NotebookLM review: excellent reader, poor filing cabinet

Filed under Review

By Gerald · 10 September 2026

A stack of books beside a lit desk lamp on a wooden desk near a window

Google renamed NotebookLM to Gemini Notebook in July 2026. The product underneath is the same tool, and everyone including Google's own help pages still calls it NotebookLM in casual reference, so that is the name this review uses. If you see "Gemini Notebook" in a menu or a help article, it is the same thing. Rebrands like this are common across the AI product space right now: the underlying capability barely changes, and the search traffic keeps using the old name for months afterward.

NotebookLM answers questions about a fixed set of documents you upload, and cites the exact source and passage for every answer. It is one of the best tools available for that specific job. It is a poor place to keep your actual knowledge long term, because you do not control the format, the retention, or what happens if Google changes the product again.

What NotebookLM actually does

You create a notebook, upload sources (PDFs, web pages, YouTube videos, Google Docs, Slides, audio files), and then ask questions. NotebookLM answers using only what you gave it, not its general training knowledge, and every claim in its answer links back to the specific source and location it came from. Beyond question answering, it can generate audio overviews (a two-host podcast-style discussion of your sources), study guides, flashcards, mind maps, and, since a February 2026 update, editable slide decks with PPTX export.

The scope restriction is the feature, not a limitation. Because it only draws from what you uploaded, it does not hallucinate facts from outside your source set the way a general chatbot can when a topic sits near the edge of its training data.

Source grounding and citation quality

A hand writing notes with a pen beside a printed page marked with yellow and pink highlighter — Grounded answers with a citation you can check, the way a highlighter marks a claim back to its source.

This is where NotebookLM earns its reputation. Ask it a question and the answer comes back with numbered citations, each one clickable to jump straight to the passage in the original source that supports the claim. In practice, checking a citation takes one click, and the citation is nearly always actually relevant to the sentence it is attached to, not a vague pointer to the right document without the right paragraph.

That verifiability is the single biggest difference from asking a general-purpose assistant the same question with a document pasted into the chat. A general assistant will confidently blend the document with its own background knowledge and rarely tell you which parts came from where. NotebookLM keeps the boundary visible, and that boundary is exactly what you want when the answer needs to survive someone else checking your work.

Where it beats a general assistant

For a research sprint, a batch of documents you need to understand quickly rather than remember forever, NotebookLM is genuinely faster than reading everything yourself and faster to trust than a general chatbot's unsourced summary. Feeding it a folder of PDFs before a meeting and asking it targeted questions is a legitimate use of an afternoon that used to take a full day.

The audio overview feature is also a real convenience for a specific kind of person. Someone who processes a discussion format better than a wall of text, or who wants a first pass on a topic during a commute before reading the primary sources directly.

Limits on sources, size, and export

According to Google's current documentation, a free notebook allows up to 50 sources. Each individual source is capped at 500,000 words or 200 MB, whichever limit is hit first. Paid tiers raise the source count, up to roughly 300 on the tier bundled with Google AI Pro and higher again on Google AI Ultra. The per-source size cap does not change on any plan.

Export is limited and manual. Individual notes can be sent to Google Docs or Google Sheets. Sources can be exported as Markdown or JSON one at a time. There is no built-in "export the whole notebook" action that reconstructs everything (sources, generated notes, chat history) as one portable archive. Third-party browser extensions exist to fill that gap, which is itself a sign that Google has not prioritized it.

Why it should not be your knowledge base

Here is the actual argument. A knowledge base is only as good as your ability to get your material back out of it in a form you control. NotebookLM was not built for that. It was built to answer questions about a bounded set of documents during the time you are actively researching something.

NotebookLM is a reading tool that happens to store what you gave it. It was never designed to be the place that knowledge lives.

If Google changes the product again, and it already has once in 2026, your notebooks are not guaranteed to survive that change in the same shape. Google says existing notebooks and shared links kept working through the rename, which is reassuring for this particular transition, but it is not a guarantee against every future one. Betting your only copy of research on a product whose name and packaging have already changed once this year is a real risk, not a hypothetical one.

The per-source size cap is a second, quieter problem. A genuinely large reference document, a full book, a long project archive, can bump into the 500,000-word ceiling, forcing you to split material in ways that were not your choice and that a permanent notes system would never impose.

How to use it alongside notes you own

The workable pattern is to treat NotebookLM as a research instrument, not storage. Do the research sprint inside it, ask your questions, get your grounded answers, and then write the conclusions, the parts you actually want to keep, into a notes system you control and can export from freely at any time.

I keep the actual output of a research session in Flow rather than leaving it inside the tool that produced it. Flow's notes are plain content you can export in full at any point, not material locked behind a product that could restructure itself again next year. Flow will not read your PDFs or answer questions about them the way NotebookLM does. That is not what it is for. It is where the answer goes once you already have it and want to keep it. That is really the same argument behind what owning your notes means more broadly.

If you also use a general assistant against your own notes rather than a fixed document set, connecting an assistant to notes you control covers a different, complementary workflow. One tool for grounded research over documents, another for querying the system where your own thinking already lives.

Who this is for

Anyone doing a bounded research task over a known set of documents
People who want citations they can actually verify, not just a confident-sounding summary
Researchers who want an audio-format first pass before reading primary sources directly

Who should skip it

Anyone looking for a permanent home for notes and reference material
Teams needing a straightforward bulk export of an entire notebook's contents
Anyone with a single source larger than the 500,000-word or 200 MB per-source cap

Frequently asked questions

Is NotebookLM free to use? Yes, the free tier (Standard) allows up to 50 sources per notebook with daily limits on chat questions. Paid access comes bundled into Google AI Plus, Pro, and Ultra subscriptions, which raise the source count and daily usage limits.

How many sources can one notebook hold? The free tier allows up to 50 sources per notebook. Paid tiers raise this to roughly 100 to 300 depending on the plan, with the highest tiers going further still. Every source, regardless of plan, is capped at 500,000 words or 200 MB.

Does NotebookLM use my documents for training? Google states that uploaded documents are not used to train its models. On personal Google accounts, providing explicit feedback on a response may allow human reviewers to see the associated content. Workspace and Education accounts are excluded from human review even when feedback is given, according to Google's documentation.

Can I export from NotebookLM? Individual notes can be sent to Google Docs or Sheets, and individual sources can be exported as Markdown or JSON. There is no native one-click export of an entire notebook's full contents as a single archive.

Is NotebookLM better than ChatGPT for research? For answering questions strictly grounded in a specific set of documents with verifiable citations, NotebookLM is the stronger tool. For open-ended reasoning, drafting, or questions that need knowledge beyond your uploaded sources, a general assistant like ChatGPT or Claude is the better fit.

My verdict

NotebookLM is the best tool I have used for asking verifiable questions about a bounded set of documents, and I reach for it during research sprints without hesitation. I would not trust it as the permanent home for anything I cared about keeping, and neither should you.