Google NotebookLM 'Better Research' Update: Full Breakdown — Gemini 3.5 × Antigravity Turns It Into an Active Research Agent

NotebookLM used to be a passive RAG tool that answered questions based only on sources users uploaded. With this update, it becomes an active research agent that autonomously searches the web for sources, writes and executes code, and outputs results in a wide range of formats including PDF, DOCX, XLSX, and PPTX. Powered by Gemini 3.5 and Google's new Antigravity coding agent framework, it achieves 4x faster processing than rival LLMs and a 65%+ overall win rate.

Google announced the major update on June 8, 2026, through the official blog post Do better research with NotebookLM. Coverage from 9to5Google and gihyo.jp quickly followed, signaling a significant shift in the AI research tool market. The rollout began immediately for Google AI Ultra and Workspace AI Ultra Access subscribers, with broader availability coming 'soon.'

The update rests on three pillars: (1) AI engine upgrade with Gemini 3.5 and Antigravity, (2) agentic research capabilities, and (3) diverse output format support. Each is detailed in the sections below.

Beyond the model upgrade to Gemini 3.5, Google integrated its new-generation coding agent framework Antigravity into NotebookLM. For more on Antigravity, see the Google Antigravity 2.0 column. The integration gives NotebookLM access to 100+ curated software skills and transparent reasoning traces. According to SiliconAngle, it is 4x faster than rival LLMs. For more on Gemini 3.5, see the Gemini 3.5 Flash + Omni column.

Google's official benchmarks show an overall average win rate improvement of 15+ points, reaching 65%+. By category: large-scale document analysis at 69.9% and web research and source discovery at 78.2%. These figures are based on comparisons against multiple rival LLMs, with the web research category showing the most pronounced advantage.

One of the most impactful changes is the allocation of a secure cloud computer (VM) to every NotebookLM workspace. Previously, NotebookLM could only read documents and generate text responses. With an embedded VM, it can now write and execute code directly. For example, it can aggregate financial data from an uploaded spreadsheet via Python, generate a chart, and export it as PNG or SVG — all within NotebookLM. This represents a clear differentiation from traditional RAG tools and delivers meaningful value for data analysts and business intelligence practitioners.

The previous requirement to pre-upload sources has been significantly relaxed. Users now only need to type a rough idea or question, and NotebookLM autonomously discovers and adds relevant primary sources via Google Search. It also searches for sources in multiple languages, streamlining research that spans Japanese, English, and Chinese materials. This feature directly competes with Perplexity Spaces' web search agent capabilities, intensifying competition in the AI research tool market.

Coupled with the VM capability, output format support has expanded substantially. Document formats include PDF, DOCX, Markdown, and plain text. Spreadsheet formats cover XLSX and CSV. Presentation output supports PPTX. Data formats include JSON and structured data. Visual outputs include PNG and SVG charts, and image outputs support PNG, JPG, and GIF. This coverage nearly spans all file types needed in typical business workflows, making it realistic to complete an entire research-to-deliverable cycle within NotebookLM.

While new features are being added, the full existing feature set remains available and is being enhanced. Audio Overviews (podcast-style audio summaries), Video Overviews, Mind Maps, Study Guides, Briefing Docs, Timelines, Flashcards, and Quizzes all continue. Source support also expands: Google Sheets, images, and DOCX files are newly supported alongside the existing PDF, Google Docs, Slides, YouTube, web URLs, and audio files.

The distinction between NotebookLM's research modes and Gemini Deep Research is becoming clearer. Fast Research focuses on quick lookups and summaries, while Deep Research handles thorough analysis and runs in the background. Gemini Deep Research covers the entire public web, whereas NotebookLM uses a hybrid approach combining personal sources with selective web searches. NotebookLM is better suited for investigations that involve confidential documents, while Gemini Deep Research is stronger for comprehensive sweeps of public information. Read this alongside Google Search's AI prioritization shift to understand how Google's entire research ecosystem is being restructured.

