Written and maintained by CASRAI Editorial Board
Last updated
NotebookLM is Google’s AI notebook tool: you upload a set of sources (PDFs, Google Docs, Slides, web pages, audio, YouTube transcripts) into a notebook, and it answers questions, summarizes, and generates study material grounded specifically in what you uploaded — not the open web. As of this check, Google’s own help documentation refers to the product internally as “Gemini Notebook” throughout its current support pages, and the notebooklm.google.com domain now 301-redirects to notebook.google.com, though “NotebookLM” remains the name researchers search for and the one still used in casual reference. This page uses NotebookLM, matching common usage, and notes the rebrand where it matters.
For a research workflow specifically — synthesizing a pile of PDFs for a lit review, prepping for a grant panel, or getting oriented in an unfamiliar sub-field fast — the pitch is that answers cite back to a specific passage in a specific uploaded source, which is a real difference from an open chatbot that answers from general training data. That grounding is genuinely useful for orientation and synthesis. It is not the same thing as a citation you can put in a manuscript, and conflating the two is the mistake this page exists to head off.
What NotebookLM actually does with your sources
Once sources are uploaded, NotebookLM can: answer questions with inline references back to the source passage; generate a structured summary or study guide across the whole notebook; generate an Audio Overview (a synthesized-voice discussion of the notebook’s contents); and, in current versions, generate Mind Maps, Video Overviews, Infographics, Slide Decks, flashcards, and quizzes. All of these operate over the notebook’s uploaded sources as a closed set — it does not go out and search the live web or a citation database unless you explicitly add a web source, which matters for a research use case: it will not surface a paper you haven’t uploaded, and it has no awareness of anything published after your sources were added.
Source limits, and what they mean for a literature-synthesis project
Per Google’s current help documentation, a free-tier notebook accepts up to 50 sources, each up to 500,000 words or 200MB for an uploaded file; a Google Slides source is capped at 100 slides, and a Google Sheets source at 100,000 tokens. Multiple web URLs added at once need to be separated by a space or a new line. Google’s paid AI plans raise some of these limits, though the exact paid-tier source cap isn’t specified in the same documentation — check your institution’s current Google Workspace/AI plan entitlement rather than assuming a number here, since Google has changed these limits before without a version-dated changelog.
The practical read for a research workflow: 50 sources is enough for a focused review of a sub-topic or a first-pass orientation read, but it is well short of a full systematic-review corpus, which routinely runs into the hundreds or low thousands of screened records. NotebookLM is not built to replace dedicated systematic review screening software for that scale of work — it’s a synthesis and orientation tool for a bounded source set, not a systematic-review platform.
What “grounded” answers guarantee — and what they don’t
NotebookLM’s citations point back to a passage in a document you uploaded. That’s a real anti-hallucination mechanism for the specific claim it’s answering — you can click through and check the source sentence yourself, which a general-purpose chatbot answering from training data doesn’t let you do. But it is a citation to your own workspace, not a formatted scholarly citation (no DOI, no journal metadata, no BibTeX/RIS record), and it doesn’t verify that the source document itself is accurate, current, or the version of record. Treat a NotebookLM citation as “here’s where I found this in my own reading pile,” not as something you paste into a reference list.
For the actual bibliography, a dedicated reference manager remains the right tool — see CASRAI’s guides on the RIS and BibTeX interchange formats for how citation metadata actually moves between tools, and Citation Management with AI Writing Tools for the accuracy risks of letting an AI tool generate citation text directly rather than pulling it from a managed library.
Audio Overviews and the other generated outputs: useful for review, not for citation
The Audio Overview — a synthesized discussion of a notebook’s sources — is genuinely useful as a first-pass orientation while commuting or between meetings, and the Mind Map/Slide Deck/Infographic outputs can speed up building a panel or committee briefing. None of these are citable outputs in a manuscript, grant, or protocol; they’re study aids generated from your sources, and like any AI-generated summary they can compress or misstate nuance the source itself gets right. Use them to get oriented fast, then verify anything load-bearing against the actual source document before it goes into something you’re accountable for.
A practical workflow: where this fits between finding papers and writing them up
NotebookLM sits after discovery and before formal synthesis: find your sources first with a literature-discovery tool (see CASRAI’s coverage of citation-mapping and AI-summarization tools in Using AI in a Literature Review Without Breaking the Method and the roundup in Chat with PDF Tools for Researchers, which covers NotebookLM alongside Scholarcy, Humata, and similar tools), load the resulting PDF set into a notebook, use it to get oriented and check your own reading comprehension against the source text, then move the actual claims — verified against the primary source, not the AI summary — into your reference manager and manuscript. Skipping straight from an AI summary to a manuscript claim, without opening the underlying source, is the failure mode every institution adopting these tools should train against.
Before you upload unpublished or sensitive research data
Uploading unpublished manuscripts, pre-decision grant text, unpublished participant-level data, or anything covered by an IRB protocol or a data use agreement into any external AI tool — NotebookLM included — is an institutional data-governance question, not just a personal workflow choice. Check your institution’s current AI-tool guidance and Google Workspace agreement before uploading anything that isn’t already public, and don’t treat a general consumer sign-in as equivalent to an institutionally-vetted enterprise agreement; the two can carry different data-handling terms. See CASRAI’s framework for choosing and governing LLMs for research for the fuller institutional checklist.
NotebookLM vs. the tools it gets confused with
NotebookLM is not a reference manager (it doesn’t export a formatted bibliography or integrate with Word/Google Docs as a citation plugin the way Zotero or EndNote do), not a systematic-review screening tool, and not a literature-discovery engine that finds new papers for you the way citation-mapping tools do. It’s closest to a purpose-built version of the general chat-with-your-documents category — CASRAI’s Chat with PDF Tools for Researchers guide covers how it compares to Humata, Scholarcy, and general chatbots with file upload on that narrower dimension.
Frequently asked questions
Is NotebookLM the same product as “Gemini Notebook”?
Functionally, yes as of this check — Google’s current help documentation for the tool uses “Gemini Notebook” as the name throughout, and the notebooklm.google.com domain redirects to notebook.google.com. “NotebookLM” is still the name most researchers use and search for; expect the branding to keep shifting as Google continues folding standalone AI products under the Gemini name.
Can I use a NotebookLM citation directly in a manuscript reference list?
No. Its citations point to a passage inside your own uploaded source document, not to a formatted bibliographic record (no DOI, journal, volume/issue, or page metadata). Use it to locate the claim in the source, then cite the actual source through your reference manager.
Does the 50-source limit make NotebookLM unusable for a systematic review?
For the screening-and-full-corpus stage of a systematic review, yes — that scale calls for dedicated screening software. It’s a reasonable fit for a bounded synthesis task: a focused sub-topic review, a grant-panel prep read, or getting oriented in a smaller, curated source set.
Is it safe to upload unpublished data or manuscripts?
Check your institution’s current AI-tool and data-governance guidance first, and confirm whether your account is under an institutional Workspace agreement or a personal consumer sign-in — the two can carry different data-handling terms. This is an institutional-policy question, not something this guide can answer generically for every institution.
