NotebookLM Short Video Overviews: turn dense documents into 60-second vertical videos

For most of 2025, NotebookLM’s marquee output was the Audio Overview — two AI hosts talking through your sources like a podcast. In 2026 that expanded into Video Overviews: narrated slide-style videos in the standard 16:9 landscape shape, running several minutes. Short Video Overviews are the newest branch of that family, and the difference is entirely about format and intent.

A Short Video Overview is a 9:16 vertical video, capped at roughly 60 seconds, generated from the documents in your notebook. The vertical frame is the tell: it’s built for a phone held upright, for the same muscle memory that carries you through TikTok, YouTube Shorts, and Instagram Reels. Google leaned into the comparison openly — the launch was widely described as NotebookLM “taking cues from TikTok.”

What makes it more than a gimmick is that the content is still grounded in your sources. The clip isn’t generic stock footage with a voiceover; it’s a compression of your material — your research paper, your quarterly filing, your lecture notes — into the single most compelling point that fits in a minute. Where the long Video Overview lets an argument breathe across several minutes, the short one is ruthless: one hook, one idea, one takeaway.

The feature rolled out in early July 2026 and is powered by Nano Banana 2 Lite, a lightweight member of Google’s Nano Banana image-and-video generation line tuned for fast, on-device-friendly rendering. (Nano Banana renders the visuals; Antigravity — Google’s agent-first runtime and the other half of NotebookLM’s June 2026 upgrade — is the complementary backend that handles the reasoning and code execution. They work together: Antigravity is the brain, Nano Banana the renderer.)

You don’t need to understand the internals to use the feature, but knowing the rough pipeline helps you prompt it well. Conceptually, generating a Short Video Overview moves through four stages.

1. Core-concept extraction. NotebookLM reads across everything in the notebook — which, thanks to the 1-million-token context window available on every plan since January 2026, can be a very large pile of material — and isolates the one narrative that will land in sixty seconds. This is the step people underestimate: the model is choosing what to leave out, and your prompt is how you influence that choice.

2. Scripting and pacing. The system drafts a tight script structured for vertical video: a hook in the first few seconds, the core argument through the middle, and a synthesis or takeaway at the end. The pacing is deliberately punchy — there’s no room for the conversational tangents that make Audio Overviews feel relaxed.

3. Narration. The script is voiced with NotebookLM’s expressive AI narration — the same lineage of natural-sounding voices used in Audio Overviews, with inflection that tracks the emphasis of the script.

4. Visual assembly. Nano Banana 2 Lite renders the visual track in 9:16, pulling charts, diagrams, and text callouts that exist in your source material and animating them to match the narration. This is why how your sources are formatted matters: clean tables and clearly labelled figures give the renderer better raw material than a wall of unbroken text.

The result streams back as a downloadable vertical video. As with every NotebookLM output, treat it as a first draft you can regenerate — the same sources with a sharper prompt will produce a noticeably different clip.

NotebookLM now gives you three ways to turn sources into media, and they are not interchangeable. The right choice depends less on the content and more on the situation the viewer is in.

A useful mental model: the Audio Overview is two academics discussing a paper in a lounge; the long Video Overview is a patient lecturer at a whiteboard; the Short Video Overview is a world-class explainer who has one minute to make you care before you scroll away. That constraint is the point — it forces the model to cut the cognitive fat and lead with the single most disruptive idea in your sources.

Education and research. A lecturer can turn each week’s reading into a 60-second prep clip students actually watch the night before class — curriculum that lives in the same feed as everything else on their phone. Researchers can distil a dense paper into a shareable hook that makes the work legible to people who would never open the PDF, widening its reach without dumbing down the source.

Corporate and investor communication. A team can feed a quarterly filing or a strategy memo into NotebookLM and get a tight vertical summary of the headline numbers — useful for internal alignment or a retail-investor-friendly recap. Onboarding is another natural fit: instead of a 50-page handbook, a library of one-minute clips, each covering a single workflow, that a new hire can watch on a phone.

Media and content. Because the clip is grounded in whatever you upload, creators can turn primary documents — a government bill, a court filing, a historical record — into factual explainers without manual video editing. The output is only as trustworthy as the sources, but that’s exactly the point of a source-grounded tool: the claims trace back to the document, not to the model’s imagination.

A caveat worth keeping in front of you: a 60-second clip is a hook, not a substitute for the source. It’s superb for retention, sharing, and getting someone to care. For anything high-stakes — a compliance decision, a citation, a medical choice — the short video should send people to the full document, not replace it.

The mechanics are simple; the quality comes from your sources and your instructions.

1. Build a focused notebook. Upload the sources you want the clip drawn from. Fewer, cleaner sources produce a sharper single-insight video than a sprawling notebook where the model has to guess what matters.
2. Open the Studio panel and choose the video output. In the notebook’s Studio tools, select the Video Overview option and pick the short / vertical format where offered.
3. Customise before you generate. Use the instructions field to set the audience, the exact thesis you want defended, and the tone. This is the single highest-leverage step — a generated clip with no guidance defaults to a generic corporate voice.
4. Generate, review, regenerate. Watch for pacing and whether it led with the right idea. If the hook is weak or it summarised the wrong thing, tighten the prompt and run it again — regeneration is cheap.

Availability tracks the broader 2026 rollout pattern: newest Studio features tend to reach Google AI Ultra and Workspace tiers first, with wider access following. If you don’t see the vertical option yet, check that you’re on the web app and that your plan and region have received the feature.

Stage 1 — Source engineering. NotebookLM reads your formatting, so format for the outcome you want. Convert messy tables into clean Markdown tables. Insert explicit callouts in the source text — a bracketed line like [CRITICAL DATA POINT: 43% efficiency gain] flags a figure the renderer can promote to an on-screen graphic. Rename files to short, clean titles before uploading; the layout engine often surfaces the source filename as an on-screen anchor, and final_draft_v3_marketing_brief_copy.pdf is not the caption you want on a polished clip. Q3 Marketing Report is.

Stage 2 — Persona and focal-point curation. In the instructions field, define the target audience, the exact thesis to defend, and the pacing style. Don’t let the model pick a generic voice unless that’s genuinely what you want. The more specific the persona and the focal point, the less the model has to guess — and guessing is where sixty precious seconds get wasted.

Stage 3 — Structural audio-visual control. Tell the model how to spend the minute. Allocating explicit time blocks — hook, core problem, evidence, conclusion — prevents awkward mid-sentence cutoffs and keeps the clip from front-loading fluff. If a specific chart should appear, label it clearly in the source (Figure 1: Customer Acquisition Cost) and reference it in the prompt: “feature Figure 1 during the core-metric segment.”