What NotebookLM Does With Your Sources

What NotebookLM Does With Your Sources

NotebookLM builds an AI workspace around sources you select. It can retrieve material, answer questions, show citations, and transform the material into formats such as study guides or Audio Overviews. The source boundary makes inspection easier, but it does not guarantee a correct or complete answer.

The most useful mental model is a chain with several places where evidence can be lost or changed.

flowchart LR
    A["Original artifact"] --> B["Imported notebook representation"]
    B --> C["Selected source set"]
    C --> D["Retrieved passages"]
    D --> E["Generated answer or artifact"]
    E --> F["Citation and human review"]
    A -. "material may be omitted" .-> B
    C -. "relevant source may be absent" .-> D
    D -. "interpretation may fail" .-> E

A notebook is an evidence boundary

A notebook contains sources plus the questions, notes, instructions, and generated artifacts built around them. The user can select which sources chat should consider.

Google's current NotebookLM overview describes an AI research assistant that can work with text, images, audio, and generated formats. It also warns that the product can make mistakes.

That warning belongs in the mental model. "Grounded in sources" means the system has a bounded evidence pool for the response. It does not mean every answer is entailed by every relevant passage.

Importing is a transformation

The original artifact and the notebook's representation are not always identical.

Google's current source-import documentation lists supported files and explains several limits. A web URL contributes the page's text, not its nested pages, images, or embedded video. Public YouTube sources contribute captions. Google-file comments and footnotes may not be imported.

This matters when the missing material contains a definition, citation, exception, warning, or method.

Record the original source, version, import method, date, and known omissions. If a page changes, determine whether the notebook has a static copy or an auto-synced Drive source before relying on it.

Source selection changes the answer space

When several sources exist, the user can include or exclude them for a chat request.

That is powerful because it supports questions such as "Compare these two policies" or "Use only the current specification." It also means an answer can look well grounded after the user accidentally excludes the source that contains the exception.

A source pack should include clear roles. A current policy establishes the rule now. An older policy establishes history. An explanatory article helps with understanding. It should not silently replace the authority.

Retrieval chooses what the model sees

The system does not necessarily pass every word of every source into every answer. It retrieves material judged relevant to the question.

A vague question can retrieve an incomplete slice. A specific question can improve the path while still missing relevant material.

Ask what evidence would contradict the proposed answer. Search for that material directly. If the source set contains a glossary, limitations section, appendix, changelog, or dissenting position, inspect it even when the first answer does not cite it.

Citations expose passages, not conclusions

Google's current chat documentation says citations can reveal quoted material and navigate to its location in context.

Opening the passage is the start of verification.

The reviewer still has to confirm the artifact, edition, surrounding conditions, date, scope, and authority. A passage can be real while the model's broader sentence overstates it. Several individually accurate passages can also be combined into an unsupported conclusion.

Use [[How to Verify an AI Answer Against Its Citations]] for a claim-level audit.

Generated formats are alternate views

NotebookLM can turn source material into chat answers, notes, briefings, study aids, visual formats, and audio.

The current Audio Overview documentation describes several formats and an interactive mode. It also says generated audio may contain inaccuracies or glitches.

An engaging explanation may help someone enter a topic. It is still a generated representation. The learner should return to the source and later demonstrate understanding without the artifact.

Do not count the number of generated outputs as progress. Count corrected misunderstandings, verified claims, independent recall, and successful application.

Account type controls the data boundary

Before adding a source, identify the exact account and applicable terms.

Google's privacy and terms page distinguishes consumer, work, school, and cloud use. Current documentation says consumer feedback can include the interaction context, while Workspace and Education accounts receive different treatment.

The user should verify administrator settings, connected services, feedback behavior, sharing, retention, export, deletion, region, and account recovery.

A public policy describes stated handling. It does not prove that a particular organization configured the service correctly or had permission to upload the material.

Sharing can expose more than the chat view

Current public-notebook documentation warns that viewers may still reach underlying sources and artifacts even when a chat-focused view hides them.

That means the sharing decision belongs at the source level. If a viewer must not access a source, do not rely on a cleaner interface to hide it.

Public sharing, work sharing, and school sharing can also operate under different terms and controls. Test with a representative viewer account before treating a link as safe.

NotebookLM and Gemini are not one evidence path

Current Notebooks in Gemini documentation says NotebookLM responses are grounded in notebook sources, while Gemini can also use web search and other tools.

The same notebook can therefore participate in different retrieval and data contexts.

Record where the question was asked, which sources and tools were active, and which activity settings applied. Product branding is not enough to reconstruct the evidence path.

A safe first test

Start with public, low-risk material you understand well enough to evaluate.

Create a small source manifest. Ask one answerable question, one question with conflicting evidence, and one question the sources cannot answer. Open every citation. Record what the product retrieved, omitted, qualified, or invented.

The purpose of the first test is not to admire the output. It is to learn the boundary.

Continue with [[How to Build a High-Quality Source Pack for an AI Tutor]]. For a related context lesson from AI coding, see [[What Cursor AI Is and How It Uses Codebase Context]].

This explainer was developed with AI assistance from the immutable E035 transcript, current Google NotebookLM documentation, and the linked product boundary record. Dalton Anderson remains the author. Product, privacy, security, accessibility, copyright, current-source, and founder review are mandatory before publication. Publication is not authorized.