What the NotebookLM podcast actually gives you, now under the Gemini Notebook name
On 16 July 2026 Google renamed NotebookLM to Gemini Notebook. It is the same product: old notebooks are still there, the Audio Overview button works the same way, and most people still search for NotebookLM, which is why both names appear in this guide. You upload sources, a PDF, a list of links, your notes, and the tool turns them into a conversation between AI hosts that you can download as an audio file.
The Audio Overview comes in four formats. Deep Dive, the default, is two hosts talking through the material. Brief is one host in under 2 minutes. Critique is two hosts giving a critical evaluation of the sources. Debate is two hosts in a formal debate. The usage split reported by the official account is Deep Dive 45.1%, Debate 26.7%, Critique 16.2% and Brief 12%. In other words, almost half of everything made with the tool is the same two host conversation, and 88% of it uses two hosts.
The controls are thinner than they look. Shorter, Default and Longer exist only in English, so in the 80+ other languages you get the default length. You can write a focus prompt with the topic and the level of expertise, but there is no documented voice selection, so every user gets the same hosts. The interactive mode is English only. Google's own note says the overviews, voices included, may contain inaccuracies or audio glitches. There is also a Video Overview that turns sources into narrated slides.
That is a lot of value for someone learning a subject. It is not the same thing as an automated AI video channel on YouTube, and the distance between an audio summary and a video that a channel can publish every week is exactly where monetization gets decided.
- Deep Dive: two hosts, the default, 45.1% of use.
- Debate: two hosts, formal debate, 26.7%.
- Critique: two hosts, critical evaluation, 16.2%.
- Brief: one host, under 2 minutes, 12%.
- Length control and interactive mode: English only. Voice selection: none documented.
Problem one: the same two voices as millions of other uploads
The voice quality is not the problem. The hosts sound natural, they laugh, they interrupt each other, they breathe. The problem is that they are the same two hosts, with the same verbal tics, the "wow", the "exactly", the "so basically", on an enormous number of uploads. Someone who has heard one NotebookLM podcast recognises the next one in five seconds, and so does a Partner Program reviewer.
The YouTube monetization policy (help center, answer 1311392) calls this inauthentic content, the name given on 15 July 2025 to what used to be repetitive content. It excludes repetitive or mass produced content, and among the examples lists AI-generated content made with generic or unoriginal templates. A channel where every episode is the same two default hosts reacting to a different PDF is that template in its purest form: only the input file changes. What the policy allows is a series with distinct episodes, your own commentary and your own narrative.
This is also why a nicer voice is not the fix. The guide on why AI narration sounds robotic covers the pacing problems that make a voice sound fake. Here the problem is the opposite: the voice is convincing, but it belongs to everybody. What protects a channel is a voice that is only yours.
In FalconVid that voice is your cloned voice or an ultra realistic premium voice chosen for the channel, and the channel DNA keeps the same voices, style and captions from episode 1 to episode 300.
Problem two: 12 minutes of audio over a still image
The Audio Overview is audio. To put it on YouTube, most people add a static cover, maybe a moving waveform, and upload. For 12 minutes the screen barely changes. The monetization policy lists slideshows with little or no narrative among the formats that do not qualify, and a still image with a waveform on top is that slideshow with a single slide.
On YouTube people watch, they do not only listen. A 12 minute conversation at podcast pace is roughly 1,700 to 1,900 words. A 12 minute video that holds attention needs a new image every 4 to 6 seconds, which means 120 to 180 scenes. The audio is maybe a third of the job.
Retention is what turns that into money. The new Partner Program rule asks for 8,000 public watch hours in 365 days for channels applying from 1 February 2027, while today it is still 4,000. A 12 minute video watched to 40% delivers 4.8 minutes per view, so 8,000 hours take 100,000 views. The same video watched to 20% delivers 2.4 minutes and needs 200,000 views for the same hours. The guide on how retention buys watch hours shows why every point of average view duration is a discount on the views you need.
In FalconVid the scenes are generated for the script, changing every few seconds, with karaoke captions in more than 15 styles on top, so the conversation has a picture that moves with it instead of a frozen cover.
