NotebookLM Advanced tutorial

Advanced · step 1 of 7

Decompose and chain

1. Split a big job into steps with a clear output each.
2. Feed each step's output into the next.
3. Check the output between steps, not just at the end.

Why this step matters: Small steps with a checked output each fail visibly and early, instead of one big answer that is wrong somewhere.

Advanced · step 2 of 7

Structured output

Ask for a fixed table or JSON so you can check and reuse it.

Return only JSON: [{"claim": "...", "source_url": "...", "quote": "...", "confidence": "high|medium|low"}]

Why this step matters: A fixed format can be checked by eye or by code and reused without retyping.

Advanced · step 3 of 7

Automation and agents

NotebookLM is an app, not an automation platform. Repeatable work comes from reusable notebooks: keep a template notebook with your standing sources and copy the prompts that work.

Why this step matters: Automation repeats mistakes at scale, so you need to know exactly what is run and what is checked.

Advanced · step 4 of 7

Failure modes to watch

• Answers that blend two sources. Click the citations.
• Audio Overviews that oversimplify or misstate a nuance.
• Missing info treated as absent. It only knows your sources.

Why this step matters: Knowing how it typically fails tells you where to look when you review its work.

Advanced · step 5 of 7

Cost control

The Standard (free) tier allows 100 notebooks, 50 sources per notebook, 50 chats a day, 3 Audio Overviews a day and 10 Deep Research runs a month. Paid Google AI plans raise this: Plus 200 notebooks and 100 sources; Pro 500 notebooks, 300 sources, 500 chats and 20 Audio Overviews a day; Ultra up to 600 sources. Each source can be up to 500,000 words or 200 MB.

Since Sep 2, 2026, some limits are compute-based: they refresh every 5 hours with a weekly cap (Google help, answer 16213268).

Free: 50 chats a day, 3 Audio Overviews a day, 50 sources per notebook. Compute-based limits refresh every 5 hours with a weekly cap (since Sep 2, 2026).

Why this step matters: Paying for features you don't use is the most common waste. Knowing the limits first tells you if the free tier is enough.

Advanced · step 6 of 7

Privacy and team rules

• Don't paste passwords, ID numbers, card numbers or other people's private details.
• Assume the company can read and keep your chats unless its official policy says otherwise.
• For health, money or legal decisions, confirm with a professional or an official source.
• Google says your NotebookLM data is not used to train models unless you send feedback. Don’t include private data in feedback.
• Shared notebooks show your sources to everyone you share with.
• Write down which data is allowed in which tool, and who approves paid plans.

Why this step matters: Anything you paste may be stored and read. Deciding what never goes in is easier than undoing a leak.

Advanced · step 7 of 7

Build a 10-task test set

1. Collect 10 real tasks with known good answers.
2. Run them now and score each pass/fail.
3. Re-run them when the model, plan or your prompt changes.
4. Keep the prompt that wins, not the one that feels better.

Why this step matters: Without a fixed test set, 'better' just means 'felt better'. Scores let you compare prompts and models fairly.