How to Prompt NotebookLM (Source-Grounded Only)

TL;DR: NotebookLM prompting runs backwards from normal prompting. The tool answers only from documents you upload, so prompts should name a source, demand a citation, and expect refusals. Google renamed it Gemini Notebook on 16 July 2026. Nineteen copy-paste prompts below, every product claim verified at Google.

Is it still called NotebookLM, and does that change your prompts?

The name changed. The behaviour did not, which is why almost nobody has noticed.

On 16 July 2026, Josh Woodward, VP of Google Labs, Gemini app and AI Studio, published the rename on Google's own blog. Google says it "remains a standalone product focused on being your premier research tool", now reaching further into the Gemini app and Search. The same post puts usage at "more than 30 million people and over 600,000 organizations".

Every old entry point still works, which is the practical answer for anyone with bookmarks:

• notebooklm.google.com returns a 301 redirect to notebook.google.com (checked 27 August 2026).
• support.google.com/notebooklm 301s to support.google.com/gemininotebook.
• Google's own support address in the help centre is still notebooklm-user-support@google.com.

So write NotebookLM if that is what your colleagues say. Nothing about a prompt breaks. But when you go looking for documentation, search the new name, because Google's help centre now says Gemini Notebook throughout, and the majority of prompting advice online was written against the old docs and never revisited.

What does source-grounded actually mean when you write a prompt?

It means the model has no library card. Google's chat documentation states it in one line: "Chat responses in Gemini Notebook only use data from your sources." The product overview puts the positive version: you "Chat with your notebook to get grounded information based on your sources with clear in-line citations for accuracy, transparency, and trust."

That single constraint invalidates a surprising amount of standard prompt engineering. In ChatGPT you are fishing in an enormous invisible corpus, so persona framing, worked examples and chain-of-thought instructions all help you steer retrieval through a vast space. Here the space is a list of files in the left-hand panel, most of them yours, all of them countable. You are not summoning knowledge. You are interrogating a shelf.

The mechanism is retrieval-augmented generation with a hand-curated index. Google's FAQ describes it directly: when your notebook holds many sources, the tool "retrieves the most relevant information based on your question, then builds a response from it". If you want the general version of that architecture, RAG vs fine-tuning vs prompting covers when each approach is the right one.

Three habits stop working immediately, and they are the three most common instructions people paste in from other tools.

The first one has a partial workaround. Configure Chat, in the chat panel, lets you set a conversational style: Default, Learning Guide, or Custom. Google's own examples of Custom are "Respond like a PhD student" and "Pretend to be a role-playing game host", plus a response length of Shorter, Default or Longer. It persists across the notebook, so it behaves like a lightweight system prompt. Set it once. Stop repeating a persona in every message.

Why does the same prompt behave differently in the Gemini app?

Because grounding is now a property of the surface, not the notebook. This is the most consequential change since the rename, and it is barely discussed anywhere.

Your notebooks now appear in three places, and Google publishes the difference between them in plain language. From the AI Mode documentation: "Gemini Notebook: Grounds responses exclusively in your notebook sources." And, in the same list, "AI Mode in Google Search & Gemini Apps: Grounds responses in your notebook sources, but may also include web search and other tools." The Gemini Apps article says the same thing in its own words.

There is a second boundary on the paid tiers. Google's chat documentation says Google AI Ultra and Pro subscribers on desktop get agentic chat that can take actions including "searching the web, running code, or creating downloadable files, charts, and images", and adds: "You can use this chat experience with or without sources." Google labels these functions experimental and in early development. If you are on a paid plan, the closed corpus is a choice you make rather than a wall the product builds for you.

Which questions can your sources answer, and which can't?

This is the skill. Everything else is decoration.

A source-grounded tool will refuse anything outside its corpus, and Google is explicit that this is designed behaviour rather than a failure: "Gemini Notebook answers questions based on the information provided in your uploaded sources. If the answer isn't in the source material, it won't provide a response." Ask for something creative, and Google's documented example of what you get back is the string "Gemini Notebook can’t answer this question".

