NotebookLM

An AI that is only allowed to read what you gave it - the clearest answer in the category to the question of where a claim came from.

Verdict

This is a documentation review, not a walkthrough: no account was opened for this page and no screenshots were captured, so nothing here is a first-hand claim about what the interface looks like today. On the documented design, NotebookLM is the most structurally honest product in consumer AI, and the reason is a constraint rather than a capability. It answers only from the sources you upload, and it attaches an inline citation to the specific passage behind each sentence. That makes checking cheap, which is the only thing that makes an AI answer usable for work that matters. The free tier is substantial because the core behaviour is the free behaviour; paid tiers mainly raise the caps. The limits to understand are that grounding reduces invention without eliminating it, that a citation proves provenance and not correctness, and that a notebook is only ever as good as the sources someone chose to put in it.

Visit NotebookLM
We have no affiliate relationship with this vendor. Link goes to the official site.

Read this first: what kind of page this is

This page is a documentation review. No free account was opened for it, no sign-up flow was followed and no screenshots were captured, so nothing below is a first-hand account of what the product looks like or how long anything takes today. Everything here comes from the vendor's published documentation, support pages and the public record of how the product has changed.

The distinction matters on this site because the other kind of page here is a free-tier walkthrough, where an account is created, the vendor's own onboarding is followed, and screenshots are captured with our own tooling and dated. Those pages say so and show the captures. This one cannot, so it says so instead.

What a documentation review is still good for: describing what a product is designed to do, how its tiers are shaped, what its documented limits are, and where the design is unusual. What it cannot tell you: whether today's interface matches today's documentation, how fast anything feels, or what the current caps are. For those, open it yourself - it costs nothing to start - and treat this page as the orientation rather than the verdict on your experience.

What it actually does, in one paragraph

You create a notebook and add sources to it: documents, pasted text, slide decks, web pages, transcripts. Then you ask questions, and the model answers using those sources and nothing else. Each sentence in the answer carries an inline citation pointing at the passage it came from, and clicking it takes you to that passage in the source.

That is the whole product, and the important word in it is "only". A general chat assistant answers from everything it absorbed during training plus whatever it can search, which means the provenance of any given sentence is unknowable. This answers from a corpus you assembled, which means the provenance of every sentence is a click away.

The design consequence is that the product is narrow on purpose. It is not trying to be a writing assistant, a coding assistant or a general-knowledge oracle. It is trying to be the thing you use when you have forty documents and a deadline, and when being wrong would cost something. The documented feature set follows that intent: summarising across sources, finding where a theme appears, drafting briefings grounded in the corpus, and generating study and audio formats from it.

Who it is for

The fit is unusually specific and worth matching honestly against your own situation.

It suits anyone holding a pile of documents they are accountable for understanding. A contract set where the question is what the termination clauses actually say across all of them. A literature or reading list where the question is which papers disagree. A year of meeting notes where the question is when a decision was made and by whom. A regulatory or policy corpus where getting it wrong has consequences and where the answer has to be defensible to someone else.

It also suits students and anyone learning from a fixed body of material, which is the use the vendor's own feature set leans into most visibly.

It is the wrong tool in two cases. It is wrong when you want the model's general knowledge, because refusing to use that knowledge is the entire design; asking it a question your sources do not cover produces a refusal or a thin answer, correctly. And it is wrong when the sources themselves are the problem - when you do not know what to read, or your corpus is unreliable - because grounding faithfully reproduces the limits of what you fed it. For that job you want search with citations across the open web, which is a different product shape and the one the Perplexity review on this site covers.

How the free tier is shaped, and why that shape is unusual

Most free tiers in AI are demonstrations: the core behaviour is withheld or throttled hard enough that the free user experiences a reduced product designed to make upgrading feel necessary. This one is shaped differently, and the difference is structural rather than generous.

The core behaviour here is grounding plus citation, and that behaviour is present on the free tier. What the paid tiers raise is quantity: how many notebooks you can keep, how many sources go into one, how much you can ask in a period, and access to the heavier generation formats. The capability is not gated; the volume is.

That shape has a sensible explanation. The value of the product comes from the constraint, and a constraint cannot be sold in tiers. There is no half-grounded version to offer. So the free tier ends up being the product, and the paid tiers end up being for people whose corpus or usage is large.

The practical consequence for a buyer is that you can evaluate this properly for nothing. That is rare enough to state plainly, and it also means the only honest advice about the paid tier is to hit the free caps first and let that tell you whether you need more. Any specific cap written on this page would be wrong within months; read the current figures on the vendor's own pricing page.

The limitation that matters most: grounding is not verification

This is the section to read if you read only one, because the product's greatest strength is also the source of its most dangerous misunderstanding.

Grounding means the model's answer is constructed from your sources. It reduces the failure mode where a model invents a plausible fact from nothing, and that reduction is real and valuable. What it does not do is guarantee the sentence is right. The model can misread a passage, merge two passages that should not be merged, miss a qualifier, or summarise a hedged statement into a flat one. When that happens, the wrong sentence arrives with a citation attached, which makes it look more trustworthy than an uncited wrong sentence rather than less.

The second layer is that a citation reports provenance, not quality. If your corpus contains an outdated policy document, the answer will faithfully reflect the outdated policy. If you left out the source that contradicts the others, nothing in the product will mention the absence. It does not know what you did not give it.

The practical discipline follows directly: click the citations on anything that matters, read the cited passage rather than trusting the quotation, and treat the corpus itself as the thing to get right. The product makes checking cheap. It does not do the checking.

Where it sits against a general assistant and against search

Three product shapes get compared here and they answer three different questions, so choosing between them is mostly a matter of knowing which question you have.

