Scanned PDF vs text PDF: why you sometimes cannot copy the text

You open a PDF, try to copy a paragraph, and nothing selects. The cursor just drags a box across it, the way it would across a photograph. Then another PDF lets you select, search and copy without complaint. Same file extension on both. So what’s different?

One of them contains text. The other contains pictures of text. Almost every baffling PDF behaviour traces back to that.

What a text layer is

A born-digital PDF, exported from a word processor or a browser or an invoicing system, records each character as a character: which glyph, which font, which position on the page. Your reader draws those characters when it renders the page, which is what lets you select them, search them and copy them out. The characters are genuinely in the file.

A scanned PDF records none of that. A scanner is a camera, and its output is an image. The PDF wrapping that image knows it’s displaying a picture that happens to depict words, but it has no idea which words. There’s nothing to select because, as far as the file format is concerned, there’s no text there at all.

How to check in five seconds

1. Drag across a sentence. Text highlights character by character in a text PDF. In a scan you get a rectangular selection box, like selecting a region of a photo.
2. Press Ctrl+F and search for a word you can see. Zero results for a word that’s plainly on the page means there’s no text layer.
3. Drop it into the PDF to text extractor. It reads whatever text layer exists, and if the document is a scan it’ll say the document looks scanned instead of quietly handing you an empty box.

What OCR does, and what it doesn’t

OCR, optical character recognition, is the bridge between the two kinds of PDF. It looks at the pixels of a scanned page, recognises letter shapes and writes a new text layer with its best guess, usually positioned invisibly behind the image so the page still looks like the original scan while becoming selectable and searchable.

Two properties of OCR are worth having in mind.

It guesses. Extraction from a real text layer is exact, because the characters are read straight out of the file. OCR is pattern recognition, and it makes mistakes: rn read as m, 0 read as O, errors that multiply with poor scan quality, unusual fonts or handwriting. Modern OCR is good. It isn’t exact.

It needs resolution. OCR engines work best on scans at 300 DPI. A heavily compressed scan can be unreadable to an OCR engine even while a human can still make it out, which is one more reason to run OCR before compressing a scan and never after.

Why extraction tools return nothing on scans

A text extractor isn’t doing OCR. It reads the text layer that’s already there, which makes it exact, fast and private, and also means it can only give you what the file contains. Point it at a scan and the honest answer is “there’s no text here”. That’s what the extractor on this site reports, instead of returning a blank page and leaving you to wonder whether the tool is broken.

Rule of thumb: extraction for born-digital PDFs, OCR for scans.

The grey areas

A few documents are hybrids, and they’re worth recognising:

• Scans that were OCRed select and search like text PDFs, but copy out with recognition errors in them. If pasted text has odd character swaps, you’re reading OCR guesses instead of the original characters.
• Text PDFs “flattened” for security, sometimes done to prevent edits, have had their text layer deliberately rasterised. They behave like scans because they now are scans.
• Mixed documents, like a digital contract with a scanned signature page stapled on the end. Extraction gives you everything except the scanned pages.

Why this matters beyond copying

The text layer is also what search indexers, screen readers and automated document systems consume. A scanned invoice is invisible to full-text search. A scanned form can’t be read aloud to a blind user. An application processed by software may fail without anyone noticing, because the pages are pictures.

Any time a document has a life beyond human eyeballs, whether it carries a real text layer is the single most consequential thing about it.