What the feature actually does
Audio Overview — the feature everyone calls "NotebookLM podcast" — reads the sources in a notebook and generates a two-host conversation that summarizes them. The output is a downloadable audio file (MP3), and in most interfaces today you can pick a format (deep dive, brief, critique, debate), a length, and an output language before generating. Interactive mode lets you join the conversation mid-playback in the consumer app.
The grounding constraint is the whole point: the hosts talk about your sources, and only your sources. A podcast over three loaded papers is a synthesis of those papers — it will not pull in outside knowledge, which is why the feature works as a commute-companion for dense material and fails completely as a general explainer on topics you haven't loaded.
Getting a good episode
- Feed it synthesis-ready sources. Three to six sources that cover the same ground from different angles produce the best conversations; twenty tangentially-related documents produce a shallow tour.
- Steer before you generate. The customize prompt (where the interface offers it) accepts instructions like "focus on the methodology disagreements" or "explain for a non-specialist" — it shapes the conversation, not just the intro.
- One audio overview per notebook. Generating a new one replaces the previous artifact; download before regenerating if you want to keep both.
- Length and format are separate knobs. "Brief" is a format and a length control exists alongside it — a short deep-dive is a different episode than a long brief.
The automation landscape
There is no consumer NotebookLM API — but there are two legitimate programmatic paths plus a gray area, and the entries on this page separate them honestly:
- Official Enterprise path — the Gemini Enterprise NotebookLM API (docs.cloud.google.com) exposes
notebooks.audioOverviews.createfor licensed enterprise tenants. Real API, real auth, narrow availability. - Unofficial automation SDKs — community packages drive the consumer web app via browser automation or reverse-engineered endpoints (Python and TypeScript SDKs exist with full artifact coverage: audio, video, slides, quizzes, flashcards, mind maps). They work today; they are not sanctioned, break when the UI changes, and entries flag that status plainly.
- Manual + export tools — the boring reliable path: generate in the app, download, publish. Tools on this page that automate the download-and-publish half (podcast feeds, file pipelines) sit on solid ground because they touch the output file, not the service.
Limits worth knowing before a workflow depends on it
Generation takes minutes, not seconds. Daily generation caps exist on free accounts. The audio is English-strongest — other languages work but the two-host chemistry is trained on English speech patterns. And the artifact belongs to the notebook: deleting the notebook deletes the audio.