NotebookLM reads only your sources. Google reads them too.

It is the best research tool Google has shipped in years. Upload a stack of documents, ask a question, and every claim in the answer links to the exact passage it came from. It is grounded, it is cited, and it is also Google — now with a copy of your sources.

Start with what it gets right

Give credit before you take it. NotebookLM, built on Gemini 3 as of this year, does one thing that a general chatbot does not: it answers only from the documents you put in front of it. No open-web guessing, no confident invention from outside your material. Every sentence carries a small numbered citation, and clicking it scrolls the source panel to the precise passage the claim came from. You can verify an answer in two clicks instead of taking it on faith.

The polish is real, too. Drop in a forty-page policy PDF and its Studio panel will produce a twelve-minute, two-host audio discussion of it that is genuinely good — better than anything a weekend project puts out. It is free, it runs in a browser, and you can be productive in ten minutes from a cold start. As a shape for a research tool, “ground the answer in the sources I chose, and show me where” is exactly right. We are not here to argue with the shape.

The distinction almost everyone skips

NotebookLM’s headline promise is that it is grounded only in your sources. People hear that as a promise about privacy. It is not. It is a promise about what the model is allowed to know when it answers. It says nothing about who read your sources to build that grounding in the first place.

To ground an answer in your contract, the contract goes to Google. To turn your quarterly reports into that tidy audio overview, the reports go to Google. The notebook lives inside a Google account, for as long as Google keeps the product and you keep the account. The citation tells you where a claim came from. It does not tell you who else has the source.

Which sources, exactly

Here is the tell, and it is in the product’s own guidance: check before you upload an NDA, unpublished financials, or health records. That caveat is the whole story. The documents people most want a private, grounded research tool for — the client contract, the patient file, the unpublished draft, the diligence folder — are precisely the ones you are told to think twice about handing over.

A research librarian who has read exactly the stack of papers you gave them, and nothing else, is a wonderful thing to have. The question the brochure never answers is who the librarian works for, and what they do with the stack after hours.

Which is the entire architecture

chanio does the same job NotebookLM does for your text and Markdown files — your own documents turned into answers that tell you which file they came from — and disagrees about exactly one thing: where the library lives.

Your sources sit on your disk. Plain files, in a notes folder you can open without us and keep after us. Not a copy inside someone’s account — the originals, at home.

The reading runs on your own Mac. The small AI that answers from your material, and shows the lines it used, runs on hardware you already own. When a question is too hard for it, you choose: send it to a bigger AI now, as a short write-up you can read and edit first, or keep it for tonight’s one batch. Either way it goes with your own Claude key, and it carries the question and a few passages — never your whole notes folder. Without a key, nothing is sent; chanio packs it for you to paste.

The notebook is portable. When a better model ships next quarter, you point it at the same folder and keep going. NotebookLM’s notebook is a Google account; you cannot carry it anywhere Google does not go.

The answer stays grounded. The sources stay home.

The honest part

Four things cut against the comfortable version of this, and you should have them.

NotebookLM’s outputs are better than ours today. Google’s Audio and Video Overviews are polished in a way a local-first tool on your own silicon does not match, and pretending otherwise would be the exact overselling we refuse to do. We are not competing on Studio polish. We are competing on who keeps the sources.

Free and instant beats “run something on your machine.” NotebookLM is ten minutes from nothing, no install, no key. chanio asks more of you up front. That friction is real — it is the price of keeping the sources on your own disk.

Grounding is a genuine safety win, and Google does it well. A small AI on your Mac is weaker than Gemini 3 at the reading itself, and it reads text and Markdown, not PDFs. chanio’s answer is the same shape — ground in your files, cite the passage, send the hard part out — but the small AI doing the reading on your Mac is not one of the biggest AIs, and we will not tell you it is.

And the fair one: for most documents — a textbook, a public report, a lecture you would happily post — none of this matters, and NotebookLM is simply the faster tool. The argument only bites when the source is one you would never hand to Google. That is a smaller set than the fear implies. It is also, almost always, the more important one.

What you own at the end

The notebook you build in NotebookLM is a set of your documents copied onto Google’s servers, useful for exactly as long as that account and that product exist. The notes folder chanio reads is files on your disk. The model that reads them is a commodity you can swap. The sources are yours, and they stay on your disk.

Everything in this stack is rented except two things: the sources, and the record of what you asked of them. Keep those at home, and the rest can be whatever is best that week.