What Online OCR Does to Your Scan (We Measured It)
Most people judge an OCR tool by whether the text came out right. That is the obvious test, and it hides the more important one: what the tool did to your document while it was in there. We ran a real three-page scan through an online OCR service and measured the output. The recognition was genuinely good. The scan itself came back at less than a quarter of its original pixel data.
What we measured, and what we did not
Being precise about scope, because studies that overstate their reach deserve to be ignored: this is one document through one service. A three-page scanned statement, originally captured at 300 dpi, run through one of the widely used free online OCR services. Every number below comes from inspecting the returned file directly — nothing here is estimated, and nothing is inferred from marketing pages.
We are not claiming this is what every OCR service does. We are showing what one widely used service did to one real document, and explaining how to check any tool yourself in about a minute.
The recognition was good
Start with the part that worked, because it matters for fairness. The text layer came back clean: 489 words and 1,145 digits extracted across three pages, with a garble rate of roughly 0.2% — a single mixed-case oddity in the entire document. For a scanned financial statement full of numbers, that is a solid result.
So this is not a story about bad OCR. If your only question is “did it read the words correctly”, the answer was yes.
The scan came back at 152 dpi
The original was scanned at 300 dpi. A letter-size page at 300 dpi is 2,550 by 3,300 pixels.
What came back was 1,275 by 1,650 pixels — exactly half in each direction. Halving both dimensions leaves a quarter of the pixel data. Measured against the page size in the returned file, that works out at 152 dpi.
The original image was not kept and annotated. The page was re-rendered from scratch as a new PNG at roughly 150 dpi, and the scan that went in was discarded. Images account for 98% of the returned file — so nearly the whole document is a re-encoded, lower-resolution copy of what you uploaded.
Why 150 dpi is a decision, not an accident
150 dpi is a reasonable default if your goal is a small file that looks fine on a screen. It is a deliberate trade, and for plenty of documents it is the right one — nobody needs 300 dpi to read a newsletter.
It stops being reasonable when the document is a record. Fine print, faint stamps, signature detail and small figures are exactly the content that a halving in each axis takes away, and they are exactly the content that matters in the documents people OCR most: statements, contracts, payslips, medical letters, legal filings.
The part worth objecting to is not the number. It is that the trade happens silently. Nothing in the process says the file you get back is not the file you put in.
What the alternative looks like
OCR does not require re-rendering the page. The text layer is invisible — it sits underneath the image and is never drawn — so the picture of the page can be left exactly as it was.
We verified this on our own tool by running its actual save path and comparing the image before and after with a cryptographic hash. The image came out byte-for-byte identical: same 2,550 by 3,300 pixels, same 300 dpi, same SHA-256. The text layer was added, and the scan was untouched. On the real three-page document, the file grew by about 4% — which is the text layer, and nothing else.
Both tools, incidentally, use the same underlying recognition approach: the text layer in each is written in an invisible font, the standard technique. The difference is not the OCR engine. It is what happens to your image on the way through.
How to check any OCR tool yourself
You do not need special software. Two checks, about a minute:
- Compare file sizes. Note the size before and after. A file that came back dramatically smaller has almost certainly been re-encoded at lower quality. OCR alone adds a small amount — a few percent — because a text layer is tiny compared with page images.
- Zoom in hard on the smallest print. Open the original and the result side by side and zoom to 400% on the same detail — a footnote, a reference number, the fine print at the bottom. Degradation that is invisible at 100% is obvious at 400%.
- Check the page dimensions if your reader shows them. A returned page that is a different size from the original has been rebuilt rather than annotated.
The size check catches most of it. People instinctively read a smaller file as efficiency. On a scan it usually means resolution was thrown away, and you cannot get it back — the only copy that still has it is the one you uploaded.
When it does not matter
Plenty of the time, honestly, none of this matters. If you are OCRing an article so you can search it, a recipe, a manual, or anything where you only need the words, a 150 dpi re-render is completely fine and the smaller file is a benefit.
It matters when the document is evidence of something. Anything you may need to send to an insurer, a landlord, a lawyer, a bank or a tribunal is worth keeping at full quality, because you cannot predict which detail someone will later need to read closely. The same goes for anything you are digitising to replace the paper — once the paper is gone, the scan is the record, and a quarter of the pixel data is what you are left with permanently.
The practical takeaway
Keep the original of anything that matters, whatever tool you use. A processed file is a derivative, and derivatives lose things quietly.
If you want the text without the trade, OCR here leaves the scan exactly as it arrived and adds the text layer underneath. If you would rather verify that than take our word for it, run the two checks above on a file of your own — that is the point of publishing the method rather than just the conclusion.
Method
The output file was inspected directly: embedded image dimensions and format read from the file structure, effective resolution calculated from image pixels against page size in points, text extracted and counted, and the text layer's render mode and font read from the page's content stream. Our own tool's behaviour was verified by running its save path on a 300 dpi test page and comparing SHA-256 hashes of the embedded image before and after.
One document, one competing service, measured on 25 July 2026. If you run the same test and get a different result, we would rather hear it than not.
Why we published the numbers instead of the conclusion
There is a lot of comparison content on the internet that is really just marketing with a table in it. The way to tell the difference is whether the method is given — whether you could repeat the test and get the same answer, or whether you are being asked to trust a claim.
So the numbers above are all reproducible from the file itself: the pixel dimensions, the page size, the format, the word and digit counts, the hash comparison. Anyone with the same two files can check every one of them, including the parts that are inconvenient for us. The fact that the competing tool's recognition accuracy was good is in here for the same reason — leaving it out would have made the piece more flattering and less true.
If you find a case where our tool does degrade an image, that is worth more to us than a favourable review, because it would be a bug rather than a trade-off. The promise is specific and testable: your page image comes back exactly as it went in.
Try it now
OCR is free, needs no sign-up, and your file is never stored.
FAQ
- Does OCR always reduce scan quality?
- No. OCR adds an invisible text layer and does not require touching the page image at all. Some tools re-render the page while they are at it, which is a separate decision from the recognition itself.
- What resolution should a scan be kept at?
- 300 dpi is the usual standard for documents you may need to read closely later. 150 dpi is fine for reading on screen but loses fine print, faint stamps and signature detail.
- How can I tell if a tool degraded my file?
- Compare file sizes before and after, and zoom to 400% on the smallest print in both. A much smaller returned file usually means the images were re-encoded at lower quality.
- Was the OCR text accuracy poor?
- No — it was good. 489 words and 1,145 digits came out with roughly a 0.2% garble rate. The finding here is about what happened to the image, not the recognition.
- Is this true of every online OCR tool?
- We measured one service on one document and are not claiming more than that. The checks described here let you test any tool yourself in about a minute.
- Does PDF Cubby keep the original quality?
- Yes. We verified it by hashing the embedded image before and after processing — it came out byte-for-byte identical, with the text layer added underneath.