NotebookLM chat mode — grounded Q&A

Most people who pick up the tool for the first time treat the chat pane like any other AI chat window — they type a question, they get an answer, they move on. The experience feels familiar enough that the key difference can take a while to land: every answer is anchored to your documents, not to a large language model's general training data. That single distinction changes what you can do with the output.

In a general chatbot, an answer about a clinical trial might draw on anything the model absorbed during training — papers, blog posts, Wikipedia edits, social media summaries. You have no way of knowing which. In this research notebook, the same answer draws on whatever you uploaded — and the footnote tells you exactly which paragraph of which document. If the paragraph is wrong, the answer is wrong in a traceable way. You can fix it. That traceability is why professionals in high-stakes fields — legal, medical, policy — have adopted the tool in contexts where a general chatbot would be unsuitable.

How citation works

When you submit a query, the indexing layer converts it to an embedding and searches the source corpus for the passages most semantically similar to the query. It retrieves the top-ranked passages — typically ten to twenty — and passes them to the generation model alongside the original question. The model composes an answer using only those passages as its evidence base, and as it writes each claim, it records which passage supported it.

The result is an answer in which every claim carries a superscript or inline citation marker. Click the marker and the source pane scrolls to the relevant paragraph, which is highlighted in yellow. You are looking at the exact text the model used — not a paraphrase of it, not a nearby section, but the specific paragraph. For a 200-page PDF this is the difference between "somewhere in here" and "third paragraph, page 94."

Citation accuracy is very high for well-formatted text-layer PDFs and Google Docs. It is somewhat lower for scanned PDFs (OCR drift can misplace paragraph boundaries) and for audio-transcribed sources where timestamps approximate text positions. The tool will occasionally cite an adjacent passage rather than the most precise one — a known limitation of fixed-window document chunking — but in practice this is uncommon enough that reviewers can verify answers quickly.

Multi-source queries

The retrieval step searches all active sources in the notebook simultaneously. A query about a theme that appears in several documents retrieves the most relevant passages from each and weaves them into a single answer with per-source citations. This is not a sequential process — the tool does not summarise document one and then document two and then combine them. It retrieves across the corpus in a single pass, which is why the answers can draw on six sources in a single paragraph without obvious stitching artefacts.

When sources disagree, the assistant is designed to surface the disagreement. Ask "what do these papers say about the efficacy of treatment X?" and if two papers reached opposite conclusions the answer will say so, with a citation to each paper's conclusion. This behaviour is deliberate — the tool is not trying to give you a confident synthesis when the evidence is genuinely split. Researchers running literature reviews find this particularly valuable because it flags disagreements they might not have noticed reading each paper separately.

Follow-up behaviour

Within a session, the tool carries conversation context. Follow-up questions using pronouns ("what else does it say about that?") or implicit references ("and the third study?") resolve correctly against previous turns because the conversation history is included in the context window passed to the model. This makes multi-turn research conversations practical — you can dig progressively deeper into a topic without restating the full question each time.

The context window has a finite size. Very long sessions — dozens of exchanges — may cause the earliest messages to fall out of the active window. When this happens, an implicit reference to something from early in the conversation may not resolve correctly. The practical workaround is to save important answers as notes (which are persistent) rather than relying on the chat history to carry context indefinitely.

When the model abstains

The grounding guarantee cuts both ways. If you ask a question and none of the retrieved passages contain a relevant answer, the model does not reach for its training data. It tells you it could not find the answer in your sources. This is the right behaviour — a confident wrong answer would be worse than an honest non-answer — but it is worth knowing about so you are not surprised when it happens.

Abstention typically occurs in three situations: the question is genuinely outside the scope of the uploaded sources, the relevant passage exists but was not retrieved (rare, usually fixed by rephrasing the question), or the source containing the answer is present but currently hidden. If you receive an abstention response and believe the answer should be in the corpus, check that no sources are hidden and try rephrasing the query with more specific terminology from the documents themselves.

Saving answers as notes

Any chat answer can be saved to the notes studio with one click on the "Save to note" button that appears on the answer card. The saved note preserves the full text and all citations. From there it can be renamed, edited, pinned, and exported to Google Docs or Markdown. Many researchers use a workflow of asking a series of targeted questions in chat, saving the best answers, and then combining and editing them in the notes studio into a coherent briefing document.

Query type reference table

Query type	Behaviour	Example
Factual lookup	Retrieves specific passage, cites it directly	"What was the sample size in study B?"
Synthesis across sources	Retrieves from multiple docs, composes unified answer	"How do the three reports differ on pricing?"
Comparative	Notes agreements and contradictions per source	"Do the authors agree on the mechanism?"
Follow-up / pronoun reference	Resolves against conversation history	"What else does it say about that?"
Out-of-scope	Abstains; says answer not found in sources	"What is the GDP of France?" (if not in sources)
Saved to note	Answer preserved with citations in notes studio	Any answer via "Save to note" button

Octavius B. Nakashima-Rowe, Investigative Journalist at Heartwood Regional Weekly in Wellington, described the chat mode's value in deadline contexts: "The citation click is the feature that earns its keep. When I am forty minutes from filing and I need to verify a claim from a source I uploaded two weeks ago, I ask the notebook. The footnote takes me to the line. I read the line. I file. That is two minutes instead of twenty."

For a policy-level perspective on citation requirements in AI-assisted professional work, the Stanford AI research portal has published useful framing on how grounding and attribution practices vary across professional domains.