Optical character recognition (OCR) is the process of converting an image of text - a scanned page, a photographed document, a screenshot - into actual, machine-readable characters that software can read, search, copy and quote.
To a computer, a scanned PDF is just a picture. The page might be covered in words, but they exist only as pixels: there is no underlying text to select, search, or feed to a language model. OCR closes that gap. It detects the regions of an image that contain text, recognizes each character, and reconstructs the words, lines and reading order as a text layer behind the image.
Why it matters
Without OCR, a huge amount of written knowledge is effectively invisible to software - old contracts, research scanned from print journals, receipts, faxes, handwritten notes, and any PDF that was photographed rather than exported. You can see the words, but you can't search them, copy them, or ask an AI about them.
OCR makes that content first-class:
- Search. Recognised text can be indexed, so semantic search and keyword search work across scanned material.
- Citing. Once each word has a known position on the page, an answer can point back to the exact passage it came from: the foundation of a real citation.
- Reuse. Text can be selected, translated, summarized, or chatted with like any digital document.
How it works
Modern OCR runs as a pipeline, and each stage affects the quality of the result:
- Pre-processing. The image is straightened (deskewed), denoised and contrast-normalized, so faint or crooked scans don't derail recognition before it starts.
- Layout analysis. The engine finds the regions that actually contain text and works out reading order - the step that decides whether a two-column page comes back as two articles or as interleaved nonsense.
- Recognition. Each detected line is decoded into characters. Early engines matched individual glyph shapes against templates; modern ones use neural networks that read whole lines at once, which copes far better with unusual fonts and slight distortion.
- Post-processing. A language model cleans up likely misreads - "rn" vs "m", "0" vs "O" - using the surrounding words as context.
The output is more than a string of text. Good OCR also records where each word sits on the page - its bounding box - and that positional data is what lets software highlight a recognized word in place, which is what makes precise citations into scanned pages possible at all.
OCR for scanned PDFs - what to expect
Not every PDF is the same underneath. A PDF exported from Word, Google Docs, or a modern scanner app usually has a text layer - an invisible grid of characters positioned exactly under the visible glyphs, which is what lets you select and search text in a normal PDF viewer. A PDF made by photographing a page, or scanned on older office hardware, is often just an image wrapped in a PDF container: no text layer, nothing to select, nothing a search box or an AI can read.
OCR is what turns the second kind into something closer to the first. Running OCR on a scanned PDF doesn't change how the page looks; it adds a text layer behind the image, aligned word-for-word with what's printed on it. The visible page still looks like a scan, but underneath, every word now has a known position, so it can be selected, searched, copied, and matched against a query.
This matters most for exactly the documents people actually need to work with: signed contracts, old research scanned from print journals, government forms, faxed paperwork, and photographed whiteboards or receipts. Without OCR, none of that is usable by search or AI - it's a picture of information, not information.
In practice, here is what OCR on a scanned PDF does and doesn't deliver:
- Clean printed scans come back near-perfect. A straight, 300 DPI scan of typed text is searchable and quotable with very few errors.
- Words survive; formatting doesn't. OCR recovers the text and where it sits, not the document's structure - tables usually come back as loose lines, and multi-column pages depend on the engine getting reading order right.
- Handwriting and stamps are hit-and-miss. Printed text is largely a solved problem; cursive notes, signatures and rubber stamps still trip most engines.
- Born-digital PDFs don't need it. If the PDF already has a text layer - exported from Word or Docs - OCR adds nothing. Good tools detect this and skip straight to the real text.
OCR accuracy
OCR accuracy is usually measured as character accuracy (the percentage of individual characters recognized correctly) or word accuracy (the percentage of whole words recognized correctly, which is stricter - one wrong letter fails the whole word). A few factors drive it more than any engine choice:
- Scan resolution. Below about 150–200 DPI, character strokes start blurring together; 300 DPI or higher is the safe zone for small print.
- Contrast and cleanliness. Faded ink, coffee stains, skewed pages, and photographed-rather-than-scanned pages all push errors up.
- Font and layout. Clean printed fonts in a single column are easiest; handwriting, dense tables, and multi-column layouts are hardest.
- Language and script. Engines trained mostly on English can miss accented characters, non-Latin scripts, and mixed-language pages unless they're explicitly supported.
Clean, high-resolution scans of ordinary printed text routinely exceed 95–99% character accuracy today. Handwriting, low-resolution photos, and heavily degraded originals can drop well below that, which is why a single "OCR accuracy" number without context is a rough guide at best.
Sidenote runs OCR on scanned PDFs automatically, so you can read, chat with and quote documents that started life as images - and because each recognized passage keeps its location, every answer can scroll to the exact source instead of asking you to trust it.
Searchable PDF vs. image-only PDF
| Searchable (text-layer) PDF | Image-only PDF | |
|---|---|---|
| What it is | Text stored as real characters, positioned under or instead of rendered glyphs | A picture of a page (scan or photo) with no character data |
| Select text | Yes: click and drag like any document | No: selecting just draws a box over the image |
| Search / Ctrl+F | Yes: finds any word instantly | No: the "words" are just pixels |
| Cite a passage | Yes: a citation can point to the exact sentence | No, until OCR adds a text layer |
| AI can read it | Yes, directly | Only after OCR |
| Typical source | Exported from Word/Docs, or scanned with OCR turned on | Photographed page, older scanner, fax, or "print to image" |
FAQ
What does OCR stand for?
OCR stands for optical character recognition - the technology that recognizes the characters in an image of text and turns them into real, machine-readable text. The name comes from early systems that optically scanned printed pages and matched each character shape against known letterforms.
What does it mean to OCR a PDF?
"To OCR a PDF" means running optical character recognition over a scanned or image-only PDF so it gains a text layer. The pages look identical afterwards, but the words become selectable, searchable and copyable. You don't usually need a separate step for this - Sidenote's free PDF to text converter runs OCR automatically whenever a page has no text layer.
Is OCR accurate?
For clean, high-resolution scans of printed text, yes: modern engines routinely exceed 95–99% character accuracy. Accuracy falls with low resolution (below roughly 200 DPI), poor contrast, unusual fonts, dense tables and handwriting, so any single "OCR accuracy" figure quoted without context is a rough guide at best.
Do I need OCR if my PDF already has selectable text?
No. If you can select and search text in a normal PDF viewer, the file already has a text layer and OCR would add nothing. OCR is only for image-only documents - scans, photos and faxes - where the words exist purely as pixels.