How to use NotebookLM, step by step

Google's AI research notebook is built around a specific workflow: you supply the material, the assistant reads it, and from that point every answer, every note, and every audio segment is grounded in what you uploaded. Understanding that loop makes the steps below click into place. Each stage described here corresponds to something the tool is doing internally — indexing, retrieving, generating — and knowing that context helps you get better results faster.

Before you begin

You need a Google account. The free tier requires no payment and has no time limit. Gather the material you want to work with — PDFs, Google Docs, links to web pages or YouTube videos, or a paste of raw text. Having three to five sources ready before you open the tool will make the session more productive than starting with a single file. The notebook is designed for corpora, not single documents. That said, a single PDF works fine for a first experiment.

The tool runs in a web browser on desktop and has native apps for Android and iOS. This walkthrough describes the web interface, but the steps map directly to the mobile apps. For AI tool safety considerations relevant to research and institutional settings, the NIST AI Risk Management Framework is worth reviewing before deploying any AI assistant in a professional workflow.

Step 1 — Create a notebook

Open the tool's home screen. You will see a grid of any notebooks you have already created, plus a prominent button labelled New notebook. Click it. A dialogue asks for a name. Give it something descriptive — "Q3 market research," "dissertation chapter 2," "product spec review." The name only helps you; it does not affect how the model processes your sources. You can rename it later from the notebook header.

The notebook opens to an empty two-panel view: sources on the left, chat on the right. Nothing happens yet because there is nothing to read.

Step 2 — Add sources

Click Add source. A menu appears with the available input types: upload a file (PDF, plain text, Markdown), link a Google Drive document or Slides deck, paste a web URL, add a YouTube link with a transcript, or paste raw text directly. Choose the format that matches your material.

You can add multiple sources before the assistant indexes anything — drag them in as a batch, or add them one at a time. A progress indicator appears below each source card while processing runs. Processing a twenty-page PDF typically takes under a minute. A long YouTube video with a dense transcript may take two to three minutes. Do not try to ask questions while sources are still processing; the model will work only from what has finished indexing.

Sources that fail to index will show an error icon. The most common causes are password-protected PDFs, image-only scans with no text layer, or web pages that block content scrapers. See the help desk page for fixes for each.

Step 3 — Preview the source index

Once a source finishes processing, click its card in the source panel. The tool opens a summary view showing the title it detected, a short description of the content, and a list of key topics it identified. This is not just cosmetic — it tells you whether the model understood the document correctly. If the summary describes something that has nothing to do with your file, the source probably failed to parse and should be re-uploaded.

Reviewing the source index before asking questions takes about thirty seconds per document and prevents a common frustration: asking a question and getting an "I cannot find that in your sources" response because a key document silently failed.

Step 4 — Ask your first question

Type a question into the chat panel on the right. Keep the first question specific — not "summarise everything" but something like "What are the three main arguments in the Smith 2023 paper?" or "Which sources discuss regulatory risk?" The assistant will produce an answer with numbered citation markers. Click any marker and the source pane will scroll to and highlight the exact paragraph the model used.

That citation behaviour is the feature that distinguishes this assistant from general chat tools. Every sentence is traceable. If a citation points to a passage that does not support the claim, that is useful information — it tells you either that the model mis-retrieved the passage or that the source itself is weaker than it appeared. Either way, the citation makes the problem visible rather than hiding it inside a confident-sounding answer.

Follow-up questions work as expected. The conversation accumulates context, so you can ask "Now compare that with what source three says" without re-stating the original question. To start a fresh line of inquiry without the accumulated context, click New chat thread in the panel header.

Step 5 — Generate a summary note

Switch to the Notes tab (the right panel has tabs for Chat and Notes). Click the Generate button — the available note types include a briefing document, a study guide, a timeline, a FAQ, and a summary. Choose Summary for a first pass. The tool reads the entire corpus and produces a structured note with section headings and inline citations.

The summary note is editable. Click anywhere in the text to add your own content, restructure the sections, or delete paragraphs you do not need. The citations remain linked to their source passages throughout. Saved notes persist in the notebook indefinitely — they are not regenerated unless you ask for a new version.

Step 6 — Produce an audio overview

Still in the Notes tab, look for the Audio Overview card. Click Generate. The tool will ask whether you want a standard overview or a customised one — if this is your first attempt, accept the defaults. Generation queues the job and typically completes in two to eight minutes. When it finishes, the card shows a playback control and a download icon for the MP3.

The output is a two-host conversation: one presenter introduces topics, the other asks clarifying questions, and the exchange covers the through-line of your sources at a pace designed for listening rather than reading. Typical overviews run eight to fifteen minutes. If you want more focus on a specific topic, use the customisation option to specify emphasis areas before generating.

The daily generation cap varies by tier. Free accounts have a lower daily limit. Plus accounts have a substantially higher cap. The counter resets at midnight Pacific time. The generated MP3 can be downloaded and loaded into any podcast app, shared as a file, or listened to directly in the browser or mobile app.

Step 7 — Share with a collaborator

Click the Share button in the top-right corner of the notebook. Enter the email address of the person you want to invite and assign a role. Viewer: the collaborator can read notes and listen to audio overviews but cannot edit anything. Commenter: adds the ability to annotate. Editor: full access including adding or removing sources and generating new content.

The collaborator receives an email invitation with a link. When they open the notebook, they see the same indexed sources you do — no data is duplicated, and no sources are copied to a new location. The share does not affect your source or generation quotas. If you are using the tool in a team setting, the Plus tier adds notebook-level retention policies and an overview of which sources each collaborator has opened.

Step 8 — Export your notes

Open any note and click the three-dot menu icon in its top-right corner. The export options are: copy as Markdown (paste into any editor), export to Google Docs (opens a new Doc pre-populated with the note content and citations formatted as footnotes), and download as plain text. For academic writing, the Google Doc export is the most useful because it brings the citation structure with it — each footnote references the source document and page where applicable.

Chat answers are not automatically saved to notes. To save a specific answer, click the bookmark icon on the answer card. This adds it to the Notes tab as a saved response, where it can be edited, combined with generated notes, and exported using the same options.

Step	What you click	What the tool does
1. Create notebook	New notebook > name it	Creates an isolated workspace with its own source index
2. Add sources	Add source > choose file / URL / paste	Chunks, embeds, and indexes each source for retrieval
3. Preview index	Click any source card	Shows the model's parsed summary and key topics
4. Ask a question	Type in the chat panel	Retrieves relevant passages, generates a cited answer
5. Generate a note	Notes tab > Generate > Summary	Synthesises the full corpus into a structured, cited document
6. Audio overview	Audio Overview card > Generate	Produces a two-host podcast covering the corpus's through-line
7. Share	Share button > invite by email > assign role	Grants collaborator access to the same indexed corpus
8. Export	Note > three-dot menu > export format	Outputs the note with citations as Markdown, Google Doc, or text

Getting more from the tool after the first notebook

Most users find that the quality of their questions improves rapidly after the first session. The model is better at synthesis than retrieval — asking it to compare two sources, find contradictions, or identify what the sources collectively say nothing about tends to produce more useful output than asking it to summarise a single document. The most productive notebook workflows treat the assistant as a reading partner rather than a search engine: you have read the material, the assistant has too, and the conversation is about what you both noticed.

Source organisation also matters. A notebook works best when the sources share a coherent scope. Mixing a business strategy document with unrelated personal notes will produce weaker cross-references than a notebook where every source bears on the same question. Many experienced users maintain one notebook per project or research question and move sources between notebooks as projects evolve.