How to Prompt NotebookLM for Research (2026): Formulas + Examples

NotebookLM is Google's source-grounded research tool, built on Gemini 3. It answers only from the sources you add — PDFs, Docs, URLs, YouTube, pasted text, audio — and attaches an inline citation to the exact passage behind every claim. That grounding changes how you prompt it: a great NotebookLM research prompt is about half choosing and curating good sources and half writing a clear, grounded ask that demands citations and admits what the sources don't cover.

This guide gives you the formula, then 10 complete prompts you can paste in and adapt. For a wider set, see the best NotebookLM prompts roundup, the dedicated research prompts collection, and the NotebookLM prompt cheat sheet for the quick reference.

Start with the sources, not the prompt

The single biggest lever on answer quality is which sources sit in the notebook. NotebookLM can only answer from what you add, so garbage in, garbage out is literal here — a clever prompt over weak sources still returns weak, if well-cited, answers.

Curate before you type a single prompt. Add the primary documents that actually contain the answers — the papers, filings, reports, transcripts, or notes — and prefer authoritative, current, first-hand sources over summaries of summaries. Give each source a clear title so citations read well. Then trim: remove anything off-topic, duplicated, or low quality before you start asking.

When a notebook holds many sources, use the source selector to focus a single query on just the relevant subset — deselect the papers that don't bear on the question so the model reads only what matters. A tight, curated set of ten strong sources beats fifty loosely related ones every time. Source quality outweighs prompt cleverness; get the corpus right first.

The NotebookLM research prompt formula

Every strong research prompt names five things. Miss one and the answer drifts — vague scope, no citations, or confident filler where the sources are silent. Use this pattern:

Read it as one line: TASK + SCOPE + FORMAT + CITATION RULE + UNCERTAINTY RULE. The last two matter most and are the ones people skip. Because NotebookLM is grounded, the citation rule costs nothing and gives you a checkable answer; the uncertainty rule turns "the sources are silent here" from a silent gap into an explicit flag you can act on.

10 example research prompts

Each prompt below is complete and paste-ready. Swap the bracketed details, and remember to select the right sources before you run it. Each note explains why the wording works.

1. Synthesis matrix across sources

Using only the selected sources, build a synthesis matrix on [research question]. Rows = each source; columns = research question, method or approach, key finding, sample or scope, and stated limitations. Fill every cell only from the sources, and put an inline citation (source name + section or page) in each cell. Where a source doesn't address a column, write "not covered" rather than inferring. After the table, give a 3-sentence synthesis of what the sources collectively show. Do not use any outside knowledge.

Why it works: A fixed matrix forces the model to read each source on the same dimensions, and the "not covered" instruction turns silence into a visible gap instead of a guessed answer.

2. Consensus vs. disagreement

Across the selected sources, tell me where they agree and where they disagree on [topic]. Give three sections: (1) Points of consensus, (2) Points of genuine disagreement — for each, name which sources take which position and cite the passage, and (3) Questions none of the sources answer. Quote or paraphrase the specific line behind each position with a citation (source + page/section). Do not resolve a disagreement the sources leave open, and do not add claims from outside the sources.

Why it works: Splitting consensus from disagreement — and demanding the source behind each side — produces an honest map rather than a smoothed-over summary that hides the conflict.

3. Cited answer to a precise question

Answer this question using only the selected sources: [precise question]. Give a direct answer first, then the supporting evidence, with an inline citation (source name + page or section) after every factual sentence. If the sources only partially answer it, say exactly what they cover and what they leave open. If they don't answer it at all, reply "The sources don't cover this" and stop. Do not fill gaps with general knowledge.

Best for: A single sharp question — leading with the direct answer and forcing a citation per sentence gives you something you can paste and defend.

4. Methodology comparison

Compare the methodologies used across the selected studies on [topic]. For each study, extract: study design, sample size and population, data collection method, analysis approach, and any stated methodological limitation. Present as a table with a citation (source + section) in every cell. Then, in 4-5 bullets, note which designs are strongest for this question and where the studies are not directly comparable (different populations, timeframes, or definitions). Only use the sources; write "not stated" where a study omits a detail.

Why it works: Pulling method details into fixed fields exposes when studies aren't actually comparable — the "not stated" rule keeps the model from inventing a sample size a paper never gave.

5. Find the gaps in the literature

Based only on the selected sources, identify the gaps in what they cover about [topic]. List: (1) questions the sources raise but don't answer, (2) populations, contexts, or timeframes none of them address, and (3) claims made in one source that no other source tests or supports. Cite the passage behind each point (source + section). Do not speculate about the wider literature beyond these sources — frame every gap as "within this source set." Rank the gaps by how central they are to [topic].

