How to make a scanned PDF searchable

You have a 40-page scanned contract and you need the one paragraph about termination. Ctrl+F finds nothing. You can see the words on the screen, so it feels like the search is broken.

It is not. There is genuinely no text in that file to find.

Why a scan has no text in it

A scanner is a camera. It records how bright each point on the page was and stores that as an image, and the PDF it produces is a container holding one picture per page. The letters you can read are patterns of dark pixels, exactly as a photograph of a road sign contains no text.

A PDF exported from Word is a different kind of object entirely: it contains actual characters, each with a font and a position. That is why search, select and copy work on one and not the other, and why the file sizes are so different for documents that look identical.

What OCR adds

Optical character recognition looks at the picture, works out which shapes are which letters, and produces text. The interesting question is what to do with that text once you have it, and there are two very different answers.

The wrong answer, which a surprising number of tools give, is to rebuild the document: recompress the page images, lay the recognised text over the top, and hand back a new PDF. The scan comes back softer than it went in, often larger, and any recognition mistake is now baked into a file you cannot compare against the original.

The better answer is to add rather than replace. The PDF format has a text rendering mode that draws characters invisibly — the text is genuinely there, at the right coordinates, and simply is not painted. Positioning the recognised words over the matching words in the image means search finds them, selection highlights the right area, and copy-and-paste returns the words. Not one pixel of the scan changes, because the scan is never touched.

Making one

1. 1Open Searchable PDF and drop your scan in.
2. 2Wait for the first page. The recognition engine loads before it can start, which takes a few seconds; every page after that is much quicker.
3. 3Check the reported confidence when it finishes. It tells you how well the recognition actually went rather than leaving you to find out later.
4. 4Download, and try Ctrl+F on something you know is in the document.

Pages that already contain real text are skipped by default, and the result says how many were. That default matters: running recognition over text that already exists leaves two overlapping copies of every word, so search results duplicate and copying returns everything twice. If a page has a bad text layer you want replaced, there is a switch for it.

What decides the accuracy

Almost entirely the scan, and hardly at all the engine. On clean 300 DPI printed text, expect 95% or better. On a faint third-generation photocopy, expect to be disappointed, and no software will change that — the information is not in the image.

• Resolution. 300 DPI is the standard for a reason. At 150 DPI the engine starts confusing similar letters; below that it degrades quickly.
• Straightness. Even a few degrees of skew hurts, because recognition works along lines of text and a sloping line drifts out of its own row.
• Contrast. Crisp black on white is ideal. Grey text, yellowed paper, and text printed over a background image are all much harder.
• Type rather than handwriting. Handwriting is not realistically recognisable by this class of engine at all — that is a different technology and a much harder problem.
• Ordinary fonts. Decorative, condensed and very small type all cost accuracy.

Words the engine is genuinely unsure about are left out rather than inserted as nonsense. That is deliberate: a missing word is obvious when you search for it, whereas a confidently wrong one silently corrupts the document.

Getting a better scan in the first place

If you control the scanning, twenty seconds of care beats any amount of processing afterwards. Scan at 300 DPI, in black and white or greyscale rather than colour, with the page square against the edge of the glass.

For phone photos, which is how most documents get captured now, the things that matter are flatness and light. Hold the phone parallel to the page rather than at an angle, get even light without your own shadow falling across it, and fill the frame. Scan documents with your phone goes through the whole process, and Scan to PDF will straighten and clean the photos up before you run recognition on them.

What you can do once it is searchable

Rather more than search, which is the part people miss. A real text layer is what every other text tool needs in order to work on the document at all.

• PDF to Text pulls the words out as a plain file.
• PDF to Word produces an editable document, which is the usual two-step route from a scan to something you can rewrite.
• PDF to Markdown gives you something clean to paste into ChatGPT or Claude — and there is a guide to why that beats uploading the PDF.
• Compare PDF can finally tell you what changed between two versions of a scanned contract.

The honest limitations

This is the one tool on the site that needs the network the first time you use it. Your document still never leaves your device — the recognition runs in your browser in a WebAssembly build of Tesseract — but the engine and the English language model are about 7MB and have to be downloaded before anything can happen. They are served from this site rather than a third party, and they are cached afterwards. Every other tool here works with the network switched off, and claiming the same for this one would be untrue.

English only for the moment. Each additional language means shipping another model file, so more will follow based on what people ask for.

Common questions

Will OCR change how my scan looks?

No. The recognised words are drawn over the page in the PDF specification's invisible text mode, so they can be searched, selected and copied while the image underneath is left exactly as it was. Nothing is recompressed and nothing is redrawn.

Is my document uploaded for the recognition?

No. The engine runs inside your browser. What gets downloaded, once, is the engine and language model — about 7MB from this site. Your file stays on your machine throughout.

How accurate is it, really?

On a clean 300 DPI scan of printed text, typically 95% or better, and the result reports the actual average confidence so you are not left guessing. Low resolution, skew, faint type, shadows from phone photos and text over pictures all reduce it. Handwriting is out of reach.

Why is the first page so much slower?

The engine and language model load before the first page can be read. After that each page is much faster, and the model stays loaded if you process another document in the same visit.

Does it make the file much bigger?

Barely. A text layer is characters and coordinates, which is tiny next to page images. Expect a few tens of kilobytes for a long document — nothing like the growth you get from tools that recompress the pages.

Can I search a scan without doing this?

Not in the PDF itself. Some document management systems and cloud storage services run their own recognition behind the scenes and search that index, which works inside those systems and does nothing for the file you email to someone. Adding the text layer travels with the document.