How to Use NotebookLM: Turn Your Documents Into a Research Assistant

What Makes NotebookLM Different From Chatbots

A general chatbot answers from everything it absorbed during training, plus whatever it can find on the web. NotebookLM answers from your uploaded sources and nothing else. Ask it something your documents do not cover and it tells you it cannot find that, rather than filling the gap with a plausible guess.

That constraint is the whole point. Most professional AI anxiety traces back to one fear — that the summary in your board pack contains an invented figure. Grounding does not make the tool infallible, but it moves the failure mode from “quietly wrong” to “visibly missing,” which is a far easier problem to manage. If the underlying mechanic is new to you, our explainer on why AI models hallucinate covers what is actually happening.

The second answer is the one you can take into a meeting. It names the document, it names the clause, and if it is wrong you will find out in four seconds rather than four weeks.

There is a cost to that constraint, worth naming early. A grounded notebook will not bring in context you did not supply, and it will not argue with your documents. Anyone hoping for an assistant that goes off and reads the internet is describing a different tool. What you get instead is narrower and more useful for work: a reader that has genuinely absorbed all eleven of your PDFs and will not pretend about what is in them.

NotebookLM sits alongside the rest of the Google Gemini family but behaves nothing like the chat app. Per Google’s July 2026 announcement of the rename, it stays a standalone research tool while syncing notebooks into the Gemini app. Worth knowing: you now reach it at notebook.google.com, though the old notebooklm.google.com address redirects there automatically.

The Clean Shelf Method Explained

Here is what most guides get wrong. They treat NotebookLM as a search box and spend their word count on the podcast feature. In practice, output quality is decided almost entirely before you type anything — by which documents you put in the notebook and which you leave out.

The Clean Shelf Method is the approach we teach at PromptPeakAI for exactly this, and it is the part of how to use NotebookLM that actually moves the needle. A notebook is a shelf. Anything on the shelf can be quoted back at you as fact, so you curate the shelf before you ask a single question. Four moves:

1. Shelve — put in only the documents that genuinely belong to this question, current versions only.
2. Scope — ask something your shelf can actually answer, and say which sources to use.
3. Spot-check — click through two or three citations before you trust the answer.
4. Ship — export into the format the recipient actually needs.

Step three is the one people drop first and regret most. Citations are only useful if somebody clicks them, and the tool cannot tell you that a source document was itself out of date. Our guide to fact-checking AI research before you put your name on it goes deeper on what to check and what can safely pass.

How to Use NotebookLM in Ten Minutes

Everything below is on the free plan. You get 100 notebooks and 50 sources per notebook, and each source can run to 500,000 words or 200MB — comfortably a 300-page manual per file. How much you can generate is metered separately, which we come back to later.

Everyone learning how to use NotebookLM starts here. Go to notebook.google.com, sign in, and create a new notebook. Resist the urge to build one giant notebook for your whole job. Name it after the decision it serves: “Q3 competitor review,” not “Research.”

Add sources from the left panel. NotebookLM accepts Google Docs, Google Slides, PDFs, text and markdown files, web URLs, pasted text, public YouTube links, and audio files. Superseded drafts and last year’s policy are the two most common contaminants — leave them out.

Ask something specific, name the output shape, and tell it what to do when the answer is not there. That last instruction is what stops a thin answer being padded out.

Using only the sources in this notebook, answer the following for [your role, e.g. an operations manager preparing a Thursday briefing].

Question: [your specific question]

Rules:
- Cite the source for every factual claim.
- If the sources do not answer part of this, say "not covered in sources" for that part. Do not fill the gap.
- Flag any place where two sources disagree with each other.

Format: short paragraphs under clear headings, no more than 400 words.

Click two or three of the citation markers. You are checking one thing: does the cited passage actually say what the answer claims it says. Ninety seconds here has saved a lot of people a bad meeting.

If your first answer comes back thin, the usual cause is scope rather than the tool. A question spanning eleven documents needs breaking into three narrower questions — the same principle covered in synthesising multiple documents into one brief.

Organising sources for cleaner citations

Once a notebook holds more than about a dozen sources, how you organise the shelf starts to affect the answers. A few habits that pay for themselves quickly:

• Rename files before uploading. Citations quote the source name, so “Kestrel MSA 2026 signed” is far more useful in an answer than “final_v3.pdf”.
• Untick sources you are not currently asking about. You can select which sources a question draws on, which sharpens answers and speeds up generation.
• One version of any document. Two drafts of the same policy on one shelf is how you get a cited answer from the superseded one.
• Split by decision, not by department. Notebooks cannot query each other, so the boundary you draw is permanent for that piece of work.

None of this is glamorous, and it is the single biggest difference between people who find NotebookLM transformative and people who try it once and conclude it is overhyped. Knowing how to use NotebookLM well is mostly librarian work.

Three Workflows That Save Real Hours

Most guides on how to use NotebookLM stop at the setup. These are the three workflows that survive contact with an actual week of work. Each one is a shelf, a scoped question, and an export.

Onboarding without the repeat questions

Build one notebook holding your policy PDFs, benefits guide, security rules, and team process docs. Generate a briefing, then share the notebook itself with the new starter so they can ask their own questions instead of queuing at your desk. HR leads tend to report the biggest saving here, because it removes the twentieth identical Slack message rather than the first.

Using only the uploaded HR and operations sources, build a first-week onboarding guide for a new [job title] joining the [team name] team.

