NotebookLM review: what Google's research tool is actually good at, measured rather than asserted

The state of the reviews, which is itself informative

Before the product, a word about the sources you are choosing between, because it explains why this article exists.

Seven of the highest-ranking reviews for this query use the name NotebookLM exclusively and none mentions the rename. That is defensible for the ones written in 2024. It is less defensible for the enterprise-software review last updated on 23 July 2026, a week after the rename, and the CNET piece updated 24 June 2026, three weeks before it.

They also disagree with each other on checkable facts. One puts the top Ultra tier at $249.99 a month; Google's own page says $199.99, and has since a price cut in May 2026. One says free-tier notebooks can be shared with Viewer and Editor permissions; another lists "no native team sharing" as a con. Reviews published six months apart give paid source limits of 300 and 600 without either explaining the change.

That last one turns out not to be a contradiction, and the resolution is the sort of thing a table settles instantly: 300 and 600 are both correct, for different tiers. Pro allows 300 sources per notebook, the 30 TB Ultra tier allows 600. Two reviewers each looked at one tier and reported it as the paid limit.

What it is, in one paragraph

You upload sources into a notebook: PDFs, Google Docs, web pages, YouTube links, audio files. You then ask questions, and the model answers using only those sources, with an inline citation on each claim that jumps you to the sentence it came from. Everything else it does, from Audio Overviews to flashcards to infographics, is generated from that same closed set. The constraint is the product. A general chatbot answers from everything it has ever seen and treats your document as a suggestion.

The thing it is genuinely best at, with a number attached

Most reviews assert that grounding makes it more accurate. One study measured it, and the result is more interesting than the marketing.

Hagar, Agustianto and Diakopoulos, in "Not Wrong, But Untrue: LLM Overconfidence in Document-Based Queries", tested three tools against a 300-document corpus on TikTok litigation and US policy, varying prompt specificity and context size, then annotated every output for hallucination type and severity. Across all systems, 30% of outputs contained at least one hallucination. Split by tool: ChatGPT and Gemini at 40% each, NotebookLM at 13%.

Read the second finding alongside the first, because it is the one that should change your behaviour. The errors were not fabricated facts or invented numbers. They were what the authors call interpretive overconfidence: the model adding unsupported characterisations of a source, or turning an opinion a document attributed to someone into a flat statement of fact.

That failure mode is worse than a fabricated statistic in one specific way. A made-up number looks wrong the moment you check it, whereas "the report found X" when the report actually said "critics argue X" survives a casual check and fails a careful one. The citation link is what saves you, and only if you click it. Our citation accuracy guide covers what to check and when.

Two caveats on that study, since it is the strongest evidence in this review. The corpus was journalism-shaped, which is a demanding but narrow domain. And it tested the product as it stood before the 2026 updates.

What people actually use it for

The praise that recurs across every source is remarkably consistent, and it is not the feature Google leads with:

• Turning dense material into an Audio Overview for a commute. The single most repeated positive, from people converting semiconductor papers into something listenable to a writer generating one from their own CV. It is also the feature most easily mistaken for studying. Listening produces a stronger feeling of comprehension than comprehension.
• Synthesis across many papers at once. Ask what three sources disagree about and you get a real answer with three citations. Tedious work, done properly.
• A recurring briefing. Load a set of sources weekly, ask for the digest. Dull, and the use with the best effort-to-value ratio.
• Anything you will be examined on or will cite. The citation trail is the entire reason to pick this over a chatbot.

What actually breaks

The recurring complaints, and which of them are real:

Source ingestion failures

• Scanned PDFs with no text layer fail. The tool does not OCR for you. The workaround that works is opening the file in Google Docs first, which does OCR it, then adding that. See our scanned PDF guide.
• YouTube videos without captions fail, with a transcript-not-available error. This is architectural rather than a bug: it reads the caption track and does not transcribe audio itself. Videos less than 72 hours old also fail, because captions have not been generated yet. Paste the transcript manually instead.
• Password-protected PDFs are rejected outright.
• Ad blockers interfere with uploads often enough that it is the first thing to try when an upload silently fails.

