You attach a PDF, ask Gemini a perfectly reasonable question about it, and get either "I can't read this file" or - worse - a confident answer that plainly isn't from your document. The same failure shows up in the Gemini app, in Gemini for Drive and Docs, and in NotebookLM, and in most cases it has one boring cause: the PDF is a scan.
The usual suspect: no text layer
Gemini answers questions about a PDF by working over its text. A born-digital PDF - exported from Word, Docs, LaTeX, a design tool - carries that text internally, and Gemini generally does fine with it. A scanned or photographed PDF carries only images of pages. Unless something runs OCR to recognise the characters, there is no text in the file at all, and every text-based tool - Gemini included - sees an empty document.
That "empty document" failure explains both symptoms:
- "I can't read this" - Gemini extracted nothing and says so.
- Hallucinated answers - Gemini extracted nothing but answers anyway, from general knowledge and the filename. This is the dangerous version, because it looks like it worked. The same failure mode affects ChatGPT: see why ChatGPT can't read your PDF.
The ten-second check: open the PDF and try to select a sentence with your cursor. If nothing highlights, it's image-only. That's the whole diagnosis.
Ruling out the other causes
If the selection test shows real text, work through the short list:
- Size and page limits. Long PDFs can exceed what your plan and surface accept - limits differ between the consumer app, Workspace, and API, and they change over time. A file that's over the line may be truncated or refused.
- The upload didn't take. Attachments occasionally fail silently, especially on mobile. Re-attach and confirm the file shows before asking.
- Restricted or malformed files. Password-protected and unusually built PDFs can fail extraction even when they display fine.
But if you can't select text, don't spend time on any of the above - it's OCR or nothing.
Fix 1 - Make the PDF searchable, then re-upload
The durable fix is to add a real text layer with OCR and give Gemini the searchable copy. Free and paid routes - OCRmyPDF, Acrobat, a Google Drive conversion - are covered step by step in how to make a PDF searchable. Once the text layer exists, Gemini can extract it like any other file.
This is worth doing when the document will be used repeatedly, or by other people and tools. The drawback is the extra round trip for every scan you meet.
Fix 2 - Read it in the browser, with citations
When you just want the answer, skip the conversion. Sidenote is a browser extension that reads the PDF you have open - including in Drive - and runs OCR automatically when pages are image-only (a paid feature, included in the 7-day trial). Ask your question and every claim in the answer carries a citation that scrolls to the exact passage; claims the document doesn't support are dropped server-side rather than dressed up.
That last part is the practical difference for scans specifically: the failure mode you're escaping is a tool answering around a document it couldn't read. A reader that must tie every claim to a recognised passage can't fail that way silently - if the scan doesn't say it, you don't get it.
Gemini vs an OCR-first reader for scanned PDFs
| Gemini | Sidenote | |
|---|---|---|
| Born-digital PDFs | Handles well | Handles well |
| Scanned PDFs | Unreliable - often blank or refused | OCRs on ingest |
| Wrong-answer risk on scans | Can answer from general knowledge | Unsupported claims dropped |
| Verify an answer | Re-read the file yourself | Click the citation, page scrolls to the passage |
The two coexist happily: Gemini is a strong general assistant, and an OCR-first cited reader handles the specific file Gemini couldn't see. If your scans live in Drive, the Google Drive OCR guide covers that path end to end.