PDF to Text: What OCR Reads Correctly (and What It Misses)

Optical character recognition looks at an image of a page, finds text regions, and maps each region to characters. A good model also classifies blocks: title, paragraph, list, table, caption. That is why the output can come back as markdown instead of one long line.

A PDF is not automatically text. Some PDFs carry a real text layer, while others are just photographs wrapped in a PDF container. The second kind behaves exactly like a scan, because to the model it is one.

If you exported the PDF from Word, Google Docs or LaTeX, the text layer is already there and extraction is close to exact. The main losses are unusual fonts, ligatures and anything rendered as an image inside the page.

A scan or a phone photo has no text layer, so every character is a guess based on shape. Straight, evenly lit pages read well. A page photographed at an angle, with a shadow across the middle, loses accuracy first in the small print: footnotes, table cells and legal text.

Tables are the hardest common case. A model that understands table structure returns rows and columns as markdown; a plain text extractor returns a pile of numbers with no alignment. Multi-column layouts have the same problem: a naive extractor reads across the gutter and interleaves two unrelated sentences.

Check the first table and the first two-column page. If those look right, the rest of the document usually follows the same pattern. If they do not, the layout is fighting the model and manual cleanup is faster than fixing the output.

Printed text is a solved problem for OCR. Handwriting is not: neat block letters often read correctly, cursive is unreliable, and signatures come back as noise. Do not use OCR as the only copy of a handwritten record.

For bad scans, the cheapest fix is a better scan. Shoot straight down, fill the frame with the page, avoid shadows from your own hand, and keep the resolution high enough that the smallest text is legible on screen.

Modern OCR handles more than 40 languages, including right-to-left scripts and languages with joined letterforms. Accuracy depends on the script and on the model having seen enough of that language in training.

Mixed-language pages work when each page is mostly one language. A page that switches languages mid-sentence can confuse the language guess, so verify names and technical terms by hand.

Extraction is step one. The output is most useful as markdown: headings become structure, lists stay lists, and tables stay tables you can paste into a doc or a spreadsheet.

For contracts and reports, read the headings and the numbers first. Those are the parts where an OCR error changes meaning rather than just spelling, and they are quick to verify against the original.