NotebookLM

Positioning: a closed-book research assistant

Most AI chatbots are open-book: they blend your documents with everything in their training data and the live web, which is powerful and quietly unreliable. NotebookLM — Google's Gemini-powered research assistant, renamed Gemini Notebook in July 2026 — is deliberately closed-book. You upload your own sources, and it answers only from those sources, citing every claim back to the document it came from. When the corpus is silent, it says so rather than inventing an answer. In my testing of AI research tools, that single design choice is what turns a chatbot into an instrument you can act on, and it is why NotebookLM remains the tool I reach for when I have a fixed pile of documents and need to be right about them.

It launched in July 2023 as the shipping version of the Project Tailwind prototype, went mainstream when Audio Overviews shipped in September 2024, gained a Notebooks surface inside the Gemini app in April 2026, and was renamed Gemini Notebook two months later. The old name still dominates search results and third-party write-ups; the product is the same. What has changed most recently is the quota system: since 2 September 2026, fixed per-feature daily caps have been replaced by a compute-based usage budget that refreshes every five hours and stops at a weekly ceiling, scaled as a multiple of the free allowance per Google AI plan. Third-party guides still publishing fixed numbers like "500 chats per day on Pro" are describing a system Google retired.

What it actually does

You create a notebook, add sources — PDFs, Google Docs, websites, pasted text, YouTube links — and the engine reads them. Each source can run up to 500,000 words. Free accounts get 100 notebooks with 50 sources each; Google AI Plus ($4.99 per month at time of writing) lifts that to 200 notebooks and 100 sources; Google AI Pro ($19.99) to 500 notebooks and 300 sources; the two Ultra tiers ($99.99 and $199.99) to 500 or 600 sources. Those caps are published as fixed numbers and have not moved with the quota change.

On top of the source set, the tool generates artifacts the first time you add material: a report, flashcards, quizzes, mind maps, infographics, slide decks and audio or video overviews. Chat answers carry inline citations to the exact passage, and there is a Deep Research mode (10 runs per month on free) for questions that need broader grounding. Audio Overviews remain the headline: a ten-to-twenty-minute conversation between two synthetic hosts walking through your material, available in more than 80 languages with an Interactive Mode that lets you interrupt and ask questions. Video Overviews extend the idea visually, and Cinematic Video is limited to Pro and Ultra tiers.

The mental model that matters in daily use is bounded synthesis. Ask NotebookLM to compare your uploaded contract set against current market data and it cannot, unless you also upload the market data. Ask it to summarise, cross-reference, quiz you on, or audio-brief the material you gave it, and it does so with a citation rate that general chatbots do not match.

One collaboration detail worth knowing before you roll this out to a team: sharing a notebook does not raise the source limit for the people you share it with. Each collaborator works under the caps of their own plan, so a free-tier colleague dropped into a 300-source Pro notebook sees it through a 50-source lens for their own additions. On the paid tiers you also get controls over how shared notebooks are used, and watermark removal on infographics and slide decks arrives only at the Ultra tiers. For education and Workspace accounts the limits differ again from the consumer tiers documented here, so check the Google support page rather than a third-party table before you promise a class or a team anything specific. The published per-feature quotas — reports, flashcards, quizzes and mind maps capped at 10 per day on free and 100 per day on Pro, Deep Research at 10 per month on free — still exist alongside the compute budget, which is why hitting a limit can feel inconsistent until you work out which meter you drained.

My testing notes

Two observations stand out.

First, grounding held up where it counts. My habit with AI research output is to cross-reference answers against the original sources. With general chatbots I find discrepancies often enough to check everything; across the notebooks I ran — a stack of diligence documents, a set of technical PDFs, a folder of meeting notes — the citation rate was effectively clean. When I deliberately asked questions the corpus could not answer, the tool declined and said the materials did not contain the information. That refusal is a feature. The one soft spot is the Audio Overview: the two hosts are confident conversationalists, and when a source is hedged the audio sometimes smooths the nuance. I never treat the audio as a substitute for reading the cited passage.

Second, the quota system is now the practical constraint, not the source limit. In one deep working session on the free tier I hit the usage budget well before the 50-source cap — the budget drains faster on synthesis-heavy prompts across many sources than on quick lookups. Since the September 2026 change the budget is compute-based and recharges every five hours up to a weekly limit, so a heavy day of report generation, mind maps and audio can exhaust it; heavier jobs like Video Overviews can be queued to run after the reset. For casual use the free tier is generous. For a working researcher, Google AI Pro at $19.99 per month is the realistic tier, and it is worth knowing that no NotebookLM plan can be bought alone — you are buying a Google AI subscription that happens to include it.

The Audio Overview workflow genuinely compressed my triage time: generate the audio from a research stack, listen on a walk, then open the notebook to drill into what mattered. Generation takes roughly five to ten minutes and runs in the background.

What NotebookLM does not do

Its strengths and weaknesses come from the same place. Bounded to your sources means no outside knowledge, ever. It is strong at retrieval, summarisation and cross-referencing, and weaker at genuinely novel synthesis — connecting two sources into an insight neither states explicitly is still a job for a frontier reasoning model on top. The free tier's 50 sources per notebook is restrictive for large projects, and 500,000 words per source is the ceiling regardless of plan. Direct upload of video and audio files is not supported; YouTube links are. Audio Overviews cap around fifteen minutes, cannot be edited or partially regenerated, and offer no control over host voices or format. Reports from users and reviewers keep repeating one friction point: the quota messaging is confusing, and plenty of published limit tables online are stale. There is also no way to remove the watermark on some outputs below the Ultra tiers. And since it is a Google consumer product riding on a subscription bundle, continuity is a question every reader should weigh for workflows they intend to keep for years.

Verdict

NotebookLM is the rare free AI tool that changes a workflow rather than demoing well. I rate it 4.3 overall: excellent depth for source-grounded research and genuinely novel audio briefings, outstanding value on the free tier, clean usability despite quota confusion, an average support story for a consumer product, and strong if Google-shaped reliability signals. For students, analysts, lawyers and researchers who live inside a defined document set, nothing else at this price answers from your sources with citations this consistently.

Sources

• https://support.google.com/notebooklm/answer/16213268
• https://notebooklm-web-importer.com/guides/gemini-notebook-usage-limits
• https://aipricecompare.org/apps/notebooklm.html
• https://notebooklm-to-pdf.com/blog/notebooklm-plus-vs-pro
• https://www.jellypod.com/blog/notebooklm-pricing
• https://github.com/authorityandbrand/notebooklm/blob/main/docs/quota-limits.md
• https://aitooltier.com/tools/notebooklm
• https://toolsdepth.com/reviews/notebooklm-audio-overviews-review-2026
• https://valueaddvc.com/blog/notebooklm-review-2026-300-sources-per-notebook-audio-overviews-and-what-googles-ai-research-tool-replaces
• https://ai.gameya.com/article/notebooklm-review-2026