDF to Word.
- Whenever accessibility matters. A scanned document is completely silent to a screen reader.
To check whether a document needs it: try to select a line of text. If your cursor draws a rectangle over a picture instead of highlighting words, it needs OCR.
Tools
What you may already have. macOS Preview performs live text recognition on scanned documents in recent versions, and iOS and Android both OCR photographed documents through their built-in scanning and camera-text features. Windows users have OneNote, which extracts text from inserted images. Adobe Acrobat's free Reader does not OCR, but Acrobat Pro does it very well.
Free and local. Tesseract is the open-source engine most other tools are built on, and OCRmyPDF wraps it in a single command that adds a text layer to an existing PDF without touching the image:
ocrmypdf input.pdf output.pdf
This is the best free option for anything sensitive or bulk, because nothing is uploaded and it can be scripted across a folder. NAPS2 is a friendly Windows front-end for the same engine.
Cloud. Google Drive OCRs automatically when you open a PDF with Google Docs. Adobe, Smallpdf, and iLovePDF all offer it. Accuracy from the major cloud providers is the best available, particularly on poor-quality scans — at the cost of uploading your document.
A note on our scope: Docento.app works on documents in your browser and does not perform OCR. Recognition is compute-heavy and needs trained models, which is genuinely impractical client-side. Once a document has been OCR'd elsewhere, it is a good place to sign or annotate it.
Getting good accuracy
OCR quality is decided almost entirely by the input. A clean scan gives near-perfect results; a bad one cannot be rescued by better software.
- Scan at 300 DPI. This is the sweet spot. Below 200, accuracy falls off sharply. Above 400 you gain almost nothing and slow everything down.
- Get the page straight. Skew is the biggest single cause of errors. Most tools deskew automatically, but starting straight is better.
- Light it evenly. Shadows across a phone capture do more damage than low resolution. See scanning documents with your phone.
- Use greyscale or black-and-white for text documents. Colour adds size without helping recognition.
- Set the language. OCR uses a language model to resolve ambiguous shapes, so telling it the document is German rather than English measurably improves accuracy. For multilingual documents, most tools accept several languages at once.
- OCR before compressing. Compression discards image detail, and detail is what recognition depends on.
Where OCR still falls down
Modern OCR is very good and it is not perfect. The failure modes are worth knowing because they are quiet:
Handwriting remains hard. Neat print is sometimes recognised; cursive generally is not. Handwriting recognition is a genuinely different and harder problem than printed-text OCR.
Tables lose their structure. The text is usually captured, but the relationship between a cell and its column header is not, so a recognised table often becomes a stream of numbers in reading order.
Multi-column layouts may be read straight across both columns, interleaving two texts into nonsense. Better tools detect columns; not all do.
Unusual and decorative typefaces degrade accuracy, as do very small type, dense footnotes, and text over images.
The errors are plausible. This is the one that catches people. OCR does not fail loudly — it substitutes rn for m, 1 for l, 0 for O, 5 for S. These pass spellcheck and read normally. In prose, a reader glides over them. In numbers, they are invisible and consequential: a misread digit in an invoice total, an account number, or a dosage is not a typo you will notice.
So: proofread OCR output before relying on it, and check figures specifically. If the document is one where a wrong number matters, verify the numbers against the image by eye. The searchable layer is a convenience; the scan remains the authoritative record.
After OCR
Confirm it worked by searching for a word you can see on the page. Then check the file size — OCR adds a text layer without replacing the image, so the file grows a little; if it shrank dramatically, the tool may have replaced your scan with a re-rendered version, which is worth knowing about.
Keep the original scan for anything that matters. And once the document is searchable, it is finally a real document: findable, readable aloud, quotable, and worth archiving properly.
