What OCR does, and when you need it
A scanned PDF looks like a document and behaves like a photograph. You cannot select a sentence, search for a name, or copy a figure out of it, because there are no letters in the file — only pixels arranged to look like letters. Optical character recognition reads those pixels and works out which characters they represent, then writes that text back into the file as an invisible layer sitting exactly over the picture. The page looks identical; it is simply searchable now.
This tool runs the recognition in your browser. The document is never uploaded, which is the difference that matters here: the files people most often need to OCR are contracts, bank statements, medical results and identity documents, and every other free OCR service asks you to send those to a server.
Is your PDF actually scanned?
The one-second test: open it and try to select a sentence with the cursor. If letters highlight, the file already has a text layer and you do not need OCR at all — use PDF to TXT, which is instant and perfectly accurate. If the cursor draws a rectangle over the whole page, the page is an image and OCR is the only route.
| What you have | Selecting text | What to use |
|---|---|---|
| Exported from Word, a browser or an accounting system | Works | PDF to TXT — exact, no guessing |
| Scanned on a copier or a phone app | Draws a box | OCR PDF |
| Scanned, then OCR'd by someone else | Works, with occasional odd characters | Nothing — it is already done |
| A photo of a page | N/A | OCR PDF accepts images directly |
| Mixed: some pages scanned, some not | Works on some pages | OCR PDF handles the whole file |
What decides accuracy
On clean printed text from a flatbed scanner, expect somewhere around 95–99% of characters correct. On a crooked phone photo of a creased receipt, far less. The things that move that number, in order of how much they matter:
- Resolution. Recognition falls off sharply below about 200 dpi. This tool renders PDF pages at roughly 220 dpi for that reason. A scan made at 150 dpi cannot be rescued by rendering it larger — the detail was never captured.
- The right language. Picking the wrong one is the single most common cause of nonsense output. A language model shapes every guess, so English settings on Thai text produce gibberish rather than a partial result.
- Straightness. A page rotated by more than two or three degrees loses a lot of accuracy. Rotate PDF fixes whole-page rotation before you start.
- Contrast. Black text on white paper is ideal. Grey photocopies, coloured backgrounds, highlighter and watermarks all interfere.
- The typeface. Ordinary serif and sans-serif print is what the engine was trained on. Handwriting is not recognised in any useful way, and decorative or condensed faces do poorly.
The first run downloads about 15 MB
The recognition engine is a WebAssembly build of Tesseract, roughly 4.7 MB, plus a trained data file for the language you pick, around 10 MB for English. They download the first time you use the tool and the browser keeps them afterwards, so a second document starts immediately. Each additional language is its own download.
Two consequences worth knowing. This is the one tool on the site that genuinely needs an internet connection the first time, because those files have to arrive. And it is slow compared with a server: expect a second or two per page on a laptop, longer on a phone. What you get in exchange is that the document never leaves your device.
Two things come out
- A searchable PDF. The original page image with the recognised text placed invisibly on top. It looks exactly like the scan, but you can search it, copy from it, and any system that indexes documents can now read it.
- A plain text file. Just the words, for feeding into a spreadsheet, a database or a script.
There is also a text-only option, which leaves out the page images entirely. The result is a fraction of the size and useful when you want the content rather than the appearance — but it no longer looks like the original document, so it is the wrong choice for anything you have to send on.
Getting a better result
- Scan at 300 dpi in the first place. It is the standard OCR target and costs nothing at scan time.
- Scan in greyscale, not colour and not pure black-and-white. Colour adds noise; 1-bit black-and-white destroys the anti-aliasing the engine uses to judge letter shapes.
- Straighten and crop first with Rotate PDF and Crop PDF. Margins full of scanner noise slow recognition and add stray characters.
- Split very long documents. This tool caps a run at 50 pages, because a browser tab holding several hundred rendered pages will run out of memory.
- Proofread anything that matters. OCR reliably confuses 0 with O, 1 with l and I, and rn with m. For a figure in a contract, check it by eye.