The complete NotebookLM guide

The AI research notebook is a Google-built tool that keeps every answer tied to the material you uploaded. Unlike a general chatbot that draws on broad world knowledge, the assistant reads only the corpus you provide and cites the specific passage behind each claim. This single design choice makes the tool genuinely useful for any workflow where accuracy and traceability matter.

Stage 1 — Setup and first access

The tool runs entirely in the browser. Sign in with any Google account, navigate to the product home page, and click "New notebook." You do not need a special invitation or a paid plan to get started. The interface presents a source panel on the left, a chat pane in the centre, and a notes panel on the right. Every notebook you create is private to your account until you share it explicitly.

Before uploading anything, decide what problem the notebook is meant to solve. A tightly scoped corpus — three reports on the same topic, for instance — consistently returns sharper answers than a sprawling collection of loosely related files. If you are new to the tool, start with two or three PDFs on a subject you already know well so you can judge citation accuracy against your existing understanding.

Stage 2 — Building your first notebook

Give the notebook a descriptive name before adding sources. Names are searchable and you will accumulate notebooks quickly. The "untitled notebook" pattern becomes painful after the fifth project. A naming convention like "Topic — Month Year" works well for most users.

Once named, the notebook is empty. You will see a source-upload prompt in the left panel. The tool accepts PDFs, Google Docs, Google Slides, plain-text files, Markdown documents, public web URLs, YouTube video links with available transcripts, and raw pasted text. Each source gets indexed individually; the full-text search runs across the combined corpus once all indexing is complete.

Stage 3 — Adding and managing sources

Upload sources one at a time or in bulk. The free tier supports up to 50 sources per notebook; the paid tier raises that to 300. After upload, each source appears in the left panel with a word count and an indexing status indicator. Indexing typically takes 10–40 seconds per source depending on file size and current server load.

Sources can be renamed, temporarily hidden from the model, or deleted at any point without losing your notes. Hiding a source is useful when you want to test how the assistant's answers change with and without a particular document — a good way to verify whether a specific paper is contributing useful signal.

Maintain one notebook per coherent topic. Mixing sources from unrelated domains does not break the tool, but it does dilute retrieval precision. The long-context model reads the whole corpus, so irrelevant material can surface in answers alongside the passages you actually wanted.

Stage 4 — Chatting with your sources

The chat pane accepts open-ended questions, structured prompts, and requests for specific document types. Ask the assistant to summarise a single source, compare two documents, extract all figures or statistics, identify contradictions between sources, or generate a set of study questions from the corpus. Every response includes inline citation markers that you can click to jump directly to the source passage.

For deeper analysis, use multi-step prompts: first ask for a summary, then ask follow-up questions that build on the summary. The assistant retains context within a conversation thread, so you do not need to repeat background context in each turn.

Stage 5 — Generating notes and briefing documents

The notes panel on the right saves every response you pin from the chat. You can also use the "Generate" menu to produce pre-built formats: briefing document, FAQ, study guide, timeline, and table of contents. These templates run the model over your entire corpus and return a structured document in seconds.

Notes can be edited directly in the panel. Changes you make are saved automatically. Once you have accumulated enough notes, click "Combine and convert to document" to merge them into a single Google Doc — the most common export path for writers, researchers, and students who need to hand the finished work to a colleague or supervisor.

Stage 6 — Audio overviews

The audio overview feature converts your corpus into a two-host spoken summary. Click "Generate" and select "Audio overview." The model produces a conversational dialogue that runs 8–18 minutes for a typical corpus. You can customise the focus before generation — specify a particular angle, ask the hosts to emphasise a certain section, or request a shorter brief. Once generated, the audio can be played in the browser or downloaded as an MP3 for offline listening.

The free tier limits the number of audio overviews you can generate per day. The paid tier provides a substantially higher daily allowance and access to extended formats including the 40-minute "Deep Dive" mode and the shorter four-minute "Brief."

Stage 7 — Sharing and collaboration

Open the share sheet from the notebook menu bar. Invite collaborators by Google account or generate a link. Three roles are available: viewer (read-only access to notes and audio overviews), commenter (can annotate notes but not edit sources), and editor (full access including adding or removing sources). Read-only links are well suited to classroom use where students should interact with the material but not modify the underlying corpus.

Stage 8 — Exporting finished work

Individual notes export to Google Docs in one click. Combined note documents follow the same path. Audio overviews download as MP3. The chat transcript can be copied in full. There is no native PDF export from the tool itself, but the Google Docs copy can be printed to PDF through the standard browser print dialog.

Stage 9 — Advanced techniques

Power users chain several workflows together. A common pattern is to upload a first batch of sources, generate a briefing document, use that document as a source in a new notebook alongside additional material, then generate a second-pass synthesis. This "notebook chaining" approach is particularly effective for literature reviews where wave one covers the background and wave two covers recent work.

Another technique is source layering: upload the primary text first, generate a study guide, then add a critical response or counter-argument paper. Ask the assistant to update the study guide in light of the new source. The resulting revision highlights exactly where the two texts agree and diverge — useful for seminar preparation, policy analysis, and legal research. See the NIST AI resource hub for guidance on integrating AI research tools into risk-managed workflows.

Stage	Action	Typical time
Setup	Sign in and create notebook	Under 2 min
Naming	Add descriptive title	30 sec
Upload	Add 5–10 PDF sources	2–5 min
Indexing	Wait for status indicators to clear	1–3 min
First chat	Ask for an overview summary	Under 1 min
Notes	Pin useful responses, edit inline	5–10 min
Audio overview	Generate and download MP3	3–5 min
Export	Push notes to Google Docs	Under 1 min
Share	Invite collaborators via share sheet	Under 2 min