Problems three and four: no control over the file, and sources that are not yours
Control first. Outside English the length is fixed. You cannot choose the voices. You cannot edit the script line by line: to change one sentence you regenerate the whole overview and get a different conversation. And when the warning about inaccuracies comes true, there is no targeted fix: a wrong date at minute 7 means living with it or regenerating and checking 12 minutes again.
Then the source. An Audio Overview summarises what you uploaded. If the sources are someone else's article, book or report, the episode is a reading of material you did not create, which the reused content policy excludes. Google does not claim the generated content, but its terms warn that it may generate the same or similar content for other people, and in the United States purely AI generated content without enough human authorship generally has no copyright protection.
The picture changes when the source is yours. If you already record your own show, the words were yours from the start, and the guide on turning podcast episodes into YouTube videos covers that path, which has none of these problems.
In FalconVid the length is a setting in any of its 63 languages, not only in English. The Studio lets you watch version 1 and fix a stretch, shorten the intro, swap a scene, change the music or open the timeline, without regenerating the whole video. And a finished project can be duplicated into another language paying only the difference, which is the opposite of a default length locked outside English.
What FalconVid does: the video podcast the NotebookLM file cannot be
FalconVid is built for the part NotebookLM stops before. You approve the calendar once, and AI specialists work in parallel: the researcher reads several sources, the scriptwriter builds an original script with your angle, and the narrator, the editor and the sound designer work at the same time, with the video ready in up to 30 minutes. Pasting your own script is optional.
The conversation format is native. Multi character with distinct voices is available on every plan, and in the AI voice narration you choose ultra realistic premium voices or your own cloned voice. Narration is billed by minute of video, not by voice: 24 credits per minute in premium, 288 for 12 minutes, and 3 credits per minute in economy, 36 for 12 minutes. The second voice costs nothing extra, and it is already inside the cost of the video.
Around the voices comes everything the audio file lacks: scenes generated for the script, karaoke captions, the thumbnail, the video SEO, a 9:16 Short of up to 60 seconds, and publishing on YouTube, Instagram, TikTok, Rumble and Facebook at the scheduled time.
Where NotebookLM is excellent, and the format that pays
None of this makes the tool bad. As a study and research tool it is excellent: it reads 20 sources faster than you, lets you hear a subject before writing about it, and shows in 10 minutes whether a topic has enough material for a 12 minute episode. For a channel, that is the ideal draft of the outline, not the episode.
The format that monetizes keeps what people like about the AI podcast, two voices talking, and changes everything that makes it generic. The script is yours, built from several sources with an angle. The two voices have jobs: a narrator who carries the story and an expert or a challenger who arrives 4 to 6 times with the number, the caveat or the objection the viewer is thinking. The guide on two voice narration versus one narrator shows why that function is what holds attention.
The math of an episode: by hand over the NotebookLM audio versus automated
By hand, the audio is the fast part. Generating it takes minutes, but you then listen to all 12 minutes to check the facts, often twice, because one error means a new generation. The visuals are the heavy part: 120 to 180 scenes, 3 to 5 hours. Syncing, captions, thumbnail, SEO and upload add 2 to 3 hours. That is 6 to 9 hours per episode, still with the same two hosts as everybody else, and one episode per week is the ceiling for someone with a full time job.
In FalconVid, on today's scale, a 12 minute video costs 1,731 credits in economy, 4,597 in balanced and 16,131 in premium, and the video arrives with scenes, captions, thumbnail, SEO and its Short. Starter, with 15,000 credits, covers 8 economy videos or 3 balanced per month, or a mix of 2 balanced and 3 economy that uses 14,387 credits. Pro covers 17 economy, 6 balanced or 1 premium across 5 channels. Business covers 54 economy, almost two per day, across 10 channels.
The manual channel over NotebookLM audio publishes one generic episode per week. The automated one publishes two or more per week from Starter, with its own voices, and scales to 109 economy videos across 25 channels on Agency or 184 across 50 channels on Scale.