A general chat assistant answers from its training plus whatever tools it can reach. It is the most capable at open-ended work - drafting, reasoning through something unfamiliar, writing code - and the least checkable, because the provenance of any sentence is not recoverable.

An answer engine searches the open web and then answers with citations to the pages it found. It is the right shape when you do not yet know what to read, and its limitation is that it chose the sources, not you. The Perplexity review on this site covers that shape in more detail.

This product inverts the last part: you choose the sources and it is confined to them. That is the right shape when you already have the corpus and the job is to understand it, and the wrong shape when the corpus is the open question.

In practice people use more than one, and the sequence is usually search to find the material, this to interrogate the material, and a general assistant to write something from the result. Treating them as competitors rather than stages is where most of the confused comparisons come from.

What the documentation suggests about getting good results

Four things recur in the vendor's own guidance and in the public record of how people use it well, and none of them requires any clever prompting.

Curate the corpus deliberately. A notebook with eight relevant sources outperforms one with forty where thirty are noise, because the model has to decide what is relevant and every irrelevant source is a chance to decide wrong. Narrow notebooks also keep you inside the caps.

Keep one notebook per question, not one per project. A notebook is a boundary around a corpus, and the boundary is the useful part. Mixing three unrelated bodies of material into one notebook degrades every answer in it.

Ask for the location, not just the answer. Questions phrased as where does this appear, which sources disagree, or what does each source say about this produce answers whose structure makes the citations easy to use, which is the point.

Name the source in the question when you want one source. Confinement to a corpus is not the same as confinement to a document, and asking about a specific one is more reliable than hoping it is weighted correctly.

Treat the generated formats as drafts. Summaries, briefings and audio overviews are derived from the corpus and inherit every flaw in it.

Privacy, data and what to check on the tier you use

This is the part worth reading before uploading anything sensitive, and it is also the part this page can least usefully summarise, because the terms differ by tier and by whether the account sits inside an organisation's workspace.

Four questions to answer from current documentation rather than from any review. Is uploaded content used to improve the models, and does the answer differ between the free tier, the paid consumer tier and an organisational account? Where is content stored and for how long after a notebook is deleted? Who inside your organisation can see a notebook, and is that governed by the product or by the surrounding workspace administration? And which regulatory commitments apply to the tier you are actually on, which is the question that matters if your corpus contains client, patient or employee material.

The general pattern across this category is that consumer tiers and organisational tiers have materially different data terms, and that the difference is the main reason an organisation should not let staff upload work material to personal accounts. That is a policy problem rather than a product flaw, but it is the most common way this category of tool causes an incident.

For anything confidential, contractual or regulated, read the current terms for your exact tier and get a decision from whoever owns that risk. A product page is not that decision.

Support, and the realistic expectation

Support is self-service in character: documentation, help pages, a community, and the general support channels of a very large vendor rather than a dedicated desk for this product. Organisational accounts inherit whatever support arrangement the surrounding contract provides, which is a different and usually better experience, and is not what a free or consumer-tier user gets.

The documentation itself is clear on the mechanical questions - what a source can be, how notebooks work, what the formats produce - and thinner on the conceptual ones, including the grounding limitation described above, which in our reading is under-emphasised relative to how much it matters.

The second realistic expectation concerns change. This is a fast-moving product in a fast-moving category, and the caps, the formats and the model behind it have all changed since launch. Anything you learn about the specific limits has a short shelf life, which is an argument for reading the current pricing and limits page rather than trusting notes, including these.

Support scores lowest of the four sub-scores at just above three. That is not a complaint about competence; it is a statement that at this price, in this category, support means documentation and a forum, and buyers should plan accordingly.

Frequently asked

Does it answer from general knowledge? No, by design. Ask it something your sources do not cover and you should get a refusal or a thin answer, which is the product working rather than failing.

Can it still be wrong? Yes. It can misread, merge or over-flatten a passage and present the result with a citation attached. Click the citation and read the passage for anything that matters.

Is the free tier enough? For most individuals with a focused corpus, yes. The caps are what you meet first, not missing capability.

What happens at the caps? Paid tiers raise them. The current figures are on the vendor's pricing page and change often enough that none are quoted here.

Should confidential documents go into it? Not before someone reads the data terms for your exact tier and owns that decision. Consumer and organisational terms differ materially.

Is it a replacement for search? No. It cannot find sources you have not given it. Use search to find material and this to interrogate it.

Why no screenshots on this page? Because no account was opened for it. That is what a documentation review means here.

Does this page earn a commission? No. There is no affiliate relationship behind it and the grade would be identical either way.

Verdict, and what this page cannot tell you

On the documented design, this is the most honest product shape in consumer AI, and the honesty comes from a constraint rather than a feature. By refusing to answer from anything except the sources you chose, and by attaching a citation to the passage behind each sentence, it makes verification cheap. Cheap verification is the only thing that makes an AI answer usable for work where being wrong costs something, and almost nothing else in this category offers it.

The free tier is unusually substantial for a structural reason: the core behaviour cannot be sold in halves, so the tiers divide volume rather than capability. That makes the product genuinely evaluable for nothing, and it means the only sensible upgrade advice is to hit the free caps and let that answer the question.

The limits are three and none of them is hidden. Grounding reduces invention without removing it. A citation proves where a claim came from, not that it is true. And a notebook inherits every weakness of the corpus somebody chose to build, including the sources they left out.

What this page cannot tell you: what the interface looks like today, how fast it feels, what the current caps are, or whether the documentation matches the product this week. No account was opened and no screenshots were captured. It costs nothing to check those yourself, and you should.

Visit NotebookLM
We have no affiliate relationship with this vendor. Link goes to the official site.