So the useful move is to learn the rewrite. Here are three pairs. The first prompt in each pair is the one people actually type. The second is the one that works.

1. The comparison you cannot make

How do these findings compare to standard industry practice?

Nothing in the notebook defines standard industry practice, so this either refuses or produces something vague you cannot trace. Make the comparison internal:

Using only the selected sources, list every place an author
compares their own approach to an alternative, prevailing, or
"standard" approach.

For each: quote the comparison, name the source, and state what
the alternative is described as being. Where an author asserts a
comparison without describing the alternative, say so.

2. The truth question you cannot ask

Is the claim that remote work reduces productivity true?

Truth is not a property the corpus exposes. Support is. Rewrite it as a support test:

Take this claim: "[CLAIM]"

Against the selected sources only, return exactly one verdict:
SUPPORTED, PARTIALLY SUPPORTED, CONTRADICTED, or NOT ADDRESSED.

Then quote the passage that produced the verdict and name its
source. If the verdict is NOT ADDRESSED, do not offer a view.

3. The recency question you cannot ask

What is the latest research on this topic?

The notebook has no idea what is latest. It only knows what is in it. Turn recency into a property of the corpus:

Order the selected sources by the publication date each one
states about itself.

For any source that does not state a date, list it separately as
"date not stated in source" rather than estimating one.

Then tell me the date range this notebook actually covers.

That last prompt does something quietly valuable. It converts an unanswerable question into a map of what you are missing.

Before you accept a refusal, spend one retry on precision. Google's FAQ names three causes for a non-answer: a safety flag, where "The content of your source contains language that triggers a safety flag"; unclear phrasing; and the answer genuinely not being there. Its fix for the middle one is specificity, because "A specific question helps Gemini Notebook locate the most relevant pieces of information." So run this ladder once:

My previous question may have been too broad. Narrowing it.

In [SOURCE NAME], section or heading "[SECTION]", what does the
text say about [NARROW TOPIC]?

Quote the relevant passage. If that section does not address it,
say "not addressed in that section" and tell me which section of
that source comes closest.

If it still refuses after that, believe it.

How do you write a prompt that names its own sources?

Google publishes the single most useful piece of prompting advice for this tool, and it is buried in the sources documentation: "When multiple sources are selected, mentioning source names in your query helps Gemini Notebook narrow its search." Google's own worked example contrasts a bare summarise-this-source request with one that names both the topic and the document.

Naming the source is not politeness. It is retrieval steering. Use it as a default shape:

In "[EXACT SOURCE TITLE]", find every statement about [TOPIC].

Return: the statement, quoted; the section or heading it sits
under; and whether the author states it as a finding, an
assumption, or a citation of other work.

Do not draw on the other sources in this notebook.

The second lever is not a prompt at all. Google's chat documentation says "you can use the checkbox on each source to include or exclude certain sources the model should use to answer your question", and its FAQ spells out how the pieces combine: "Sources: Always used in either the entire set or the subset you select." Notes are used "Only when you specifically select it", and conversation history is "Used to generate responses."

Deselecting sources before you type is the highest-leverage action in the entire product. A five-source subset answers a narrow question far better than a fifty-source notebook does, and you do not have to hope the retrieval step guesses right.

When you have narrowed the set, say so, so the answer records its own scope:

The sources currently selected are the only ones in scope.

Before answering anything, list the titles of the sources you can
see. If that list does not match what I expect, I will fix the
selection rather than accept the answer.

Then answer: [QUESTION]

And one reusable header worth keeping in a snippet, because it converts a sloppy request into a checkable one:

Rules for this answer:
- Every claim carries an inline citation
- Anything supported by only one source is labelled "single-source"
- Where sources disagree, present both rather than resolving it
- Anything you cannot support from these sources goes in a final
  section headed "Not covered by these sources"

Now: [REQUEST]

One structural limit to plan around. Google states that "Each notebook is independent. Gemini Notebook can’t access information across multiple notebooks at the same time." There is no cross-notebook query. If two projects need to talk to each other, they need to be one notebook.

