NotebookLM: answers locked to your sources

What NotebookLM is for

NotebookLM is a source-grounded research tool from Google. You create a notebook, add your sources (documents, PDFs, Google Docs, pasted text, web pages and more) and then everything you ask is answered from those sources only, with inline citations that link to the exact passage the answer came from. It's a different promise from a chat: instead of "here's a plausible answer from everything I know", it's "here's what your documents actually say, and here's where."

That constraint is the feature. Because a notebook can't reach past its sources, it's far harder for it to wander off into a confident invention, and when it does summarise or answer, you can click the citation and check the claim against the original in one move. This is grounding made strict: the model is pinned to your material. It doesn't make NotebookLM immune to error, since it can still misread a table or over-summarise, but it changes the failure mode from "made something up from nowhere" to "misread a specific passage you can go and check."

We'll build a notebook for Fernway, where you're Tom Elliott getting across the customer-feedback rollout before a meeting. The paperwork is scattered across a brief, some rough notes, an email thread and a policy, exactly the situation NotebookLM is built for.

Building a notebook over a source set

Create a new notebook and add your sources. For the Fernway feedback rollout, add the project brief, the meeting notes, the email thread and the remote-working policy. NotebookLM accepts a wide range of source types (the current list is on Google's add sources help page) and the durable rule is what matters: once a source is in the notebook, it's fair game for grounding, and anything not in the notebook is invisible to it.

Now ask a question that spans the whole set. Because the notebook can only answer from what you added, you get a synthesis across your documents with citations to each source.

Across these sources, summarise the Fernway customer-feedback rollout: the problem it solves, who owns which deliverable, and the key dates. Cite the source for each point, and note anything important that the sources don't actually settle.

A good answer names Priya as sponsor and Maya as lead from the brief, pulls the "customers never hear back" problem from both the brief and the meeting notes, lists the week-1 dates, and flags, for instance, that the holiday-cover rota is still unassigned. Every point carries a citation you can open. That last habit, clicking through, is the whole reason to use a grounded tool: the checking is built in, not bolted on.

Audio Overviews: get across it hands-free

NotebookLM's best-known feature is the Audio Overview: it generates a spoken, podcast-style discussion between two AI hosts who talk through the key points of your sources. You can generate one for a notebook and listen while commuting, walking or doing something else, a useful way to get across a dense set of documents when reading isn't practical. Newer versions also let you steer the overview (asking it to focus on particular topics or pitch it at a particular audience) and, on some accounts, join in and ask questions; the current capabilities are on Google's Audio Overview help page.

The honest framing: an Audio Overview is a briefing, not a source of record. The two hosts are engaging and clear, which, exactly like a polished written report, can make the content feel more authoritative than the underlying documents warrant. It's excellent for getting the shape of a topic into your head before a meeting; it is not the thing to quote a figure from without opening the notebook and checking the citation. Use it to get oriented fast, then read the grounded answer for anything load-bearing.

Newer capabilities: reasoning, code and generated outputs

A June 2026 upgrade widened what a notebook can do beyond answering questions and generating audio. On accounts that have it, NotebookLM can apply more advanced reasoning to your sources, run code in a secure, sandboxed cloud environment, and turn what it finds into charts, spreadsheets and slide decks rather than only text. In practice that means you can ask not just "what do these sources say" but "build a chart of the figures in this report" or "draft a short slide deck summarising these documents", with the output still grounded in the sources you added.

As ever, this arrived unevenly: it landed first for Ultra subscribers and specific Workspace accounts, so if you don't see it, that's an availability matter tied to your plan, not a fault. The honest checking habit doesn't change either. A generated chart or deck is only as right as the sources behind it, and a figure the notebook computes for itself is a claim to verify against the source, not a result to trust on sight.

When a notebook beats a chat

You now have two overlapping tools, an ordinary Gemini chat and a NotebookLM notebook, and the Power User skill is knowing which fits. Reach for NotebookLM when:

• You have a fixed, defined set of sources and want answers grounded strictly in them: a contract, a policy, a project's paperwork, a stack of research papers.
• Traceability matters. You need to click from every claim back to the exact passage, because you'll be quoting it, defending it, or acting on it.
• You're getting across a lot of material and want a reliable summary or an Audio Overview rather than a general explanation.

Reach for an ordinary Gemini chat (or a Gem) when you want the model's broad knowledge, the live web, or a quick draft, for when the answer shouldn't be limited to a fixed document set. Deep Research, which you met at Practitioner level, is the mirror image of NotebookLM: it goes out to the open web of unknown quality, whereas NotebookLM stays in your trusted sources. A rough rule: if the honest answer is "it depends what my documents say", use NotebookLM; if it's "it depends what's true in the world", use a chat or Deep Research.

From these sources, produce a one-page briefing for a manager: three-sentence overview, a table of decisions already made with the source for each, and a separate list of open questions the sources leave unresolved (with who would need to answer them). Cite everything.

Do any of these sources contradict each other, or leave an important question unanswered? List each conflict or gap, quote the relevant lines with their source, and say what additional information would resolve it.

Try it now

Common mistakes

• Treating the Audio Overview as the source of record. It's a briefing, not a citation. Two fluent AI hosts make thin material sound settled. Get oriented from the audio, but quote figures only after opening the grounded answer and checking it.
• Assuming grounded means correct. NotebookLM can't invent from nowhere, but it can still misread a table, over-summarise, or miss a caveat buried in a source. The citations exist to be clicked. A claim you'll act on is one you check against the passage.
• Forgetting what's not in the notebook. A notebook only knows its sources. If you didn't add the file, its silence isn't reassurance. It simply can't see it. Missing an important source skews every answer, quietly.
• Using it when you needed the open world. For a general question, current events or a quick draft, a source-locked notebook is the wrong tool. It can only tell you what your documents say. Match the tool to whether the answer lives in your files or out in the world.
• Adding a stale version of a document. A notebook is only as current as the sources you put in it. Add an out-of-date brief and you'll get confident, well-cited answers about the wrong plan.

Keeping current

NotebookLM is developing quickly. Supported source types, Audio Overview controls, the newer reasoning, code and generation features, sharing and availability all change, and much of it is still rolling out by plan. Trust your screen over this lesson and check Google's official pages: NotebookLM Help, Add or discover new sources and Generate an Audio Overview. Accurate as of 14 July 2026.