You open a PDF, press Ctrl+F, search for a word you can see on the page, and get zero results. Or you feed it to an AI tool and it responds as if the file were empty. The document looks perfectly normal to you, which makes the failure feel like a bug. It isn't. Your PDF is a stack of photographs, and the fix is a fifty-year-old technology with an unhelpful acronym: OCR.
OCR in plain English
OCR, optical character recognition, is software that looks at an image of text and works out what the letters are. It finds the shapes that look like characters, decides "that's an r, that's an e", and reconstructs the words, turning a picture of a page into actual text a computer can select, search, copy, and quote.
The distinction that matters is between two kinds of PDF that look identical on screen:
- Born-digital PDFs were exported from software (Word, LaTeX, a browser's print dialog). The text inside them is real text; the file knows every character and where it sits on the page.
- Scanned PDFs came from a scanner or a phone camera. Each page is one big image. To you it reads fine; to software there are no words there at all, just pixels.
Search, screen readers, copy-paste, and every AI reading tool operate on the text layer. When there isn't one, they all fail the same way: silently, on a document that looks fine.
How to tell which kind you have
The ten-second test: open the PDF and try to drag-select a sentence with your cursor.
| Test | Born-digital PDF | Scanned PDF |
|---|---|---|
| Drag your cursor across a sentence | Text highlights line by line | Nothing selects, or the whole page selects as one block |
| Ctrl+F for a word on the page | Found | Zero results |
| Zoom in to 400 percent | Letters stay crisp | Letters go fuzzy or pixelated |
| Copy and paste into a text editor | Clean text comes across | Nothing, or garbage |
Two failures show up in the wild often enough to name. First, the mixed file: a born-digital report with scanned appendices, which passes the selection test on page 1 and fails it on page 40. Second, the bad-OCR file: someone ran OCR years ago with weak software, so text selects but pastes as gibberish. Both behave unpredictably in AI tools, and both are fixed the same way, by redoing the text layer properly.
How to fix a scanned PDF
The fix is always the same operation, getting real text into or out of the file. The right tool depends on the job:
- You need the text out, once, for free. Our PDF to text converter runs OCR on a scanned PDF in your browser and gives you the extracted text. No account, no upload ceremony, done.
- You need a searchable PDF that still looks like the original. Desktop tools (Adobe Acrobat, or free options like OCRmyPDF) add an invisible text layer under the page image, so the file looks identical but becomes searchable. Our full OCR extraction guide walks through the options and the accuracy traps.
- You need to read and question the document, not convert it. Sidenote OCRs scanned PDFs in place as you read them in the browser, then treats the recognized words like any other text: summaries, explanations, and answers, each with a citation. In-place OCR is a paid feature, and worth knowing about before you build a convert-upload-repeat loop out of free tools.
Why other AI tools say your PDF is unreadable
If you've hit "NotebookLM imported my PDF as blank" or "Copilot can't find anything in this file", you've met this exact problem wearing different logos. Most AI tools read only the text layer, and most don't run OCR for you: NotebookLM tends to import scanned PDFs as empty sources, and Copilot goes quiet on scanned files in SharePoint until the text reaches the search index. The tool isn't broken; it was handed a stack of photos.
What OCR means for citations
There's a second-order benefit that's easy to miss. Once a scanned document has a real text layer, an AI's claims about it become checkable. A grounded tool can quote the exact recognized sentence an answer came from and point you at it, which is source-grounding doing its job: with Sidenote, clicking a citation scrolls to the exact passage, so even on a document that started life as a photograph, you can verify every claim against the source. Without OCR there is nothing to cite, so any "answer" about a scanned file is a guess by construction.
That's the practical takeaway. OCR isn't just about making Ctrl+F work; it's the difference between a document your tools can only look at and one they can be held accountable to.
Frequently asked questions
Is OCR free?
Often, yes. Our PDF to text tool is free for extracting text from a scanned PDF, and free desktop options exist for producing searchable PDFs. Paid OCR (Adobe's, or Sidenote's in-place OCR while you read) buys accuracy on difficult scans and the removal of the convert-and-re-upload step, not the basic capability.
How accurate is OCR?
On a clean, straight, 300-DPI scan of printed text, modern OCR is routinely above 98 percent character accuracy, which reads as near-perfect. Accuracy falls with photographed pages, skewed or low-contrast scans, small print, tables, and handwriting. The failure mode to watch for is the plausible substitution ("1" for "l", "rn" for "m"), so check names and numbers against the image.
Why does my scanned PDF search find nothing even though I can read the page?
Because you and the search box are reading different things. You see the page image; search reads the text layer, and a scanned PDF doesn't have one. Run the file through OCR and search starts working, along with copy-paste, screen readers, and AI tools.
Can AI tools read scanned PDFs directly?
Most can't; they read the text layer and treat an image-only file as blank. A few, Sidenote included, run OCR automatically as you open the document, so scanned files behave like any other PDF, with answers cited to the recognized text. In Sidenote that's a paid feature; the free plan handles born-digital PDFs, and the free converter covers one-off scans.