What do NotebookLM's citations actually guarantee?

More than most tools, and less than the marketing implies. Both halves matter if you are going to lean on it.

What Google documents: the tool "uses direct quotes, text, and images straight from your sources as citations to answer your questions and perform actions", you can hover a citation to read the full quoted text, and selecting one navigates you to the quote in context. That is a genuine improvement over a chatbot that names a paper it never read.

What Google also documents, in the FAQ, under the heading "I don't always get citations": "If your source content is too short, Gemini Notebook references the entire document without a cited individual text from your source."

So citation granularity varies with source length. A pasted paragraph gets cited as a whole document. A ninety-page PDF gets cited at the passage. If you are building a bibliography, check which kind you have before you trust a locator. This prompt makes the tool tell you:

For each citation in your previous answer, tell me whether it
points to a specific passage or to a whole document.

List them as: claim | source | passage-level or document-level.

Flag any claim where the only support is a document-level
reference, because I cannot verify those without reading the
whole file.

And a discipline prompt for anything you intend to publish:

Rewrite your last answer so that every sentence containing a
number, a date, a name, or a direct assertion carries its own
citation.

Delete any sentence you cannot cite. Do not replace it with a
hedged version. I would rather have four cited sentences than
ten sentences of which six are unsupported.

What gets lost between your file and the model?

The notebook does not contain your documents. It contains an import of your documents, and Google publishes exactly what the importer drops. This is prompting-relevant, because half of what looks like a model failure is a source that never arrived intact.

From Google's sources documentation, accessed 27 August 2026:

• Web URLs. "Only the text content of the given HTML webpage is scraped for use as a source. Images, embedded videos, or nested webpages are not imported. Paywalled webpages aren't supported."
• YouTube. Only public videos with captions are supported, and "Only the text transcript of the video is imported as a source." Nothing visual survives. A prompt asking what a chart in the video showed will fail, correctly.
• Google Docs and Sheets. "Gemini Notebook does not import footnotes or comments from Google files." If your review comments carry the argument, they are not in the notebook.
• Broken links. "Inactive sources will count towards source limits but will not be referenced throughout your notebook." A source you lost Drive access to still eats a slot and answers nothing.

Audit the import before you audit the answer:

For each selected source, tell me what you can actually see:

Source | Type | Roughly how much text | Does it contain tables? |
Does it contain figures or images you can describe? | Does it
have visible section headings?

If a source appears empty, truncated, or is only a transcript,
say so explicitly. I need to know what did not make it in.

How many sources can one notebook hold?

Google publishes the caps, and marks them "Subject to Change". As of 27 August 2026:

Per source, the ceiling is 500,000 words or 200MB for local uploads, and Google states there is no page limit. Imports fail on three things: exceeding the word limit, exceeding the file size, or a copy-protected PDF.

The number that shapes behaviour is not the source cap. It is fifty chats a day on the free tier. Iterating on a vague prompt costs the same as running a good one, so drafting the prompt properly before you spend a query is not fussiness, it is budgeting.

How do you prompt the Studio formats?

Most people only ever prompt the chat box. The Studio panel produces Notes, Audio Overviews, Video Overviews, Mind Maps, Reports, Data Tables, Flashcards, Quizzes, Slide Decks and Infographics, and Google documents a free-text prompt box on most of them. Google's Mind Maps article documents no prompt box, so treat that format as unsteerable until Google publishes otherwise.

A useful habit: Google notes you can select "View custom prompt" in the three-dot menu next to a generated artefact to see the prompt that produced it. That is how you reverse-engineer a good default into a saved template.

Reports. You can pick FAQ, study guide, briefing document, an AI-suggested type, or write your own brief.

Write a decision memo for someone with ten minutes and no
background in this material.

