What you feed NotebookLM determines what you get out of it. These practices reliably improve output quality.
Best: PDFs with selectable text (not scanned images). Good: Google Docs, clean web URLs, YouTube transcripts. Avoid: Scanned PDFs (OCR degrades quality), very long documents over 500,000 words (content gets truncated), and sources that require login to access.
NotebookLM supports up to 50 sources per notebook. In practice, 5–20 focused sources outperform 40+ loosely related ones. More sources dilute the signal — the model has a harder time synthesizing when content is too broad or redundant.
3–8 high-quality sources on a specific question. Every source should be directly relevant to what you're analyzing.
8–25 sources. Group by theme across multiple notebooks if coverage exceeds 25 papers — then do a synthesis notebook with your notes from each.
Use one notebook per project or topic. Don't mix a dissertation chapter with personal notes with a work report — separate notebooks keep context clean and citations meaningful.
Add a brief "context note" as a text source: paste a paragraph explaining what you're trying to accomplish. For example: "I'm a PhD student analyzing labor economics papers for a literature review. I'm looking for arguments about wage stagnation, and I want to identify methodological differences between papers." This orients NotebookLM's responses without restricting what it can access.
NotebookLM gives vague answers: Your sources may be too general. Add more specific sources or ask more specific questions. It can't find something you know is there: Large PDFs sometimes index incompletely. Try re-uploading the specific pages as a separate source. Audio Overview is off-topic: The model summarizes all sources equally — if you want the Audio Overview to focus on a subset, create a new notebook with just those sources.