Build a Notebook That Answers From Your Sources Only

A notebook that answers from your sources only is the entire value proposition. If the model can reach past your PDFs, every answer becomes a claim you have to verify from scratch, and you have bought yourself a slower chatbot.

Gemini Notebook, which Google renamed from NotebookLM on 16 July 2026, gets closer to that guarantee than most tools. It does not get all the way, and the gaps are documented in Google's own help centre if you read the pages against each other.

The grounding promise lives in a Tip box, not in a spec

The strongest statement Google makes appears in Use chat in Gemini Notebook, formatted as a tip:

Chat responses in Gemini Notebook only use data from your sources. If you explicitly ask the model to do something more creative, like "rewrite the end of my short story," you may receive, "Gemini Notebook can't answer this question" as a chat response.

The FAQ restates it from the failure side, under the heading "My question wasn't answered": "Information not in sources: 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."

Both sentences are about retrieval, not about generation. Nothing in the help centre says the model's own trained knowledge is switched off while it writes the sentence around a retrieved passage. Treat the guarantee as scoped to where evidence comes from, and not to how the prose gets built.

Three things reach the model, and your sources are only one of them

The FAQ carries a small section titled "How sources, chat context, and notes are used in prompts". It is the most useful paragraph in the documentation and it is easy to miss:

Notes: Only when you specifically select it. Sources: Always used in either the entire set or the subset you select. Conversation history: Used to generate responses.

That third line is the one that catches people. Conversation history is an input. Anything you typed into the chat box, including a half-remembered figure you were asking the model to check, is now in the context and can come back to you phrased as an answer. It carries no citation, because it never was a source.

Practical consequence: do not paste facts into chat. Ask questions. If you need the model to consider a claim from outside the corpus, add the claim as a source so it arrives with provenance attached, or start a new chat. The three-dot menu in the chat panel has Delete Chat History, and the docs note that "Chat history is retained and kept private to you."

The source-level control is the checkbox on each source in the Sources panel: "you can use the checkbox on each source to include or exclude certain sources the model should use to answer your question." Subsetting is the mechanism you actually have. Use it as the default, not as a special case.

Google's Pro and Ultra chat documentation contradicts the grounding tip

This is a real conflict between two sections of the same help page, and it is worth quoting rather than summarising. Further down the same article that promises chat "only use data from your sources":

As a Google AI Ultra or Pro subscriber, you can access agentic capabilities and more advanced reasoning in chat for Gemini Notebook to take actions on your behalf, such as searching the web, running code, or creating downloadable files, charts, and images. You can use this chat experience with or without sources.

Searching the web from inside the chat panel is the opposite of source-grounded, and "with or without sources" says so outright. Google does not reconcile the two statements, does not say whether the agentic path is opt-in per query, and does not document a setting to disable it. Its own warning on that section is unusually direct: "These new functions are experimental and in early development. Your supervision is important to help prevent unintended and potentially harmful actions."

If you are on a free or Plus account, the documented agentic behaviour does not apply to you, and that is genuinely an argument for staying on the free tier for grounded work. If you are on Pro or Ultra, assume any chat answer might have web provenance until you have clicked its citations and confirmed they point into your own sources.

Converting a note to a source launders model output into evidence

The notes documentation describes two paths from generated text to source material. For one note: "Open the note. Select Convert to source." For all of them: "In the 'Notes' section within the 'Studio' panel, click More. Select Convert all notes to source."

Both are useful. Both are dangerous in a notebook whose whole purpose is that everything in it is real.

A note created with Save to Note is a chat response. Convert it to a source and it joins the corpus that the docs say is "Always used". From that point the model can cite it, and the citation will look exactly like a citation into a peer-reviewed PDF. You have created a document whose only provenance is an earlier generation, and buried the fact.

The rule that keeps a notebook honest: convert your own written notes if you like, and never convert a saved chat response. If you need a synthesis to persist, export it out with Export to Docs and bring it back as a clearly named Drive source such as SYNTHESIS not evidence 2026-07-31. Note the documented one-way door on that export: "Any modifications made within the newly created Docs or Sheets file will not synchronize with your original Gemini Notebook note."

Import from the web only through a step that leaves a trail

The Sources panel can pull in material you never chose. The sources documentation describes two discovery modes. Fast Research runs a query where you "Choose Web or Drive" and imports what comes back. Deep Research goes further: "Deep Research is an agentic feature that can automatically browse up to hundreds of websites on your behalf, think through its findings, and create insightful multi-page reports in minutes."

Deep Research is capped at 10/month on the free tier, 20/day on Pro. Its output is a generated report plus "a list of all relevant sources (cited and not cited)". Importing the report puts a synthesis into your corpus alongside your papers, and after that the notebook cannot tell you which is which.

For a grounded notebook, set Fast Research to Drive rather than Web when you use it at all, import the underlying sources from Deep Research rather than the report, and rename anything web-derived at import so its origin is visible in the citation. Web URL import also has a documented blind spot: "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."

Prove the lock works with a question you know it cannot answer

Before you trust a notebook, run a negative control. Ask it something plainly outside the corpus but adjacent enough that a chatty model would want to help, and see what comes back.

Using only the sources in this notebook, state the 2024 replication rate reported for this literature. If no source in this notebook reports it, reply with exactly: NOT IN SOURCES.

A correct notebook returns the refusal string, or Google's own "Gemini Notebook can't answer this question". A notebook that produces a plausible number with no citation, or with a citation you click through to a passage that says nothing of the kind, has failed the test. Do not troubleshoot that with a better prompt. Check whether an agentic chat tier is in play, whether a converted note is sitting in the source list, and whether the answer came from something you typed earlier in the same conversation.

Run the control again after every batch of imports. The corpus changes and the guarantee changes with it.

The gaps you are working around, stated plainly

Google does not publish the model version behind the notebook. The Upgrade page's "Gemini Model" row reads only "Access to Gemini models", "Higher access to Gemini models" and "Highest access to Gemini models". It does not publish how many passages retrieval pulls per query, how citation spans are selected, or whether the grounding tip is enforced by architecture or by prompting. Any of those would change how much a clean negative control is worth, and none of them are documented.

Data handling is documented, and it is better than the silence elsewhere: "Your data is protected, and is not used to train Gemini Notebook unless you provide feedback. When you share feedback to help us improve, we may review the full context of that interaction, including your queries, uploads, and the model's responses." Workspace and Workspace for Education accounts get a stronger line: uploads, queries and responses "will not be reviewed by human reviewers even when you provide thumbs up or down feedback, and will not be used to train AI models." So the thumbs-down button is not a free action on an unpublished manuscript in a consumer account.

None of this makes the tool unusable. It makes the negative control non-optional, and it makes provenance a thing you maintain by hand rather than something the product guarantees for you.