Folia Tools
What OCR can and cannot read, and what a searchable PDF is actually made of
OCR is the one operation on this site that guesses. Understanding what it is guessing from tells you when to trust the result and what to fix before running it.
A searchable PDF is a picture with words written invisibly over it
This is the part that surprises people. OCR does not convert your scan into text. It leaves the scan exactly where it is and adds an invisible layer of text on top, positioned so that each recognised word sits over the picture of that word.
You still see the original scan — the coffee stain, the staple mark, the handwriting in the margin. But a search finds words, a text selection returns characters, and a screen reader can read the page aloud, because all of those operate on the invisible layer. It is a genuinely clever arrangement: you keep the evidential picture and gain the searchable text, and neither compromises the other.
OCR PDF builds this by rendering each page to a JPEG about 1100 pixels wide at quality 0.8, recognising it, and rebuilding it as that image with the recognised words written invisibly over it; the single-page results are then merged into one document. So what you download is a new file built from page images, and the original's links, bookmarks, annotations and form fields are not carried across.
What that resolution means in practice
On A4, 1100 pixels across works out to about 133 dpi. That is ample for recognition — modern engines do not need 300 dpi to read ten-point type — and comfortable for reading on screen. It is well below what an archival scan would be.
Two consequences follow. The first is welcome: because every page is re-encoded at that size, a large scan usually comes back noticeably smaller, which is a real benefit if the file was going to be emailed. The second is a reason to keep your original: where the resolution itself matters — evidence, signed contracts, anything going into an archive or being enlarged for a detail — the OCR'd copy is not a replacement for the scan, it is a searchable companion to it.
The whole document is processed, too. There is no way to recognise only pages 4 to 9, so a 400-page scan is a 400-page job.
What wrecks accuracy, in order
Skew. Recognition works by finding rows of text, and a page rotated even two or three degrees smears those rows together. This is the single biggest lever and the easiest to fix. Deskew PDF measures the tilt by rendering each page, sampling about 20,000 points for ink, test-rotating from 8° left to 8° right in quarter-degree steps, and taking the angle that makes the horizontal ink histogram spikiest. Below 0.2° a page is declared straight and copied through untouched, so a 200-page scan with four crooked pages comes back with 196 pages nobody has been near. Running it before OCR is close to free — recognition renders everything to around 1100 pixels wide anyway, near what the deskewed raster already holds.
Contrast and noise. A photograph of a page taken at an angle in poor light, with a shadow across the gutter, is a much harder problem than a flatbed scan. If the source is a phone photo, Scan to PDF is the right way in.
Layout. Multiple columns, tables, sidebars and callouts all challenge the reading-order logic. The characters are usually right; the order they come out in may not be.
The font itself. Decorative faces, script, very condensed type, and text printed over a background image are all materially harder than a book face on white paper.
Choosing a language, and why only one
One language per run. English and Korean ship with the site and need no download, and the English + Korean option loads both at once — the only mixed-language combination available. The other thirteen fetch about 12 MB of recognition data once, after you agree to it, and it stays cached in your browser afterwards. Your document still never moves.
This is not an arbitrary restriction. A recognition model is trained on a character set and a language model, and running two at once means both are competing to claim every glyph. For a document that mixes languages, choose the one most of the words are in and accept that the other will come out poorly.
It also means the wrong choice is expensive. Running an English model over a French document does not fail — it produces confident, plausible, wrong output, with accented characters silently mapped to their nearest unaccented neighbours. Check the language before you start a long job.
When not to run OCR at all
Before anything else, try selecting a word in your PDF. If the text selects, there is nothing to recognise, and OCR would replace a crisp vector page with a 1100-pixel image for no gain whatsoever. This is the most common wasted operation on any PDF site.
The other case is where recognition is already happening for you. PDF to Word and PDF to Excel both recognise text-less pages during the conversion, so a scan converts to editable text in one step — though that built-in pass covers English and Korean only, which is the reason to run OCR separately when your document is in something else.
Compare PDF goes the other way and refuses to run on two scans: the text layer OCR adds is precisely what it needs to work, so OCR first is not optional there. And PDF to Markdown does not run OCR at all — it reads the existing text layer, so a scan converts happily and hands back a file with page marks and no words. The character count under the download is the tell: a forty-page scan coming back under a thousand characters means there was no text layer to read.
Questions
Does OCR change how my scan looks?
Slightly. The page is re-encoded as a JPEG about 1100 pixels wide at quality 0.8, so on close inspection it is softer than a 300 dpi original. What you see is still the scan, not a reconstruction — the recognised text is invisible and sits over the top.
Why can I search my OCR'd PDF but the words look untouched?
That is exactly how it is meant to work. The visible page is the original image; the searchable text is a separate invisible layer positioned over it. Searching, selecting and screen readers all use the invisible layer.
How accurate is it?
On a clean, straight, single-column scan of ordinary type, very. On a skewed phone photo of a multi-column page in a decorative face, poor. Accuracy is dominated by the input, which is why deskewing first is worth more than any setting.
Can I OCR just a few pages?
No — the whole document is processed. If you only need pages 4 to 9, extract them first with Extract pages, run OCR on the extract, and keep the original alongside it.
Do my files get uploaded for recognition?
No. Recognition runs in your browser. The only thing that is downloaded is the recognition data for the language you pick — about 12 MB for the languages that are not bundled, fetched once and cached — and the document itself never moves.