NotebookLM is different from the chat window you already use, but not in the way most write-ups describe. It answers only from the sources you give it, and that single constraint moves the work: the clever part stops being the wording of the prompt and becomes the choice of what goes into the notebook.
For a seller that is good news, because you already sit on the raw material a grounded model needs. If you want the general-purpose tool pages, Claude prompts for sales and Perplexity prompts for sales cover those, and Gemini Deep Research, Copilot Researcher and Gemini Gems cover their own named features. This page stays on the source-grounded notebook.
NotebookLM answers from your sources and nothing else, which changes the prompt
A general model fills a gap. Ask what a company's procurement process looks like and it produces something plausible from the shape of other companies. A grounded model tells you the sources do not say. That refusal is the feature.
So write prompts that treat refusal as valid. Instead of "summarise their priorities," ask "list the priorities stated in these sources, and say which are quoted directly and which I am inferring." Instead of "what are the risks," ask "which risks appear in more than one source, and which in only one." The prompts that work here let the model come back shorter than you wanted.
Prompts for the notebook worth building: what to put in and what to leave out
Bad sources produce confident nonsense with citations attached, which is worse than nonsense without them. Build the notebook deliberately, one deal or account each.
In: call transcripts and your own notes, the RFP or security questionnaire, the current contract and any amendment, the buyer's last two filings, their published pricing. Out: internal battlecards, marketing copy, anything that states the conclusion you want the model to reach, and anything you would not want quoted back at you. Then run: "Inventory these sources. For each one, say what it covers, its date, and one question it cannot answer." That tells you what to add before you ask anything that matters. The privacy rules for what not to paste apply here too.
Account research from documents you already have
Most account research prompts go out to the open web. This one runs the other direction: twelve documents already in hand and no time to read them. The account research brief is the output, built from your own sources rather than a search.
Try: "Across these sources, build a one-page brief with four sections: what this organisation says it is trying to do this year, who is named as accountable for it, what they say constrains them, and what has changed since the oldest source here. Cite every line." The last clause is what makes it worth reading. Uncited lines are the ones you go and verify.
Prompts that interrogate a long document for the three things that matter to your deal
Nobody reads a 60-page annual report before a call. You need three things out of it, and a grounded model finds them without inventing a fourth.
Shape the prompt around your deal rather than the document: "I sell [what you sell]. Find every passage in this filing that touches [the problem you solve], the budget line it would sit under, and any stated initiative it would compete with for funding. Quote each passage and give the page." Then the follow-up most people skip: "What in this document argues against buying what I just described?" That is the answer you take into the room, because your buyer's finance partner has already read it.
Turning a pile of call transcripts into one deal narrative
Six calls across four months, three with different people, and your CRM has the shape of the deal but not the story. Ask for the narrative rather than the summary.
"These are transcripts from one deal, in order. Build a timeline of what each person said they needed, and flag every point where a requirement changed or a new name appeared without explanation." Changed requirements and unexplained new names are the two things that wreck forecast accuracy, and both are visible in transcripts and invisible in a pipeline report. Call notes prompts do the per-call pass; mining transcripts for patterns covers the across-deals version.
Prompts for the questions a grounded model can answer and a general one cannot
This is the part worth memorising, because it tells you which tool to open. A grounded model is better at "what do these documents say," and worse at "what should I do about it."
Good here: does their stated policy allow what we are proposing; where do these two contracts disagree; which of my notes contradict the questionnaire they returned; what did they ask for in March that nobody has mentioned since. Bad here: write me the follow-up email, what is the market doing, how should I price this. Those belong in a general model, and the full prompt set by role and stage covers them.
When a citation is not evidence
Here is the honest limit. A citation proves a sentence exists in your sources. It does not prove the sentence is true, current, or the whole picture.
Three failure modes. A perfectly cited quote from a filing two years old, describing a strategy they abandoned. A quote from your own notes, which were wrong when you typed them. A correct quote from a passage whose next sentence reverses it. The habit that fixes all three: whenever a citation is about to change what you say to a buyer, open the source and read the paragraph around it. Citations make verification fast, not unnecessary.
What a source-grounded model still cannot tell you about the deal
It cannot tell you who actually decides, unless somebody wrote that down. It cannot tell you whether your champion will spend political capital on you. It cannot tell you why the deal went quiet in week six. None of that is in the documents, because nobody put it there.
So the notebook is a preparation tool, not a diagnosis tool. It gets you to the call knowing what the sources say, so the call can be spent on what only a person will tell you. Prompts for that half sit in Meeting Prep & Discovery and Deal Strategy & Stakeholder Management; the research prompts live in Account Research & Buyer Intelligence, and tool-level workflows in AI-Era / Emerging Sales Work and Sales Operations, CRM & Productivity.
Every prompt above is a starting point to edit before you run it. Start with a free prompt, then Browse the library → or read the full taxonomy. Pro is $12 per month or $89 per year, and lifetime access is $149. See pricing.
Frequently asked questions
What are good NotebookLM prompts for sales?
The ones that ask what your sources say rather than what the model thinks. Useful shapes: inventory these sources and name what each cannot answer; build a cited one-page account brief; find every passage in this filing that touches the problem I solve; show me where two of these documents disagree. Ask for citations on every line.
How do I use NotebookLM for account research?
Build one notebook per account from documents you already hold: transcripts, your notes, the current contract, recent filings, published pricing. Then ask for a cited brief covering what they say they are trying to do, who is accountable, what constrains them, and what has changed since your oldest source. Inventory the notebook before you trust the brief.
Is NotebookLM better than ChatGPT for sales?
Neither is better; they answer different questions. A grounded notebook is stronger on "what do these documents say" and tells you when the sources are silent. A general model is stronger on "what should I do about it": drafting, framing, roleplay. Most reps use both: the notebook for preparation, the general model for the writing that follows.
What should I put in a NotebookLM notebook for a deal?
Call transcripts and your own notes, the RFP or security questionnaire, the current contract and amendments, recent filings, published pricing, and competitor material your champion sent you. Leave out internal battlecards, marketing copy, anything asserting the conclusion you want, and anything you would not want quoted back.
Can I trust NotebookLM's citations?
Trust them to show where a sentence came from, not that it is true. A perfect citation can point at a two-year-old strategy, at notes you got wrong, or at a sentence the next line contradicts. The working rule: whenever a citation would change what you say to a buyer, open the source and read the paragraph around it.