Include:
- What they must complete on day one, with the policy each item comes from
- Who approves what, and the named process for each approval
- The five questions new starters in this role most commonly need answered

Cite the source document for every item. Mark anything the sources do not cover as "confirm with HR".

Competitor synthesis you can defend

Load competitor reports, launch announcements, and public YouTube keynotes into one notebook, then ask for a structured comparison. The advantage over doing this in a chatbot is that every row of the resulting table points at a document, so when your director asks where a claim came from you have an answer.

The first question asks the notebook for a judgement it has no basis for. The second asks it for extraction, which is what it is genuinely excellent at — and leaves the strategy call with you, where it belongs.

RFPs and contract extraction

Put your product specs, security documentation, and prior approved answers on the shelf, then work through requirements one at a time. For contracts, ask for a comparison table across documents rather than a summary of each. Our full workflow for writing RFP responses with AI pairs well with this as the drafting layer.

From the uploaded contracts only, extract the following for each agreement:

- Document name
- Termination notice period
- Liability cap
- Governing law

Present as a table, one row per agreement, with a citation in each row.
If a field is absent from a document, write "not stated" rather than inferring a market-standard value.

Turning Research Into Shareable Deliverables

The Studio panel on the right is where a notebook stops being a private thinking space and becomes something you can send. Everything it produces is generated from your sources, so the citations carry through.

That last column matters more than it used to. Since September 2026, Google has metered Gemini Notebook on a compute budget rather than a fixed number of generations per day, weighing prompt complexity, chat length, how many sources are attached, and which feature you pick.

Your budget refreshes every five hours under a weekly ceiling, and Studio shows an expected cost bar before you generate. Video Overviews and Slide Decks are the two you can queue for later when you run out — which tells you plainly which ones cost most.

Studio outputs improve considerably if you brief them rather than accepting the default. Most people click Report and take whatever appears; you get a far more usable draft by telling it who the reader is first.

Create a briefing report from the sources in this notebook for [audience, e.g. a director who has not read any of these documents].

Structure:
- A five-line summary of what the sources collectively say
- The three findings that would change a decision, each with its citation
- Any point where the sources conflict or leave a gap

Keep it under one page. Plain language, no jargon. Do not include anything that is not in the sources.

Audio Overviews get the attention, and they are genuinely useful for a commute before a client meeting. Mind Maps are the quieter winner — they show how themes connect across documents, which is exactly the thing that is slow to see by reading. Deep Research, which goes and finds public sources for you rather than working from your shelf, is one of the more expensive things you can run against a free budget; our guide to Gemini’s Deep Research mode covers when that trade is worth making.

What About Your Confidential Company Files

Before you upload a single client file, settle this. It is the question that decides whether you can use the tool at all, and nearly every guide gets it wrong by framing it as free versus paid. It is not. The line runs between account types, and there is a third tier most articles miss entirely.

Google’s position, set out in the Gemini Notebook privacy and terms documentation: your content is not used to train the foundational models — unless you submit thumbs-up or thumbs-down feedback. Do that, and trained reviewers may see the full context of that interaction, retained up to three years and disconnected from your account.

That second row is the trap. Being on a work account does not automatically buy you enterprise handling — on Business Base, Essentials Starter, Workspace Individual and a few others, NotebookLM is an additional service with consumer-grade data terms. On Business Starter and above it is a core service, like Gmail or Docs, and the stronger protection applies. Check for a plan badge next to your profile picture in the app, and if you are unsure, ask your admin before uploading anything sensitive.

If your company has no AI policy yet and you are quietly using tools anyway, you are in very normal company — and our piece on shadow AI at work covers how to get approval rather than forgiveness.

What NotebookLM Cannot Do

The grounding that makes it trustworthy is also the source of every one of its limits. This is the honest list.

There is also a straightforward capability boundary. For sustained reasoning over one very long document — a contract you want genuinely interrogated rather than extracted from — a long-context chat model often does better, as covered in our comparison of the best AI tools for long documents. Use the grounded notebook for many sources, and a chat model for deep work on one.

When Your Notebook Returns Nothing Useful

Four failures cover most of what goes wrong in the first week of learning how to use NotebookLM. None of them mean the tool is broken.

The last row is the one to internalise. Vague output is almost never a model failure — it is a scoping failure, and it is fixed by asking a smaller question rather than by rewriting the prompt three more times.

Where to Take This Next

You now know how to use NotebookLM properly, so start with one notebook this week, on a question you already have to answer. Build the shelf carefully, ask one scoped question, click two citations. That single pass teaches more than any amount of reading about the feature list.

The professionals who get the most out of this are not the ones who learned every Studio feature. They are the ones who built the habit of asking, before any research task, whether this is a shelf-and-question job.

• The monthly competitor check — same shelf, refreshed sources, one standing question.
• The quarterly policy review — every current policy on one shelf, asked what changed.
• The onboarding pack — built once, reused every intake, shared with the new starter.

Recurring work is where knowing how to use NotebookLM compounds, because the shelf is already curated and you are only changing the question.

Once a few notebooks are earning their keep, the natural next move is treating them as a permanent layer rather than one-off projects — which is the subject of building an AI-powered personal knowledge base for your job. If you also want to know which plan tier is worth paying for as your usage grows, our breakdown of Gemini free versus Google AI Pro and Ultra lays out the decision by use case.