How to OCR a PDF — turn scans into searchable, structured, copy-able text
OCR (optical character recognition) is the bridge between a PDF that's just an image of pages and a PDF you can search, copy, and analyze. Our AI OCR layer is more than character recognition — it also detects document structure (headings, paragraphs, tables, lists) and preserves it. That's the difference between OCR'd text you can paste into a chat and text that needs an hour of cleanup first.
What AI OCR does
First we render every page to a high-resolution bitmap. Then a multi-stage model recognizes text glyphs across mixed scripts and fonts (Latin, Cyrillic, Greek, CJK, Arabic, Hebrew, Devanagari). On top of that, a layout model identifies regions: what's a heading, what's a column, what's a table, what's a footnote. The recognized text is positioned back into a hidden text layer aligned with the original pixels — so the PDF still looks identical to the scan, but you can now select and search the text.
Optionally we also export the text as a clean .txt or .docx with the structure detected. That's useful when you don't need the original layout, just the content.
When OCR is the right step
OCR is the prerequisite for almost any other operation on a scanned PDF:
- Scanned contracts. Make them searchable, copy-able, and chat-able.
- Photographed documents. Phone photos of receipts, IDs, or whiteboards become real text.
- Old archives. Decades of scanned paper records become a searchable corpus.
- Faxes. Yes, people still receive faxes. Faxed PDFs are usually scans, and OCR makes them workable.
- Image-only exports from older software. Some older 'PDF printers' produced image-only output. OCR adds the missing text layer.
Common mistakes
Patterns that produce poor OCR results:
- OCR'ing low-DPI scans. Below 200 DPI accuracy drops fast. Below 150 DPI it's not worth running. Rescan if you can.
- Skipping deskew on tilted scans. OCR engines work best on orthogonal text. Scans with a 5° tilt lose 5-10% accuracy. We deskew automatically; toggle it off only if you have a reason.
- Running OCR after compression. Strong compression softens edges, which hurts character recognition. OCR first, compress second.
- Not telling the engine the source language. Auto-detection works well, but specifying the language for known cases (e.g. mixed English/Spanish) can boost accuracy by a few percent.
- Trusting OCR'd numbers without verification. 0/O, 1/l, 5/S, and 8/B are the most-confused pairs. Spot-check any extracted figures that matter.
Tips for the best results
Most of the OCR-quality battle is won at scan time, not OCR time:
- Scan at 300 DPI, grayscale. Color scans are larger and don't OCR better. 300 DPI grayscale is the sweet spot for accuracy and file size.
- Scan flat, with the page parallel to the platen. Phone photos at angles are why so much OCR fails. Use a document scanner app that detects edges.
- Avoid photocopier scans of photocopies. Each generation loses contrast. OCR a fresh scan, not a fourth-generation copy.
- Confirm orientation. Pages scanned upside-down or sideways need rotation before OCR. Auto-rotate fixes this most of the time, but check.
- Inspect the confidence overlay. We highlight low-confidence words. A quick visual scan catches the substitution errors automation never will.
Limits and supported scripts
We OCR Latin, Cyrillic, Greek, CJK (Simplified and Traditional Chinese, Japanese, Korean), Arabic, Hebrew, Devanagari, Bengali, Tamil, Thai, and 30+ other scripts. Free tier: up to 20 pages free, then 2 credits per page. Files up to 200 MB. Output: PDF with hidden text layer (looks identical to the scan, but searchable), plus optional plain text or Word export.