OCR Explained: Turning Scanned PDFs Into Searchable, Editable Text

How to tell whether a PDF needs OCR

Open the PDF and try to select a single word with your cursor. If the word highlights on its own, the PDF already contains a text layer and you can simply extract the text. If instead the whole page highlights as one block, or nothing highlights at all, the page is an image and needs OCR. Another quick test is to press Ctrl+F (Cmd+F on a Mac) and search for a word you can see on the page; if it is not found, there is no text layer.

What happens inside an OCR engine

Modern OCR runs in several stages. First the image is cleaned up: it is converted to black and white, straightened (deskewed) and stripped of speckles. Next the engine analyses the layout, finding columns, paragraphs, lines and table cells so that text is read in the correct order. Then each line is recognised. Older engines matched individual character shapes against templates; current engines use neural networks that read a whole line at a time and use context from neighbouring letters, which is why they cope far better with unusual fonts and slightly damaged text.

Finally, a language model checks the result against a dictionary and grammar, correcting likely mistakes such as reading 'rn' as 'm' or '0' as 'O'. This last step is also why OCR accuracy drops for names, product codes and account numbers: there is no dictionary to fall back on.

Ten ways to get more accurate results

1. Scan at 300 dpi. Lower resolutions lose the fine detail that separates similar letters; much higher resolutions just make files bigger.
2. Photograph pages flat and square-on, in even daylight, with no shadow from your phone across the page.
3. Fill the frame with the page and crop away the table or desk around it.
4. Straighten the page. Even a few degrees of tilt reduces accuracy on long lines.
5. Avoid heavy compression. JPEG artefacts around letters look like ink to an OCR engine.
6. Tell the tool which language the document is in, especially for accented or non-Latin scripts.
7. Process one document type at a time; a batch of identical invoices is recognised more consistently than a mixed pile.
8. For faxes, request 'fine' resolution from the sender when you can — standard-mode faxes are roughly half the vertical resolution.
9. Remove highlighter and coloured stamps where possible; they can hide the characters underneath.
10. Always proofread numbers, dates, names and totals. These are exactly the places where OCR has no language context to help it.

Handwriting, tables and forms

Printed text in a clear font is the easy case, and good engines routinely exceed 99% character accuracy on it. Handwriting is much harder: neat block capitals often recognise well, joined-up handwriting much less so. Tables are a layout problem rather than a recognition problem — the words may be read correctly but land in the wrong cells. If your goal is to get table data into a spreadsheet, use a data-extraction tool that understands rows and columns rather than plain OCR, and check column alignment before relying on totals.

Searchable PDF or plain text?

OCR output comes in two forms. A searchable PDF keeps the original page image and places an invisible text layer behind it, so the document looks identical but can be searched and copied. This is the best choice for archives and records. Plain text, or a Word document, discards the image and gives you only the words, which is better when you need to edit or reuse the content. If you are unsure, keep the searchable PDF as your record and export text from it when needed.

Privacy when using online OCR

Scanned documents are often the sensitive ones — contracts, medical letters, identity documents, bank statements. Before uploading, check that the service uses an encrypted connection, deletes files promptly, and does not reuse your documents for training. On Free PDF FAX, files are deleted automatically within an hour and are not used to train third-party models.