NotebookLM for literature review: a source-grounded workflow

NotebookLM (currently branded Gemini Notebook in Google’s help pages) can organize a bounded source set and draft answers with citations. Treat those citations as pointers to inspect, not proof that an interpretation is correct. This walkthrough produces a short, auditable evidence table; the optional MCP connection is a separate, third-party integration.

Define the question and source set

Start with one review question narrow enough to answer from a small group of documents. For example: “How do the included protocols define a positive control, and where do they differ?” Choose documents you are permitted to process in a cloud service. Keep an external source manifest with each document’s title, version or publication date, stable URL or identifier, and the question it informs.

Guided setup (no example papers are supplied). Select two methods papers or protocols you already have lawful permission to upload; use public, non-sensitive material and do not upload restricted or unpublished work without institutional approval. Record each title, version/date, and stable link in your manifest. Live use requires a signed-in Google Account in a supported location and age range; access to managed work or school accounts can depend on administrator settings. On a computer, open Gemini Notebook, create a notebook, choose Add sources, then upload supported local files or select permitted Drive files; wait for import to finish and confirm both titles appear in the source panel. The interface can change, so consult Google’s current source-import instructions if labels differ. Select only the two intended sources before asking your comparison question. If either source fails to import or you cannot establish permission, stop and choose another permitted document rather than proceeding with an incomplete pair.

1. Create a notebook named for that question. Add only the selected materials and wait until each has finished importing.
2. Ask a comparison question that requests one row per source, a short quotation or passage pointer, and an “unclear or absent” value when the source does not answer it. Avoid prompting the system to fill gaps from general knowledge.
3. Open each citation in the original document. Check that the passage supports the statement and that qualifiers, units, population, assay conditions, and version have not been lost.
4. Record accepted evidence in a table or notes document. Separate what the source states from your interpretation; preserve disagreements instead of averaging them away.

| Claim | Source title / version | Passage locator | Source wording or faithful paraphrase | My interpretation | Status |
|---|---|---|---|---|---|
| [claim] | [title, date/version] | [section, page, or citation opened in source] | [what the source says; include relevant qualifier] | [what you infer, separately] | [supported / contradicted / unresolved] |
| Illustrative only: protocol defines a positive control | Example Protocol A, v1 | Methods, “Controls” paragraph | “A known positive sample was included in each run.” | This describes a run-level control; it does not specify acceptance criteria. | unresolved pending direct source check |

NotebookLM can produce missing, incomplete, or mistaken citations, and Google notes that it may not always provide citations. If a source pointer cannot be located or does not support the claim, revise or remove the claim and consult the underlying source directly. This workflow organizes reading; it does not establish biological validity or replace a formal systematic-review method.

Ask questions that keep sources visible

Use a prompt that specifies the comparison unit and what to do when evidence is missing. Substitute your question and actual source labels:

Using only the sources in this notebook, compare [method or claim]
across the included documents. For every row, give the source title,
version/date, a short answer, and the citation or passage pointer.
Write “not stated in this source” when appropriate. Keep conflicting
statements in separate rows. Do not infer a biological outcome.

Check whether the answer really stays within the source set. For a methods comparison, do not collapse distinct cell types, sample preparation steps, instruments, or endpoints into one generic recommendation. Use the original paper or protocol when context is needed.

Optional: connect an MCP client

The MCP add-on described here is maintained outside Google. It is not an official NotebookLM API: the project documentation warns that it uses undocumented internal interfaces that can change and may not align with Google’s terms. Decide whether that dependency is acceptable before installing it. You can use the source-review workflow above without MCP.

The project currently documents a unified Python distribution named notebooklm-mcp-cli, which supplies the nlm CLI and notebooklm-mcp server executable. This optional route requires pipx installed and available on your PATH. Check with pipx --version; if missing, follow the official pipx installation instructions for your operating system, then open a new terminal if you changed PATH. On Ubuntu 23.04 or newer, install with sudo apt install pipx and run pipx ensurepath. If the version check still fails, stop and resolve pipx/PATH before installing the optional client. Install the client in the isolated pipx environment:

pipx install notebooklm-mcp-cli
nlm --help

These commands install and inspect the tool; they do not prove the service is authenticated or usable with your chosen client. Project instructions use nlm login for sign-in and nlm setup for client registration. Follow the current prompts and the documentation for your exact client rather than copying an old menu path or config filename. Quit and restart the client after changing its MCP configuration, then verify that it reports the server connected and exposes its tools.

If your client requires manual configuration, its local-server entry must use the executable and arguments specified by the current project instructions. Do not paste account cookies, tokens, or credentials into a prompt, source file, or shared repository. The project’s authentication flow handles sensitive account session material; review its credential storage and permissions before proceeding.

Use a conservative privacy and verification policy

Gemini Notebook processes notebook content as a cloud service. Google’s current help says personal-account data is not used to train Gemini Notebook unless feedback is submitted; when feedback is sent, Google may review the interaction context, including uploads, queries, and responses. For Workspace and Workspace for Education accounts, Google says uploads, queries, and responses are not reviewed by human reviewers or used to train AI models, and the applicable Workspace terms differ. These statements do not replace your institution’s data-governance review.

• Do not upload patient-identifiable or otherwise restricted information unless your organization has explicitly approved that service and use.
• Check copyright, license, consent, and institutional rules before uploading manuscripts, protocols, or unpublished data. A freely accessible file is not automatically cleared for cloud processing.
• Do not treat an AI-generated summary, citation, or extracted value as a verified scientific result. Confirm it in the original source and preserve relevant caveats.

Google changes product limits and availability. Check its current usage-limit guidance and source-import instructions before planning a large notebook; the fixed source-count and word-limit figures in older tutorials can become stale.

Diagnose common failures

Practice and self-check