What is OCR?

The distinction that saves the most time

A PDF does not necessarily need OCR. PDFs come in two kinds: ones containing real text characters, and ones containing a picture of a page. Try to select a word in a PDF viewer — if it highlights, the characters are already there and can be copied exactly, with no model and no error rate. Only a scanned page needs OCR.

People routinely run scanned-document workflows on text PDFs and inherit an error rate they did not need to have.

Four things that still defeat it

• Faded thermal paper. Receipts fade fastest where the ink was densest — often the total. Nothing recovers information that is no longer on the paper.
• A warped baseline. A photo of a curled or angled document scrambles decimal points more than it scrambles letters, because a misplaced dot is still a plausible number.
• Tables without ruling lines. Column boundaries inferred from whitespace break when one cell's content is long.
• Handwriting, still, outside of narrow cases like forms with boxed characters.

The practical consequence: how you capture the image matters more than which OCR you use. Flat, evenly lit, straight on, as soon as possible.

Where it runs, and why that matters

OCR good enough for documents runs as a model, and models run somewhere. Free web converters run it on their servers, which means your document is uploaded. Some applications run a local model. Neither the tools on this site nor TableDI 2 does OCR at all: they read files that already contain text — a spreadsheet, a CSV, a PDF with a text layer — and say so when a file turns out to be a scan.