How to Use NotebookLM (Now Gemini Notebook) for Study and Research

It's the night before an exam, or the morning a literature review is due, and you have 14 PDFs, three lecture slide decks and a YouTube lecture you never watched. Asking a general chatbot to "summarise all this" gets you a smooth paragraph with no way of knowing which page it came from.

NotebookLM, which Google now calls Gemini Notebook, is built for exactly that situation. You load your own material as sources, and every answer comes back with citations pointing into those sources. This post is a working routine for using NotebookLM for study and research, with the limits as stated on Google's help pages on 2026-10-08. I haven't run a study project through it for this article, so the workflow below is built from Google's documentation and the prompts are templates, not recorded results.

What changed: NotebookLM is now Gemini Notebook

Google announced the rename on 16 July 2026, in Google's own announcement. Google's support site now lives at support.google.com/gemininotebook, and its own pages say Gemini Notebook throughout. Reports describe the rename as branding plus new features, notably a "cloud computer" per notebook that can write and run code, rolling out first to Ultra and some Workspace plans. I haven't verified the code-execution rollout beyond that, so check your own account before planning around it.

Everything below uses the current help-center wording. If you see "NotebookLM" in your old bookmarks, same thing.

How do I choose sources that give good answers?

Most bad results come from the notebook, not the question. A notebook is a pile of sources the model searches; the pile decides what it can say.

One notebook per question or course unit, not per life. The help page notes that with many sources the tool retrieves the most relevant passages for your question and builds the answer from them. A 50-source notebook on three unrelated topics gives it more room to retrieve the wrong thing.

Know what actually gets imported. From Google's source help page:

Two consequences you can act on today. First, a slide deck full of diagrams will only give you what the extractable text carries, so add your own notes as a pasted-text source. Second, if a web page is paywalled, copy the text in manually (check you have the right to).

Name sources so you can address them. Google's own tip: when several sources are selected, mentioning source names in the question helps narrow the search. "What does the 'Pharmacology Week 4 slides' source say about half-life?" beats "what about half-life?". Rename sources on upload if the filenames are final_v3.pdf.

Use the source checkboxes. In chat you can select a subset of sources. Use it to compare two papers (select only those two), or to check whether a claim exists in the primary source rather than your own notes.

Delete the filler. A source that is a bad scan, or a 90-page report where you need 6 pages, adds noise. Cut the PDF to the relevant pages first, or paste the relevant section.

How should I set up chat before asking anything?

In the chat panel, open Configure Chat. Google lists three conversational styles: Default (general research and brainstorming), Learning Guide (educational content, for grasping new concepts and skills), and Custom, where you describe a style or role. There's also a response-length setting.

For study, switch to Learning Guide or a custom role. A custom instruction I'd start from:

You are a tutor for a second-year pharmacology course. Answer only from the
selected sources. When a term is not defined in the sources, say so instead of
explaining it from general knowledge. Keep answers under 200 words unless I ask
for more. End each answer with one question that checks my understanding.

The "say so instead" line matters because the tool is source-grounded: Google's help page says that if the answer isn't in the sources, it won't provide one. You want it to tell you about the gap rather than refuse silently, and then you decide whether to add a source.

One caveat from the same page: very short sources may be referenced whole, without a citation to a specific passage. If a citation looks like the entire document, that's why.

What prompts work for studying?

The principle: make the notebook test you, not just summarise for you. Reading a good summary feels like learning and mostly isn't. The basics from the clarity and specificity lesson, stating the task, the level and the format, apply here too.

Map the territory first (select all sources):

List the 8 to 12 main concepts across these sources, ordered from foundational
to advanced. For each: one-sentence definition, and which source covers it best.

Find disagreement (select two or three sources):

Where do these sources disagree, or use the same term differently? Quote the
conflicting passages and say which one is more specific about method or data.

Explain-back (the one that exposes gaps):

I will explain [concept] in my own words. Compare my explanation with the
sources. List what I got right, what I missed, and what I got wrong, with
citations. Do not rewrite my explanation for me.

