How to make a scanned PDF searchable
You scan forty pages of old contracts, save them as one tidy PDF, and go looking for the word "indemnity". Ctrl+F finds nothing. You can see the word — it's right there on page twelve — but your computer can't, because as far as it's concerned that page contains no words at all. It contains a photograph.
Three kinds of PDF that look identical
Almost every confusion about scans starts here. Three very different files can appear the same on screen:
- A born-digital PDF. Exported from a word processor. It stores actual characters with actual fonts. Search, selection and copy all work because the text is genuinely text.
- An image-only PDF. From a scanner or a phone camera. Each page is one big picture. There is no text in the file — only pixels arranged to look like text. This is what defeats Ctrl+F.
- A searchable PDF. The scan plus an invisible text layer. It looks exactly like the image-only version — same picture, same page — but text sits behind it, so search and copy work.
The third kind is what you want, and OCR is how you get there.
What OCR actually does
Optical Character Recognition looks at the pixels, finds shapes that look like characters, and decides which characters they are. Modern engines don't match letters one at a time like a stencil; they work along lines of text, using the shapes and the surrounding context to resolve ambiguity. That context is why a good engine reads "rn" as two letters in "modern" but might still stumble over it in a garbled scan — it's making an informed guess, every single time.
That word — guess — is the honest core of OCR. It is a probabilistic reading of an image, not a lossless conversion. Usually a very good guess. Never a guaranteed one.
Why a searchable PDF is usually what you want
You could OCR a scan and keep only the text, but you'd throw away the document: the signature, the letterhead, the coffee ring, the layout — everything that makes it the original rather than a transcript. And since OCR is imperfect, a pure-text version has no way to check itself.
A searchable PDF avoids that trade. The scanned image stays exactly as it was, untouched. The recognised text is added as an invisible layer positioned over the words it corresponds to. You see the original; your computer sees the text; when you copy a paragraph you get the words, and when a recognition error creeps in the true page is still right there to compare against.
What actually drives accuracy
OCR quality is decided long before the engine runs — mostly at the moment of scanning:
- Resolution. Around 300 dpi is the sweet spot for printed text. Below about 200, characters lose the detail needed to distinguish them. Above 400 you're mostly making the file bigger, not the reading better.
- Contrast. Crisp black on clean white is ideal. Grey photocopies, coloured paper and shadows across the page all cost you accuracy.
- Straightness. Skew is a real enemy. A page photographed at an angle, or with the curve of a book near the spine, gives the engine bent baselines to follow.
- Print quality. Clean machine print reads well. Faxes, third-generation photocopies, dot-matrix output and stamped text read considerably worse.
- The right language. The engine uses a language model to resolve ambiguity. Running an English model over a French document degrades results even though the alphabet matches.
- Handwriting — don't. This engine is built for machine-printed text. Handwriting recognition is a genuinely different problem and is not a supported use case. If your document is handwritten, OCR will produce confident nonsense.
If a scan comes out badly, the highest-value fix is almost never a different OCR setting. It's rescanning: flatter, brighter, straighter, at 300 dpi. Ten seconds at the scanner beats an hour of correcting output.
Searchable PDF or plain text?
Two different goals, two different tools, and picking the wrong one wastes time:
- You want the document to stay a document — searchable, copyable, archived, still looking like the original. That's Searchable PDF. Best for contracts, records, anything you may need to produce later as it was.
- You want the words out — to paste into an email, feed into a spreadsheet, quote in a report. That's Extract Text, which gives you plain text and discards the layout.
If you're unsure, choose the searchable PDF. You can always pull text out of it later; you can't recover a page you discarded.
How to do it
OCR in CyvoDocOps runs on your own device — the scan is never uploaded to a server. It's part of CyvoDocOps Pro, and because Pro is purchased through the App Store or Google Play rather than on the web, the OCR tools live in the mobile app.
- Get the best scan you can. 300 dpi, flat, good light, straight. This step decides your result more than any other.
- Open the document in the CyvoDocOps app. See the app if you don't have it.
- Choose your tool. Searchable PDF to keep the page and add the hidden layer; Extract Text to pull the words out; OCR English or OCR Arabic for the language you're reading.
- Check the result. Search for a word you know is on page one. If it isn't found, the recognition struggled — which usually points back at the scan.
Being straight about the limits
- OCR is never 100% accurate. Even on clean scans there are errors — a digit misread, a hyphen lost. Good is not perfect, and the gap matters when the number is a payment amount.
- Don't trust a negative search. If Ctrl+F doesn't find "indemnity", that may mean the word isn't there, or that OCR read it as "indernnity". Absence of a hit is not evidence of absence in an OCR'd document — an important distinction if you're searching records for legal or medical purposes.
- Tables and columns are hard. Recognition follows lines of text; complex layouts can interleave columns or scramble table cells even when every individual character is read correctly.
- The text layer can be wrong while the page is right. That's the design working as intended — the image is the record, the text is the index — but don't quote from the layer without looking at the page.
Arabic, and other scripts
Recognition difficulty varies enormously by script. Arabic is markedly harder than English: letters join, their shapes change by position in the word, diacritics carry meaning, and the text runs right to left. A model trained for English is useless on it — you need the Arabic model, via OCR Arabic. If that's your use case, working with Arabic PDFs covers the specifics.
Why on-device matters here
Think about what people actually scan: contracts, medical letters, bank statements, passports, tax records. Scanning is what you do to documents that exist on paper because they're important. Sending that pile to someone else's server for processing is a strange bargain — and a common one, because most OCR services have no other way to work. Running recognition on your own device keeps the document where it started. Nothing is uploaded, so there's nothing to retain, log, or leak.
Frequently asked questions
Why can't I search my scanned PDF?
Because it contains no text. Each page is an image, and search looks for characters. OCR is what turns the picture into words your computer can find.
Does making a PDF searchable change how it looks?
No. The scanned image is preserved exactly. The recognised text is added as an invisible layer behind it.
Will OCR work on handwriting?
No. The engine is designed for machine-printed text. Handwriting is a different problem and isn't a supported use case.
Why is OCR in the app rather than the browser?
It's part of CyvoDocOps Pro, and Pro is purchased through the App Store or Google Play — there's no web checkout — so the tools live in the app. The recognition runs on your device either way.
Is my scan uploaded for OCR?
No. Recognition runs entirely on your device and the document is not sent to a server.