Why scans are different
A PDF exported from Word or any other program contains its text as characters with positions. A scanned PDF contains a picture of a page. Nothing can compare the words on two scans until something has read them, and that is what optical character recognition does: it reads the image and adds an invisible text layer over it, so the page looks the same but the words can be selected, searched and compared.
A quick test: open the PDF and try to select a sentence with your cursor, or search for a word you can see on the page. If you can, there is a text layer. If the whole page selects as one block, or search finds nothing, it is an image only.
What works
If your scanner, copier or document management system applies OCR when it scans, the PDF has a text layer and the comparison works on it. Many office copiers do this when set to "searchable PDF". Most phone scanner apps do it too, though some only when you turn it on.
The engine detects when text came from OCR and treats formatting differences with suspicion, because OCR guesses bold and italic from the shape of the letters and gets it wrong often. Those differences are ranked low so they do not crowd the list. Substantive text changes are still reported normally.
What does not work
- A scan with no text layer at all. There is no text to compare, so the file is refused with a message saying so, rather than producing a comparison of nothing.
- Very poor scans, where OCR misreads characters. Those misreads show up as changes that are not really there. Rescan at 300 dpi or better.
- Handwritten additions. OCR does not read handwriting reliably, so a handwritten change to a printed contract will not appear in the change list. Check signed copies visually as well.
What OCR noise looks like
OCR errors have a pattern, and once you know it you can read past them quickly. Common ones: the letter l and the number 1 swapped, a capital O read as a zero, "rn" read as "m", a comma read as a full stop, and words split or joined where the scan was faint.
Say the original says "Clause 11.2" and the scan of the signed copy comes back as "Clause l1.2". That shows up in the change list, and it is obviously noise. Now say the original says "payable within 30 days" and the scan says "payable within 60 days". A 3 and a 6 can be confused by OCR on a bad scan, but they can also be a real edit. That is the kind of entry you check against the paper or the image, not dismiss.
The rule of thumb: dismiss noise that changes a character into something that is not a real word or number in context. Investigate anything that produces a different, plausible value.
Getting a better scan
Most bad comparisons of scanned documents come from bad scans, and the fixes are simple.
- Scan at 300 dpi. Lower resolutions save space but make OCR guess.
- Scan in black and white or greyscale for text documents. Colour adds file size without helping recognition.
- Put pages on the glass straight. Skewed pages produce more errors at the edges.
- Set the OCR language to the language of the document, if your scanner offers the option.
- Keep files under 20 MB each. A long document scanned at a high resolution in colour can exceed that easily.
Recommended workflow for signed copies
The most common reason to compare a scan is to check a signed copy against the version that was sent for signature. Compare the OCR-layered scan of the signed copy against the original digital PDF, not against another scan, so only one side carries OCR noise.
Expect a small amount of noise. Anything substantive in the change list is either a real edit or a misread, and you can tell which by looking at the same spot on the scanned page. Then look over the pages for handwriting, since that will not be in the list at all.
If you check signed contracts regularly, the guide to comparing contract versions in PDF covers the rest of the workflow, and HR teams can see how returned employment contracts are checked.
Questions people ask
Can the tool run OCR for me? Not on the web tool. Run OCR in your scanner software, a desktop PDF editor or a scanner app first, then compare.
Why did my scan get refused when I can read it fine? Because you are reading a picture. The tool needs a text layer, and the file does not have one.
Will formatting changes be reported for scans? They are detected but ranked low, because OCR is unreliable at telling bold from regular text.
Try it on your own documents. Upload two versions on the homepage. Documents up to 30 pages need no account, and nothing is stored.