What OCR produces, precisely
The distinction that matters in practice is searchable versus correct. A document can be fully searchable — every word findable — while containing a wrong digit in a contract amount. Searchability is what OCR guarantees; correctness is what you have to verify where the text carries legal or financial weight.
What determines accuracy
Printed text rendered at 300 dpi is comfortably readable by a recognition engine; at 150 dpi, characters start merging.
PDFnoted: Resolution is detected before recognition and reported, so a file that will recognise poorly is flagged rather than processed silently.
A page scanned two degrees off square pushes baseline detection off, and accuracy falls faster than you would expect.
PDFnoted: Pages are deskewed before recognition runs. This is the cheapest large accuracy win available.
A Chinese contract run with English-only settings produces confident nonsense.
PDFnoted: Set one or more languages; mixed Chinese–English documents should have both. This is a configuration choice, not a quality ceiling.
Grey backgrounds, show-through from the reverse side, stamps and signatures sitting on top of the text.
PDFnoted: Denoising removes background speckle. Where a stamp covers text, the mark has to be handled before recognition — a stamp read as characters is the most common source of garbage in an otherwise clean result.
How PDFnoted runs OCR
- Inspect the pageResolution, skew and whether a text layer already exists are checked first — no recognition is run on a page that is already searchable.
- Deskew and denoiseRotation is corrected and background speckle removed. Cheap, and the largest single accuracy gain.
- RecogniseCharacters are recognised in the languages you set and placed at the coordinates of the original glyphs.
- Write a layered PDFThe page image is untouched; the text layer sits invisibly on top. The document still looks exactly like the scan.
Where OCR output needs a human
✓ Worth doing
- Check every amount against the original page
- Check reference numbers, dates and long identifiers
- Compare one complex table cell by cell
- Scan at 300 dpi where you have the choice
- Set both languages on a mixed-language document
✗ Not worth the risk
- Assume a searchable document is an accurate one
- Rely on OCR for handwritten annotations
- Feed a 150 dpi fax into a legal workflow without checking it
- Drop the page image after recognition
Compared with the usual options
| Desktop OCR suite | Typical online OCR | PDFnoted | |
|---|---|---|---|
| Where the file goes | Local | Uploaded to a server | Your browser for basic work |
| Layered output (image kept) | Yes | Sometimes | Yes |
| Automatic deskew | Yes | Sometimes | Yes |
| Mixed Chinese–English | Manual setup | Rarely handled | Multi-language setting |
| Flags low-resolution pages | No | No | Yes |
| Text stays positioned on the page | Yes | Sometimes | Yes |
Frequently asked questions
What is a searchable PDF?
A PDF with two layers: the original page image, and an invisible text layer placed over it at the coordinates where the characters were recognised. It looks identical to the scan, but you can search it, copy from it and select text. This layered form is the right output for legal and archival use, because the image of the original page is preserved underneath.
How accurate is OCR?
On a clean 300 dpi scan of printed text, character accuracy is high enough that you will not notice errors in prose. On a 150 dpi fax, a skewed phone photo, or a dense table, it drops. Digits and single-character symbols are where errors concentrate — which is exactly why amounts and reference numbers need checking even when the sentences read perfectly.
Why is my scan recognised badly?
Four usual causes, in order: resolution below about 200 dpi; skew — even two degrees hurts; the wrong language setting; and noise, such as a grey background, show-through from the back of the page, or a stamp overlapping the text. PDFnoted deskews and denoises before recognition, which fixes the first two automatically.
Can OCR read handwriting?
Not reliably, and you should assume not. Handwriting recognition is a different problem from printed-text OCR and is materially less accurate on cursive. Machine-printed text, forms and typed documents are what OCR is for. Handwritten annotations are usually better handled by review than by automation.
Does OCR change the look of my document?
No. The page image is untouched; the recognised text is added as an invisible layer on top at the right coordinates. Nothing is re-rendered, so the document still looks exactly like the scan — including stamps, signatures and handwriting.
Can I OCR a PDF that already has text?
You can, but you would be adding a second, worse text layer over a good one. PDFnoted checks for an existing text layer and tells you if the page is already searchable, rather than running recognition you do not need.
Which languages are supported?
Recognition covers English and Chinese among others, and mixed-language documents can be run with multiple languages set. For mixed Chinese–English contracts, set both — otherwise one of the two will be recognised poorly.
Is there a page limit?
No quota on our side. Recognition is the slower operation in document processing — roughly a second or two per page — so a 300-page bundle takes a few minutes rather than being instantaneous.
Try it on your own file
Upload a document and see what comes back. Free to start, nothing to install.
OCR a PDF now — free to startKeep going
Sources and further reading: ISO 32000 (PDF specification); Adobe PDF Reference; Schema.org structured data; Aggarwal et al., GEO: Generative Engine Optimization (arXiv:2311.09735). This page is a technical guide, not legal advice.