NotebookLM can now find its own sources via Google Search

NotebookLM has always had a defining constraint: it only knew what you gave it. Upload a document, ask a question, get an answer grounded in that document. The loop was closed and controlled. Google has now broken it open.

The Decoder reports that NotebookLM has been upgraded to run on Gemini 2.5 Flash, execute code on its own cloud computer, and search the web via Google Search to find sources autonomously. In internal evaluations, the new system outperformed the previous version in 78.2 percent of test cases. That is not a marginal improvement to a side product; it is a structural change to how one of the more widely used AI research tools assembles its answers.

What the architecture shift actually means

The old NotebookLM was, in effect, a retrieval system operating over a user-supplied corpus. The new one is an agent: it can decide that the supplied sources are insufficient, query Google Search to find additional material, and incorporate what it finds before generating a response. Code execution compounds this. The tool can now run calculations, parse datasets, and validate outputs against live information rather than static uploads.

For users doing competitive research, policy analysis, or literature reviews, the experience will feel like an upgrade. For organisations that assumed NotebookLM's outputs were bounded by what they explicitly loaded into it, the calculus has changed. The model is now choosing sources, not just reading them.

That distinction matters enormously for any institution where cited sources carry institutional weight. A multilateral organisation using NotebookLM to support a policy brief can no longer assume the tool's citations are limited to pre-approved documents. The model may surface additional sources from the open web, including sources that contradict, contextualise, or displace the originals. The governance question is not hypothetical.

The citation mechanics behind the number

The 78.2 percent win rate in internal tests is the kind of figure that sounds impressive and reveals little on its own. What it tells us is that Google's own evaluation framework now rewards the agentic version consistently over the constrained one. The implication is directional: Google is building NotebookLM toward autonomous research, not away from it.

For B2B brands, this is the relevant shift. NotebookLM is used by analysts, strategists, and communications teams to synthesise information quickly. When those users ask the tool about a sector, a company, or a policy issue, the tool will now reach out to the web to fill gaps. Whether a brand appears in those retrieved sources, and in what framing, is now a material question for visibility. A financial services institution or an industrial group that has invested in published research, white papers, and indexed content stands a better chance of being pulled in as a supporting source. A brand that has not will be absent from answers it might otherwise have shaped.

The deeper structural point is that NotebookLM's upgrade is one instance of a broader pattern: the agent layer is eating the upload layer. Across the major AI research tools, the model is progressively taking over the task of source selection that users previously controlled manually. As that happens, the quality, discoverability, and indexability of a brand's published knowledge assets become ranking inputs, not just marketing materials.

Google has not published the full methodology behind the 78.2 percent evaluation figure, so the precise conditions under which the new system wins are not yet clear. What is clear is that a tool used by knowledge workers to produce cited research now finds its own citations from the open web. Brands that are not legible to Google Search are not legible to NotebookLM's new agent. The SEO implications and the AI visibility implications are, for once, exactly the same problem.