Export, which was the biggest complaint until two days ago

Until very recently native export went to Google Docs or Google Sheets and nowhere else, which ruled out PDF, Markdown and plain text. Third-party browser extensions exist purely to work around it, which is a reasonable proxy for how much it annoyed people.

That changed on 4 August 2026, when Google finished rolling out chat-generated file outputs to all AI Pro subscribers. You can now ask the chat for a PDF, XLSX or PPTX and get the file. Two things it does not fix: chat history export, and getting an Audio Overview out as a file. And it is Pro and Ultra only. On the free tier, the old complaint stands in full.

Audio Overview quality

The most persistent quality complaint is the format rather than the fidelity: two hosts, always, with a chattiness that wears thin on a long listen. Length presets exist now. Output is available across a wide language range, and speakers of several languages report the intonation is noticeably less natural than the English version. That is plausible, and not something we can verify at scale.

The naming, which is a genuine mess

Google now uses near-identical names for three different things: Gemini Notebook (the renamed standalone product), Notebooks in Gemini (a feature inside the main Gemini app, shipped April 2026), and notebooks surfaced in Search. They sync, so a source added in one appears in the other, but they are not the same surface, and the overlap confused specialist commentators on rename day. If you have ever wondered why the thing you opened does not have Video Overviews, this is why.

The limits, which decide this more than the features do

From Google's published table, which Google labels as subject to change:

                        Free    Plus    Pro     Ultra20  Ultra30
-----------------------------------------------------------------
Notebooks               100     200     500     500      500
Sources per notebook    50      100     300     500      600
Chats per day           50      200     500     2,500    5,000
Audio Overviews/day     3       6       20      100      200
Video Overviews/day     3       6       20      100      200
Reports/quizzes/day     10      20      100     500      1,000
Deep Research           10/mo   3/day   20/day  75/day   200/day

Per source, all tiers: 500,000 words or 200 MB, whichever is smaller.

The number that binds in practice is sources per notebook. Fifty is generous for a taught module and insufficient for a literature review. Everything else on that table is headroom most people never reach. Notice too that the notebook count stops rising above Pro, so paying for Ultra buys depth per notebook rather than more notebooks.

Accepted source types: pasted text, PDF, .docx, .txt, .md, Google Docs, Google Slides (up to 100 slides), Google Sheets (capped at 100,000 tokens), ePub, public web URLs, public YouTube URLs, plus audio and image files. Web URLs bring text only. Images, embedded video and anything paywalled do not come across. Output is available in 80-plus languages, up from 50-plus in early 2025.

Price, and what the rename did not change

It is not sold on its own at any price. Access arrives bundled inside a Google AI subscription: $4.99 a month for AI Plus, $19.99 for AI Pro, $99.99 for the 20 TB Ultra tier and $199.99 for 30 TB, checked against Google's US page on 6 August 2026. Full breakdown in our pricing guide.

On the rename itself: notebooklm.google.com still redirects to notebook.google.com, so old links and bookmarks keep working. We covered that URL migration separately. Notebooks, sources and generated Audio Overviews carried over with no reported data loss. Google's own developer documentation still uses notebooklm-enterprise in its URL paths. The consumer brand changed and the namespace did not, which is a decent hint about how deep the change went.

Search behaviour has not moved either. "NotebookLM" still draws roughly ten times the monthly search volume of "Gemini Notebook." Both names will be in circulation for a long time, which is why this review uses both.

Privacy, stated precisely

The distinction here is genuinely important and usually reported wrong.

On a personal account, free or paid, Google's documentation states that notebook content is not used to train its foundational models unless you provide feedback. If you do click thumbs up or down, that submission can be reviewed by trained staff after being disconnected from your account, and retained for up to three years. Paying does not change this: there is no stronger default privacy on Pro than on free.

On a Workspace or Education account, the guarantee is unconditional: uploads, queries and responses are not reviewed by humans even when you submit feedback, and are not used to train models. If the material is confidential, that difference should decide which account you sign in with.