1. What these sources establish, cited
2. Where they disagree, with both sides quoted
3. What a reader would need that is not here
4. The single decision this evidence is strong enough to support
5. The single decision it is not

End with one line stating how many sources this draws on.

Data Tables. Google says you can "describe the rows and columns for a more focused, precise table", and that exporting to Sheets puts the data in the first tab and citations in the second. That second tab is the reason to use this format over asking for a table in chat.

Rows: one per source.
Columns: Source | Year stated | Method | Population or dataset |
Headline finding | Stated limitation.

Write "not stated" in any cell the source does not report. Do not
infer, estimate, or carry a value across from another source.

Slide Decks. Formats are Detailed Deck or Presenter Slides, with short, default or long length.

Presenter slides for a 15-minute internal review.

One slide per contested question, not one slide per source.
Each slide: the question, what the sources say, and where they
split. Put figures on the slide only if a source states them.

Infographics. Google's example prompt is "Use a blue color theme and highlight the 3 key stats." Detail runs Concise, Standard or Detailed (beta), with orientation and visual style options.

Concise, landscape, professional style.

Show only figures that appear verbatim in the sources, with the
source named beside each one. No invented percentages, no
rounded-up approximations, no illustrative numbers.

Flashcards and Quizzes. Difficulty is easy, medium or hard, and Google says the prompt box lets you "add a high-level outline, or guide the audience, style, and focus."

Hard difficulty. Cards should test whether I can distinguish
between the positions different sources take, not whether I can
recall a definition.

Prefer questions of the form "which source argues X, and on what
evidence" over "what is X".

Audio Overviews. Formats are Deep Dive, The Brief, The Critique and The Debate, with a prompt box for focus and expertise level.

Audience already knows the basics and does not need definitions.

Cover only [TOPIC A] and [TOPIC B]. Skip background entirely.
Where the sources disagree, say so out loud and name both.

What source-grounded prompting still can't do

Worth being blunt, because the gap is where people lose time.

It is not a drafting tool. Creative rewriting gets refused by design. Write in ChatGPT or Claude, then bring the draft back here for a grounded fact-check pass. The handoff is covered in how to summarize 50 papers without losing citations.

It cannot compare your sources to anything outside them. If you want a benchmark, the benchmark has to become a source.

It cannot see what the importer dropped, which means it cannot warn you that the argument lived in a footnote.

And it cannot reason across notebooks. One project, one notebook, or split your question.

Where a prompt library helps here, and where it doesn't

Straight answer: we make prompts, not research tools. Prompt Architects is not a NotebookLM alternative and does not compete with it. It has no notebook, no corpus, no citations, and it will not read your PDFs.

What it is good for is the part of this workflow that repeats. Look at the nineteen prompts above: almost none of them mention a subject. "Audit what the importer actually captured" and "verdict the claim, then quote the passage" are the same prompt whether the notebook holds twelve immunology papers or twelve customer interviews. Only the bracketed slots change. That is exactly the case for a prompt template with named variables rather than a scratch file you keep losing. Our free plan includes 5 prompt enhancements per day, forever, per our FAQ; paid pricing is on the pricing page.

Two habits matter more than any tool. Version a prompt when you change it, so a rerun in three months is comparable to the original. And store the source list alongside the prompt, because a source-grounded prompt without its corpus is not reproducible. That second one is the difference between a note and a method.

If you want the prompts themselves rather than the reasoning behind them, the companion piece has twenty-eight, grouped by research job: NotebookLM prompt templates for source-based research.

Sources. Every product claim above was verified against Google-owned pages on 27 August 2026: the rename announcement on blog.google (published 16 July 2026), and the Gemini Notebook help centre pages for chat, the product overview, frequently asked questions, sources and import limits, notebooks and Studio outputs, plan limits, Audio Overviews, Video Overviews, Slide Decks, Infographics, Flashcards and Quizzes, notebooks in Gemini Apps and notebooks in AI Mode in Search. Google marks its usage limits "Subject to Change", so re-check a number before you quote it.