The new agentic research and VM features are initially exclusive to Google AI Ultra ($199.99/month) and Workspace AI Ultra Access subscribers. According to felloai.com's pricing summary, the current plan tiers are Free, Plus at $7.99, Pro at $19.99, Ultra at $99.99, and Ultra at $200. The timeline for rolling out new features to lower-tier plans is listed only as 'soon,' with no specific date. Organizations considering business deployment should calculate the cost-benefit ratio of Workspace AI Ultra Access in advance.

Here is a quick competitive overview. Against ChatGPT Projects, NotebookLM has an edge in source attribution and citation transparency. Against Claude Projects, the gap in reasoning quality has narrowed since the Gemini 3.5 upgrade, making it a more balanced rivalry (see also AI Consulting Services). Perplexity Spaces is the most direct competitor given its web search agent feature, though NotebookLM holds the advantage in hybrid analysis combining private documents and web sources. Obsidian + AI plugins are strong for local processing and data privacy, but cannot match NotebookLM's cloud-scale throughput. Refer to the Google AI Optimization Guide when aligning tool selection with your organization's information security policies.

Key scenarios where Japanese enterprises can leverage the new features include: (1) Market and competitive research — type a broad question, let the agent collect Japanese and English primary sources, and generate a DOCX report. (2) Internal training materials — upload PDF manuals and use Audio Overviews for commute-friendly audio summaries. (3) Client proposals — auto-generate a PPTX deck and attach an XLSX data appendix with supporting evidence. (4) Multilingual research — receive Japanese-language summaries from English and Chinese technical papers. One important caveat: processing confidential data on a cloud VM requires advance review by your information security team.

Several questions remain unanswered in the official announcement. These include: the timeline for full Japanese-language interface and source support (multilingual improvement is an evaluation metric but is not explicitly stated); specific dates for rolling out new features to non-Ultra plans; VM specs, storage limits, and data retention policies; and the scope of administrator governance settings in Workspace environments. Continued monitoring is recommended.

Q1. Can free-plan users access the new features? A. Currently the rollout is limited to Google AI Ultra and Workspace AI Ultra Access. Expansion to Free, Plus, and Pro plans is described only as 'coming soon' with no specific date. Q2. Will data processed on the VM be used to train Google models? A. This is not explicitly stated in the official announcement. If you handle confidential information, review the Workspace terms of service and Data Processing Agreement. Q3. Will the agent automatically collect Japanese-language primary sources? A. Multilingual source discovery is included as a feature, and Japanese is expected to be supported, but it is not explicitly confirmed. Q4. Is Gemini Deep Research different from NotebookLM's Deep Research? A. Yes. Gemini Deep Research covers the entire public web, while NotebookLM uses a hybrid of personal sources and web search, making it stronger for analyses that include internal documents. Q5. Are existing Audio Overviews still available? A. Yes, they continue to be offered. Q6. What is the quality of PPTX output? A. Official demos show basic slide structures. Fine-grained design customization will still require editing in Google Slides or similar tools. Q7. How should organizations choose between NotebookLM and local tools like Obsidian? A. If data privacy is the top priority, local processing with Obsidian + AI plugins is preferable. If scale, speed, and multilingual support matter most, NotebookLM has the advantage. Q8. Which competitor poses the greatest threat? A. Perplexity Spaces is the most direct competitor due to its web search agent feature. However, NotebookLM maintains an advantage in hybrid analysis combining private documents with web sources.

The NotebookLM 'Better Research' update is more than a performance upgrade — it is a paradigm shift in what the tool fundamentally is. The three pillars of Gemini 3.5 × Antigravity engine upgrade, agentic research with autonomous source discovery, and VM-powered diverse output formats transform NotebookLM from a passive RAG tool into an active research agent. The current Ultra-only restriction is a real limitation, but once broader rollout happens, this has the potential to significantly reshape information gathering and document creation workflows for enterprises of all sizes. Japanese organizations should verify their information security policies and run an Ultra trial to assess workflow fit. Feel free to reach out via the AI Consulting Service for guidance.

Do better research with NotebookLM (Google Official Blog) / 9to5Google — Gemini 3.5 & Antigravity upgrade / SiliconAngle — Google upgrades NotebookLM / gihyo.jp — NotebookLM Research Update / felloai.com — NotebookLM Pricing 2026 / jetstream.blog — 3 New Features