When something else is the better tool

The honest version of an alternatives section is not a ranking. Each of these beats it at something it was not built for:

• Perplexity Spaces: when the answer is not inside the documents you have. This one is closed-corpus by design and will not go and look.
• Elicit: for systematic literature review. It searches over 100 million papers and screens thousands against inclusion criteria; a notebook only ever sees the sources you uploaded by hand.
• Consensus: when the question is what the published evidence says overall, rather than what your PDFs say.
• SciSpace: for reading one dense paper closely. It lets you highlight an equation and ask about that passage, where a notebook works at whole-document level.
• Claude Projects: once the sources are vetted and the job is constructing an argument. Better at synthesis and prose; weaker at citing.
• ChatGPT Projects: when the deliverable involves running code or iterating on an artifact rather than answering from sources.
• Obsidian with AI plugins: when you want notes you own as local files, permanently, with no vendor holding them.

Our alternatives guide goes deeper on each.

Who should not use it

• Anyone who needs current information. It knows your sources and nothing else. Ask about last week and you get nothing.
• Anyone doing a large literature review on the free tier. Fifty sources per notebook is a hard wall, and splitting across notebooks defeats the point, because it cannot synthesise across them.
• Anyone wanting a note-taking app. There is no personal knowledge graph, no backlinks, no accumulation. Notebooks are project-scoped containers, not a second brain.
• Anyone who will not click the citations. The whole safety mechanism is that you can verify. If you intend to trust the output unchecked, a 13% interpretive-error rate is not meaningfully better than 40%, because you will catch neither.

Verdict

It does one job better than anything else available, and the job is narrow: making a set of documents you have chosen interrogable, with a verifiable trail back to the text. The measured error rate is roughly a third of what general chatbots produce on the same corpus, the free tier is unusually generous, and the export weakness that dominated criticism for two years was partly fixed this week for paying users.

It is not an AI assistant, a search engine, or a notes app, and most disappointment with it comes from expecting one of those three. Judge it on the one thing and it earns the recommendation.

People also ask

Is NotebookLM worth using in 2026?

Yes, for a specific job: questioning documents you have selected, where accuracy matters and you will verify. A peer-reviewed comparison measured its hallucination rate at 13% against 40% for ChatGPT and Gemini on the same corpus. It is not worth using for open-web research, note-taking, or literature review at scale.

Is NotebookLM the same as Gemini Notebook?

Yes. Google renamed NotebookLM to Gemini Notebook on 16 July 2026. Same product, new name and logo; notebooklm.google.com redirects to notebook.google.com and existing notebooks carried over. Confusingly, a separate feature called Notebooks in Gemini also exists inside the main Gemini app, and the two sync without being the same thing.

What is NotebookLM's biggest weakness?

It cannot go beyond the sources you give it, which is also the reason to use it. After that, the practical weaknesses are the 50-source free-tier cap and export, which only became workable for Pro and Ultra subscribers on 4 August 2026 and is unchanged on the free tier.

Does NotebookLM hallucinate?

Less than general chatbots, and not never. The measured rate was 13% of outputs against 40% for ChatGPT and Gemini on the same 300-document corpus. The errors were not invented facts but interpretive overconfidence: presenting an attributed opinion as established fact, which is harder to spot than a wrong number. Click the citations.

Is NotebookLM free?

There is a genuinely usable free tier: 100 notebooks, 50 sources each, 50 chats and three Audio Overviews a day. Paid access is bundled into Google AI subscriptions from $4.99 a month; it cannot be bought on its own.

Is NotebookLM better than ChatGPT?

For answering questions about documents you uploaded, yes, measurably. For anything needing general knowledge, current information, code execution or open-ended drafting, no. That is what ChatGPT is for. They are not really competing products.

Does Google train its AI on what I upload?

On a personal account, free or paid, Google states your content is not used to train foundational models unless you submit feedback, and feedback can be human-reviewed and kept for up to three years. On Workspace and Education accounts the protection is unconditional. Paying for a consumer tier does not upgrade your privacy.

Can I upload a Google Doc as a source?

Yes, along with PDFs, .docx, .txt, .md, Google Slides up to 100 slides, Google Sheets up to 100,000 tokens, ePub files, public web pages, public YouTube links with captions, and audio and image files. Each source is capped at 500,000 words or 200 MB, whichever is smaller.