NotebookLM, in depth

NotebookLM answers questions using only the sources you give it, with citations back to the exact passage. That one constraint is the whole product, and it is genuinely valuable. It is also why it could not do the job we first asked of it.

What it is

A research notebook that refuses to guess

You give it a fixed set of sources — documents, pages, uploads — and it answers questions from those, citing which source produced each claim. It will also generate summaries, study material and audio overviews of the same set.

The discipline is the feature. A general assistant will happily fill a gap from general knowledge; this one stays inside the material you supplied, which makes it checkable in a way that ordinary chat is not. Google has been folding it into the wider Gemini product line and it now also appears under the name Gemini Notebook, so expect the branding to keep moving.

The question we actually asked

"Can this be our shared memory? Drive is chaotic."

That was the real starting point, and it is the same question most organisations arrive at: our files are a mess, could an AI layer sit on top and make sense of them. It is an appealing idea because it promises to skip the boring part.

We wrote the answer down, and it was no. The reasoning is the useful bit:

A tool that grounds its answers in your sources inherits the quality of those sources completely. If nobody can say which version of a document is current, the citation tells you where the claim came from but not whether it is true. Grounding is not the same as governance.

Where it landed

A rendered view, never operating memory

Rather than drop it, we narrowed it. The rule we settled on is simple enough to state in a sentence: the notebook is a rendered view of the source of truth, never the source of truth itself. If the notebook and the canonical record disagree, the record wins and the notebook gets rebuilt.

That reframing made it genuinely useful. Pointed at a curated, clean source set, it is very good at exactly what the team likes it for: grounded question and answer over a defined body of material, and audio overviews for the times when reading is not practical. Pointed at everything, it is a liability.

How we keep it fed

We maintain a notebook holding a curated export of our knowledge base, refreshed from the canonical record rather than edited in place. Two refresh behaviours exist and the difference matters more than it sounds: one creates a fresh notebook each time, the other replaces the contents of the existing one. Choosing the wrong one silently orphans whatever people had already been working against.

That is a mundane operational detail, and it is precisely the kind of thing that decides whether an AI tool is trusted six months in.

Getting data in and out

Automation is the weak point

We initially recorded that there was no programmatic way in at all. On checking properly, that was wrong, and the corrected picture is worth knowing before anyone plans around it:

• Manual use — fully supported, and for most teams entirely sufficient.
• An official programmatic route — exists, but gated behind the enterprise tier of Google's AI licensing rather than being generally available.
• Community integrations — several exist, and we ruled them out. They work by driving the product as if they were a logged-in browser, which means borrowed credentials, undocumented behaviour that can change without notice, and terms-of-service exposure. Not something to put company infrastructure on.

Ruling that third option out cost us convenience and was still obviously correct. A workflow that depends on an unofficial integration is a workflow with a hidden expiry date.

Where it stops

What we would not use it for

• As a system of record. It is a reader, not a registry. Nothing should exist only inside it.
• As a fix for disorganised files. It will make disorganised files feel more accessible without making them more correct — which can delay the cleanup rather than replace it.
• As an automation backbone. Unless you hold the enterprise licence, plan around manual curation.
• For anything sensitive without a decision about what goes in. The source set is the exposure surface. That is a governance question, not a product question.

Used inside those lines, it earns its place. It is one of the clearest examples we have of a tool that got more useful once we asked it to do less.