I uploaded six months of board decks into NotebookLM and asked it what we’d been avoiding. The answer was uncomfortable. Not because it told me anything I didn’t already know, but because it pulled together threads I’d written across different months and never read back-to-back. Concerns I’d noted in March about a particular client concentration. A note in May about the same client coming up in a different context. A risk paragraph from August that, read alongside the other two, suddenly looked less like a paragraph and more like a pattern.

Most AI tools are bad at strategy because they don’t know our business. NotebookLM is useful precisely because it doesn’t pretend to. It can only work from what we give it, and that turns out to be the whole point. Most strategy work, in my experience, is the patient business of noticing what’s already in your own evidence and acting on it before someone else has to point it out. New insight matters, but it’s a smaller share of the job than people think. Most of what counts is paying attention to what you’ve already written, and most of us are bad at it, because nobody has time to re-read half a year of their own commentary on a Tuesday.

(A quick note on names: Google renamed NotebookLM to Gemini Notebook in July 2026 — same standalone product, deeper ties to the rest of the Gemini ecosystem. Everything in this piece applies to both, and I’ll use the name you still know it by.)

Why general AI is the wrong tool for this job

If you've tried using ChatGPT or Claude for strategic thinking — really thinking, not drafting an email — you'll know the disappointment. You ask a sharp question about your market positioning and you get a plausible-sounding answer with no relationship to your actual business. The model has never met your customers. It doesn't know what your last three board meetings argued about. It's filling the gaps with sector averages and confident generalisations, and the result reads as competent only because it isn't grounded in anything specific enough to be wrong.

This is why most experiments with AI as a strategy partner stall. We’re asking a model to invent facts about a business it has no knowledge of, and then evaluating the output as if it were analysis.

NotebookLM inverts the problem. You upload your sources first, and the model reasons only over what you’ve given it. A September 2025 arxiv paper (“Not Wrong, But Untrue”) found NotebookLM hallucinated on about 13% of document-grounded queries, against roughly 40% for ChatGPT and Gemini — and the residual errors were interpretive, not invented numbers. That’s the kind of error you can spot and correct, rather than the kind that poisons your thinking without you noticing.

That limitation is also the point. NotebookLM can’t tell you about a market trend it hasn’t read about, and it can’t invent a competitor your sources don’t mention. What it does well is the bit that’s actually rate-limiting for most teams: reading everything you’ve already written, and answering specific questions about it.

How to set it up for strategy work

The workflow that's worked for me is unglamorous. There's no clever prompt to share.
  1. Build the corpus deliberately. Upload the documents that actually carry the thinking: board decks, the last few quarterly reviews, customer research, competitive notes, the strategy doc itself if you've got one, financial summaries. Skip the operational noise. The free tier gives you 50 sources per notebook and the paid tiers run into the hundreds, but for most strategic work you'll need somewhere between 10 and 30 well-chosen documents, not a document dump.
  2. Lead with the question you've been avoiding. "What are we treating as a fact that this evidence doesn't actually support?" is a better opening prompt than "summarise our position." The first uses what's distinctive about the tool. The second wastes it.
  3. Use it to find contradictions, not consensus. Ask where your different documents disagree with each other. Where the optimistic Q1 narrative is in tension with the Q3 customer feedback. Where the strategy doc says one thing and the financial reality says another. This is genuinely hard for humans to do unaided, because we read documents sequentially and forget the earlier ones.
  4. Demand citations and check them. Every answer comes with citation chips that link back to the source passage. Click them. Especially on anything that surprises you. NotebookLM's interpretive errors tend to come from softening a hedged statement into a confident one, and you'll catch most of those by reading the actual passage it pulled from.
  5. Listen to the audio overview before a big meeting. Two AI hosts talking through your sources for ten or fifteen minutes sounds like a gimmick, and is, in fact, the single feature I'd recommend trying first. I can't fully explain why this works as well as it does — something about hearing your own material discussed aloud, by voices that aren't yours, makes patterns audible that you'd never notice reading. I've used it before board meetings to absorb material I'd seen too many times to read properly.