Best for: Scoping a review or a project — bounding the gaps to "within this source set" keeps the answer grounded instead of drifting into speculation about all research everywhere.

6. Build a dated timeline

From the selected sources, build a dated timeline of [event, project, or line of research]. For each entry give the date, what happened, and a citation to the source and section it comes from. Order chronologically. Distinguish dates the sources state explicitly from ones you infer, and label any inferred date as "(approx., inferred)". Where two sources give different dates for the same event, show both and cite each. If a key moment is referenced but undated in the sources, note it as "date not given."

Why it works: Separating stated dates from inferred ones, and surfacing conflicting dates side by side, keeps the timeline accurate to the sources rather than a tidy fiction.

7. Cited Briefing Doc in Studio

Generate a Briefing Doc from the selected sources on [topic] for someone who hasn't read them. Cover: the core question or purpose, the key findings, the main points of contention, and the practical takeaways. Keep every claim tied to an inline citation (source + section). Add a short "What the sources don't cover" section at the end listing open questions. Write in plain, precise language — accurate over polished — and don't include anything not supported by the sources.

Best for: Onboarding yourself or a teammate to a corpus — the Briefing Doc in Studio produces a cited overview, and the closing gaps section stops it from reading as more complete than it is.

8. Evidence-strength table

For each major claim about [topic] in the selected sources, rate how strong the supporting evidence is. Build a table: Claim | Supporting source(s) | Type of evidence (e.g. RCT, observational, expert opinion, anecdote) | Strength (Strong / Moderate / Weak) | Citation. Base the strength rating only on what the sources describe about their own methods. Flag any claim supported by just one source, or contradicted by another. Do not rate a claim the sources don't actually make. End with the 3 best-supported and 3 weakest claims.

Why it works: Rating evidence by the source's own described method — not by how confident it sounds — tells you which claims to trust and which to double-check before you cite them.

9. Appraise a specific claim against the sources

Appraise this claim against the selected sources only: "[paste the claim]". State whether the sources Support, Partly support, Contradict, or Don't address it. Quote the specific passages for and against, each with a citation (source + page/section). If support is partial, separate the part the sources back from the part they don't. If no source speaks to it, say "Not addressed by the sources" — do not judge it from outside knowledge. Finish with a one-line, evidence-based verdict.

Best for: Pressure-testing a claim before you rely on it — a fixed verdict scale plus required quotes makes the model point at evidence instead of asserting.

10. Audio Overview Deep Dive focus prompt

Generate a Deep Dive Audio Overview from the selected sources, focused on [specific angle, e.g. "the methodological disagreements between the two main studies"]. Prioritize the evidence and points of contention over background, keep it grounded in the sources, and don't introduce claims the sources don't make. Spend most of the time on [the sub-topic that matters most to me]. Where the sources disagree, present both sides fairly rather than settling it.

Why it works: The Audio Overview focus box accepts an instruction like this — naming the angle and the sub-topic to dwell on turns a generic two-host summary into a targeted briefing on exactly what you need.

Get better citations

NotebookLM cites by default, but you get cleaner, more useful citations by asking for them explicitly. Add a line like "cite the source name and page or section for every claim" to any prompt — that writes the reference into the text, which matters when you paste the answer elsewhere and lose the clickable inline links.

For documents and PDFs, the inline citation jumps to the exact passage, so ask for the page or section rather than just the source name. For YouTube sources it points to the spot in the transcript. The verification step is non-negotiable: click the load-bearing citations and confirm the passage actually says what the answer claims. The grounding makes fabricated references rare, but the model can still paraphrase loosely or cite a tangential line — the citation exists to be opened, not counted.

Mistakes to avoid

Most weak NotebookLM results trace back to one of these:

• Adding low-quality sources. A notebook full of blog posts and summaries returns blog-quality answers. Load primary, authoritative documents.
• Vague asks. "Tell me about this" wastes the grounding. Name the task, the format, and the citation rule.
• Asking outside the sources. Questions the corpus can't answer produce thin results or refusals — either add the source or accept it's out of scope.
• Not deselecting irrelevant sources. With a big notebook, an unfocused query reads everything. Use the source selector to narrow to what bears on the question.
• Trusting an answer without clicking the citations. Grounded is not infallible. Open the load-bearing links and read the passage yourself.
• One giant multi-part ask. Ten sub-questions in one prompt get shallow, tangled answers. Break the research into focused prompts and build on each.

Once a prompt lands close, refine it the way you'd brief a research assistant: tighten the scope, name the exact sources, and insist it flag what they don't cover. For the wider set, browse the best NotebookLM prompts roundup and the student prompts collection, and keep the cheat sheet open while you write.