Why Your Scanned PDF Isn't Searchable (And What OCR Can't Fix)

You scan a twenty-page contract, open the PDF, press Ctrl+F to find a clause, and get nothing. The words are visibly there on screen. The search finds none of them.

This confuses people reasonably often, and the explanation is simple once stated: your scanner produced a photograph of each page. To the PDF, each page is a single image — a grid of coloured dots. There is no text in the file, only a picture that a human eye interprets as text. Searching it is like asking your computer to find a word in a holiday snapshot.

Optical character recognition is the process that changes this.

What OCR actually produces

OCR examines the image, identifies shapes it recognises as characters, and produces machine-readable text. In a PDF, that text is then written into the file as an invisible layer positioned precisely over the corresponding words in the image.

The result is a document that still looks exactly like your scan — same appearance, same imperfections — but which now has real text behind it. Search works. Copy and paste works. Screen readers can read it aloud. The visible page is unchanged.

This is why an OCR'd PDF is sometimes called a "searchable image" or "image with hidden text". You are not converting the scan into a Word-style document; you are annotating it with what it says.

Our OCR tool does this in your browser. It is worth being upfront that browser-based OCR is slower than server-based processing for long documents, because the work happens on your own machine — the trade you get in exchange is that the document never leaves it.

Why accuracy varies so much

OCR quality is decided almost entirely before the software runs. The scan is the variable that matters.

Resolution. 300 DPI is the practical minimum for reliable recognition of normal body text. At 150 DPI, character shapes blur into each other and error rates climb steeply. Above 400 DPI you get diminishing returns and much larger files. If you control the scanner, 300 DPI is the answer for text documents.

Straightness. Even a two-degree skew measurably hurts accuracy, because character recognition assumes roughly horizontal baselines. Most OCR software attempts automatic deskewing, and it usually helps, but a page scanned straight beats a page corrected after the fact.

Contrast and background. Crisp black on white is the ideal. Grey photocopies, yellowed paper, coffee stains, highlighter and shadowing from a curved book spine all degrade recognition, sometimes severely.

Typeface. Standard serif and sans-serif text in the 10–12pt range is what these systems are trained on. Decorative fonts, script, very small print and heavily condensed type produce more errors. Handwriting is a fundamentally different and much harder problem, and general-purpose OCR should not be expected to handle it.

Language and character set. Recognition depends on knowing which characters to expect. A document containing accented characters, mathematical symbols or a non-Latin script needs OCR configured for that content, or those characters will be approximated with whatever the engine considers closest.

Errors that survive spellcheck

The failure mode worth understanding is that OCR errors are frequently valid words.

Recognition confuses characters that look alike: rn reads as m, l and 1 and I blur together, 0 and O swap, cl becomes d. So "modern" becomes "modem", "burn" becomes "bum", "1,000" becomes "l,OOO".

None of these will be flagged by a spellchecker, because they are real words. In a scanned invoice, a misread digit in an amount is both plausible and invisible to automated checking.

This is why OCR output should not be treated as authoritative for anything consequential. It is excellent for making a document findable. It is not a substitute for reading the original when the exact figures matter — and for legal, financial or medical documents, the scan remains the record.

The tables problem

Tables deserve their own warning. OCR reads the page as lines of text and has limited understanding of column structure, so a table often emerges as a run-on sequence with the relationship between cells lost. Multi-column layouts have a related problem: text from adjacent columns can interleave into nonsense.

Specialised document-structure tools do better here, but general OCR on a complex table is likely to produce something you would not want to rely on without checking. If you need tabular data from a scan and accuracy matters, extracting manually is often genuinely faster than correcting an automated attempt.

What to do before you scan

If the document is not yet scanned, a few choices decide the outcome:

• Scan at 300 DPI.
• Use greyscale for typical text documents rather than colour. It captures what OCR needs and produces much smaller files. Pure black-and-white is smaller still but can lose faint text.
• Get the page flat and square against the platen. For bound books, press the spine down or use a scanner with an edge-flattening mode.
• Prefer a flatbed over a phone photograph when you have the choice. Phone captures introduce perspective distortion and uneven lighting, both of which hurt recognition. If a phone is all you have, a dedicated scanning app that flattens perspective is considerably better than a plain photo.

When the source isn't a scan at all

Here is a case people miss. If your PDF came from an export — from Word, from a design tool, from a web page — it already contains real text, and you do not need OCR at all. A search that fails on such a file usually indicates a font encoding problem rather than an absence of text.

You can tell the difference in two seconds: try to select a line of text with your cursor. If the selection highlights individual words and letters, the text is there. If you get a rectangular selection over the whole page area, it is an image.

That test also tells you whether OCR has worked after you have run it.

Realistic expectations

On a clean 300 DPI scan of printed text in a standard font, expect accuracy in the high nineties by character — good enough that search reliably finds what you are looking for.

On a photocopied fax of a fax, or a phone photo of a curled page under a desk lamp, expect considerably worse, and expect the errors to cluster in exactly the places where the image is hardest to read.

OCR is a genuinely useful technology that turns an archive of dead images into something you can search. It is not a transcription service, and treating its output as verified text is where people get into trouble. Use it to find the document; read the document to rely on it.