The prompts that actually do work

Once the corpus is in, the difference between getting a generic summary and getting something useful is the question you ask. A few that have worked for me, grouped by job:
  • For contradictions: "Where do these documents disagree with each other? Specifically, where is the optimism in the strategy doc not supported by the customer research?"
  • For surfaced assumptions: "What are we treating as settled fact that we've never actually tested?"
  • For board prep: "Argue the strongest case against the position taken in the Q3 board deck, using only evidence from the other documents I've uploaded. Cite the passages."
  • For theme synthesis: "Across the customer interviews, what do customers keep saying that the strategy doc doesn't address?"
  • For red-teaming a decision: "If we made the call described in [document X], what's the case in the other documents that this is the wrong call?"
The pattern is to ask questions the documents are uniquely positioned to answer, and to force the model to cite where it found things. Don't ask it for opinions it hasn't earned.

What it's good for, and what it isn't

It's strongest when the work is essentially editorial — distilling, comparing, surfacing, stress-testing. Preparing for a board meeting from quarterly material. Synthesising customer interviews into themes. Pressure-testing the assumptions inside a strategy doc by asking NotebookLM to argue against them using only the supporting evidence you've uploaded. Reading six months of your own commentary in one sitting without actually reading any of it.

It falls down on anything requiring real-time information — current market data, today’s competitor moves, anything that isn’t in the corpus you’ve uploaded. It’s not the tool for creative ideation from a blank page. And while it wasn’t built for quantitative modelling, the July 2026 update is starting to change that: every notebook now gets a secure cloud computer that can write and run code against your sources, rolling out from the Ultra tiers down to Pro. Use NotebookLM when the answer has to come from your documents. Use Claude or ChatGPT when you need anything beyond them.

I’m not sure whether the gap between document-grounded and general models will narrow as the others get better at grounding. For now they’re different tools for different jobs, and the people I’ve seen get most value from AI tend to use all three. The ones who pick a favourite and force everything through it are doing worse work without realising it.

The blind spot most teams have

The reason this matters beyond a tool review is that most organisations have a blind spot the size of their own filing cabinet. The documents aren't bad; nobody reads them all at once, and nobody is expected to. The CEO has seen the board decks. The finance director has seen the financial reports. The marketing director has seen the customer research. But almost nobody, the CEO included, has read all of it together with a sharp question in mind. Which means the most expensive thinking inside your business is sitting there, unread, on a shared drive someone half-tidied in 2024.

NotebookLM doesn’t do the thinking. It just makes it cheap to sit down with everything you’ve already written and ask it something specific. Until quite recently, that was an afternoon’s work most teams skipped.

If you’ve already explored NotebookLM as a way to turn documents into slide decks, this is the deeper use case. Slides are output. Asking sharp questions of your own documents is input, and that’s where the real leverage sits.

The uncomfortable thing about my own board-deck experiment wasn’t that NotebookLM was clever. It was that the answer had been in my own writing for months and I hadn’t seen it. It just made me read what I’d already written.

Quick answers

Is NotebookLM better than Claude or ChatGPT for strategy work?

For different things. NotebookLM is better when you want answers grounded strictly in your own documents with citations you can verify, and you’d rather have a model that refuses to invent than one that obliges. The general-purpose assistants are better when you want broader reasoning, real-time information, or ideation from a blank page. Most strategic workflows benefit from using both, with the rough rule that NotebookLM is for interrogating what you already have, and the other two are for figuring out what to do about it.

What kinds of documents are worth uploading? The ones that actually carry the thinking — board decks, strategy documents, customer research, anything you’d reluctantly re-read before a big decision. Twenty well-chosen documents beats two hundred indiscriminate ones, and operational noise actively dilutes the signal.