NotebookLM capabilities — a deep-dive inventory

Describing the tool as "an AI research assistant" covers the marketing but obscures the engineering. What the assistant actually does is run a three-phase pipeline — index, retrieve, generate — inside a boundary defined by the sources you uploaded. Every capability listed below is a variation on how that pipeline is configured and what it produces. Understanding the pipeline makes it easier to predict when the tool will perform well and when it will need human oversight.

The boundary is the key concept. Unlike an open-domain chatbot, this research notebook refuses to answer from general knowledge when grounding is active. If the answer is not in your sources, the tool says so. That refusal is not a weakness — it is the mechanism that makes the output auditable. A lawyer or a scientist cannot use "a chatbot told me" as a source; they can use "page 14 of this paper, as cited by the notebook." The grounding boundary is what enables the latter.

Retrieval-augmented reasoning

When you submit a query, the tool converts it into an embedding and searches the indexed source corpus for passages with high semantic similarity. It retrieves a ranked set of passages — typically the top ten to twenty, weighted by relevance — and passes them to the Gemini generation model alongside the original query. The model composes an answer using only those passages as evidence, citing each one. This retrieve-then-generate pattern is known as retrieval-augmented generation (RAG) and it is the foundation of every capability in the tool.

The reasoning quality depends on what is in the passages retrieved. For factual queries with a clear answer in the sources, accuracy is high. For complex inferential questions — "what does the combination of sources A and B imply about the mechanism described in source C?" — the model synthesises across the retrieved set, which works well when the relevant passages are semantically similar enough to land in the same retrieval batch.

Summarisation

The tool summarises at several levels of granularity: a single source, a selected subset of sources, or the full notebook corpus. Summaries are generated via the notes studio and are grounded in the same retrieval pipeline as chat answers. Each paragraph in a generated summary links back to the passage that motivated it. The tool also performs implicit summarisation in every chat answer — composing a response that synthesises multiple retrieved passages into a single coherent paragraph.

Audio generation

The audio overview capability transforms the notebook into a scripted dialogue between two AI hosts. This is not a text-to-speech pass over a written summary — the tool generates a separate dialogue script optimised for audio comprehension, with question-and-answer structure, repetition for clarity, and pacing that suits listening rather than reading. Four modes (Brief, Standard, Deep Dive, Critique) target different audience needs. The audio overviews page covers this in full detail.

Citation mechanics

Every generated output — chat answer, briefing doc, study guide, audio overview script — includes citations. In chat and notes the citations are inline: a superscript or footnote-style marker that, when clicked, opens the source pane at the highlighted passage. The citation points to a specific paragraph or sentence, not to the document as a whole. This granularity matters: for a 300-page PDF, "see this document" is not useful; "see paragraph 4 on page 47" is actionable.

Citation accuracy is generally high for text-layer PDFs and Google Docs. It degrades somewhat for scanned PDFs (OCR artefacts can shift paragraph boundaries) and for audio-transcribed sources where timestamps approximate rather than exact text positions. The tool will also occasionally produce a citation that points to a slightly adjacent passage rather than the most precise one — a known limitation of retrieval systems that chunk documents into fixed-size windows.

Multi-source synthesis

One of the genuinely distinctive capabilities is cross-document synthesis. A query like "how do sources A, B, and C differ on the question of inflation targeting?" will retrieve the most relevant passages from all three documents and compose a comparative answer with per-source citations. The model is also designed to surface disagreements rather than paper over them — if two sources contradict each other, the answer will note the contradiction rather than silently choosing one view.

Multi-source synthesis is also what powers the audio overview feature at the notebook level: the overview is a synthesised narrative across the entire corpus, not a per-document summary stitched together. That is why overviews from notebooks with ten diverse sources tend to be more interesting than overviews from notebooks with ten near-identical sources.

Language support

Source ingestion is language-agnostic. Chat and note output follow the prompt language. Cross-language operation — French sources, English questions — works well because the retrieval step operates on multilingual embeddings that map semantically equivalent phrases across languages to similar vector positions. Audio overviews added nine language-specific host pairs in 2025; the output language is chosen independently of the source language at generation time.

Sharing and collaboration

Sharing is a capability in the same sense that retrieval is: it extends what the tool can do to additional people. A shared notebook makes the indexed corpus and all saved notes available to collaborators within their assigned role. Viewers can run queries and listen to overviews without editing. Editors can add sources and generate notes. The Plus tier adds analytics and retention policies. Read-only public links allow broader distribution — a researcher can share a notebook publicly so anyone can query it without signing in.

Export

The tool exports notes to Google Docs and copies them as Markdown. Audio overviews download as MP3. These export paths are the tool's interface to the outside world — the mechanism by which outputs leave the notebook and enter other workflows. Export does not include the source index itself; you cannot export the full notebook as a portable file. What exports is the generated content.

Capability matrix

Capability	How it works	Notes
Grounded Q&A	RAG: retrieve relevant passages, generate cited answer	Will abstain if answer not in sources
Summarisation	Notes studio generates grounded prose summaries	Supports single source or full corpus
Audio overview	Generates dialogue script, renders with voice models	4 modes; 9 languages; MP3 download
Citation	Inline passage-level links from output back to source	Accuracy varies for scanned PDFs
Multi-source synthesis	Retrieval across all indexed sources in one query	Surfaces contradictions between sources
Language support	Multilingual embeddings; prompt-language output	Cross-language queries supported
Sharing	Role-based notebook access; public read-only links	Analytics on Plus tier
Export	Google Docs, Markdown, MP3	Index not exportable
Mind Map	Visual concept graph of source relationships	Plus tier only
Abstention	Declines to answer when evidence absent	Core grounding guarantee

Ramona C. Ashcroft-Mbeki, Producer at Gulden Research Collective in Cape Town, put the value succinctly: "The multi-source synthesis is the capability that changed our workflow most. We can now load a full conference's papers into a notebook and ask a single question that draws on twenty of them simultaneously. Six months ago that would have taken two researchers a full day."

For a broader policy-level analysis of how retrieval-augmented tools fit responsible-AI frameworks, the OECD AI policy observatory publishes updated guidance that covers grounding, citation, and auditability requirements for AI-assisted knowledge work.