My explanation: [paste]

Question writing for exams:

Write 10 questions on [chapter/topic] at the level of the exam questions in the
'Past paper 2025' source. Mix short-answer and application questions. Give
answers separately at the end, each with a citation.

Google's Studio panel also generates Flashcards and Quizzes from your sources. Per the help page you can set difficulty (easy, medium, hard), type a prompt to steer focus, choose fewer/standard/more items, and it tracks which flashcards you marked "Got it" or "Missed it" so you can re-practise only the ones you missed. Flashcards can be downloaded as CSV, though notebooks that include a Play Books ebook may be restricted by the publisher. Mind Maps, Audio Overviews, Video Overviews, Slide Decks and Reports are also in the help center; Audio Overview is fine for a commute but I wouldn't treat a podcast-style summary as study. Quiz yourself instead.

Also useful: save good answers with Save to note. Google says tables and inline citations are kept in the note. Notes are only used in later prompts when you select them.

How do I use it for research rather than revision?

Research has a different failure mode: you're not checking recall, you're checking claims.

1. Build the notebook from real papers, not summaries of them. Upload the PDFs. Add the abstract-only pages separately, and label them, so you can tell which claims rest on full text.
2. Ask for evidence tables.

Make a table with one row per source: research question, method, sample size,
main finding, stated limitations. If a field is not stated, write "not stated".

The "not stated" instruction forces gaps into the open instead of letting the model smooth them over.

3. Click through the citations that matter. Google's help says citations use direct quotes from your sources and you can hover to see the full quote or select to jump to it in context. For any number or claim going into your writing, open it. This is the habit that makes the tool trustworthy rather than the tool itself. See the avoiding hallucinations lesson for why grounded tools still need verification.
4. Use "discover sources" as a lead generator, not a bibliography. You can type a research question in the Add sources search to find web sources and Workspace files. Treat what it finds like search results: read them before you cite them.
5. Keep a gaps note. Whenever it says the sources don't cover something, add it to a pasted-text note called "Not covered". That list is your reading plan.

If you need breadth across the open web rather than depth in a fixed set, a different tool fits better; see the ChatGPT deep research guide and the broader AI research workflows post.

What are the current limits and plans?

From Google's plans page and usage-limits page, read on 2026-10-08. Google marks these "subject to change", so look at the page before you rely on a number.

Other limits worth knowing:

• Usage is compute-based, not a fixed number of chats. Google says it factors in prompt complexity, the models and features you use, and chat length. The quota refreshes every 5 hours until you reach a weekly limit.
• Studio items show an expected usage cost. When you're out of quota you can choose Generate later (web only), which Google says may take a couple of hours.
• Sharing a notebook doesn't raise a collaborator's source limit.
• If you live in India, South Korea or Vietnam, Google says a visible watermark is automatically applied to generated images and videos.
• Duplicating a notebook was "not supported yet" and deleted notes couldn't be recovered, per the FAQ. Export what you care about.

I'm deliberately not quoting prices. They vary by country and bundle, and I didn't verify them on an official page.

One more caveat on terms: Google says consumer data isn't used for training unless you submit feedback, and Workspace accounts aren't human-reviewed. Check which account type you're using before uploading anything confidential, such as unpublished data or client documents.

NotebookLM or a regular chatbot: which should I use?

Note that the chat page now also describes agentic features (web search, running code, creating files such as charts, CSV and PowerPoint) that work with or without sources, and calls them experimental. In that mode, the "only your sources" guarantee isn't the point, so don't assume citations mean the same thing.

A one-hour routine for an exam or paper

1. Create one notebook for one unit. Upload only what's in scope. Rename sources. (10 min)
2. Set Learning Guide or the custom tutor instruction above. (2 min)
3. Run the concept-map prompt and fix gaps by adding sources. (10 min)
4. Run explain-back on the five hardest concepts. (20 min)
5. Generate a quiz at "hard" with a prompt naming the weak areas, and redo the missed flashcards. (15 min)
6. Save the best answers as notes, and download the flashcards as CSV. (3 min)

The point isn't that the tool does the studying. It's that every answer is checkable, and the explain-back step makes you do the part that sticks.