共有:
Research Agent

NotebookLM is no longer
an AI that only reads.

Answering your uploaded documents with citations—that was the NotebookLM of yesterday. The new NotebookLM searches the web on its own, verifies numbers with code, and "researches and confirms" in a single loop. From a citation-based answering tool to a research agent. We unpack that shift with diagrams.

AI Navigate Editorial·2026.06.11·6 min read
BEFORE Answers your docs with citations Research and checks done by you AFTER Reason Web search Run code Runs autonomously in one loop
01
The Old NotebookLM

Until now, it
"only answered with citations"

Three months ago, NotebookLM was a tool with a clear specialty. For the PDFs and documents you uploaded yourself, it would answer while showing, with citations, where the information was written—a reliable "reader," strong at research where you want to confirm the evidence.

But its scope was confined to the material at hand. Finding missing sources on the web, verifying the numbers it surfaced through calculation—all of that was the user's own work. It was strictly a "citation-based answering tool," with no autonomy to go investigate on its own. That is why many people split their workflow: "NotebookLM when I want citations, ChatGPT for web research."

The old NotebookLMThe new NotebookLM
Answers your docs with citationsFinds the sources it needs from the web itself
Number checks done manually by the userVerifies numbers on the spot by running code
Human bridges research → analysis → verificationReasoning, search, and execution in one loop
A passive "reader"An autonomous "research agent"

From an AI that reads documents,
to an AI that goes and investigates on its own.


02
How It Works

The "research and confirm" loop

For a single question, the new NotebookLM moves back and forth between reasoning, web search, and code execution, assembling the answer on its own.

Question Plan it out by reasoning Web search Finds sources on its own Run code Verifies numbers on the spot Integrate & re-reason Back to search if needed Repeats until satisfied Verified answer Citations + numeric backing
FIG. Starting from reasoning, it assembles the answer by moving between web search and code execution
01

Plan by reasoning

On receiving a question, it first works out on its own "what to investigate and what to confirm by calculation." This is the starting point for judging whether the material at hand is enough or whether it needs to consult the web.

02

Move between search and execution

It discovers the sources it needs on its own via Google Search, and runs code on an independent "cloud computer" to verify the numbers. The novelty is that it can run research and calculation in a single flow, without a human in between.

03

Repeat until satisfied

It integrates the results it obtains and, if they fall short, goes back to search and execution. At the end of this back-and-forth, it produces an answer backed by both citations and numbers.

03
Under the Hood

The revamp is
a "new brain" and an "execution environment"

The autonomy comes down to two things: swapping the foundation to Gemini 3.5 Flash, and giving it a dedicated, independent environment for running code.

BRAIN Gemini 3.5 Flash Instruct Google Search Cloud computer Independent env for code Agentic workflow The agent ties search and execution together
FIG. The new brain calls Google Search and the code-execution environment, tying them into an agentic flow
3.5
The underlying Gemini Flash
78.2%
Advantage over the prior version (internal eval)
¥0
No extra charge with a Google account

The first is the overhauled foundation. NotebookLM has been moved to a Gemini 3.5 Flash base, integrating an independent "cloud computer" for code execution and autonomous source discovery via Google Search into an agentic workflow. The "researching" and "calculating" that humans used to bridge have been built in as moves on the model's side.

The effect shows in the numbers too. In internal evaluation, it is said to have outperformed the prior version in 78.2% of cases. From a citation-backed answering tool to a research agent that runs reasoning, web search, and code execution in a single loop—its very character has changed dramatically.

04
In Practice

Who it lands with, and how

If you have a Google account, you can try it at no extra charge. The more your workflow boundary shifts, the bigger the benefit of this change.

Unify your research work

It lands with people who want to complete the flow of "analyze while pulling sources, verify numbers with code" in a single tool.

Want citations and verification together

For research that demands both the origin of the evidence and numeric backing at once, you can rely less on splitting work with ChatGPT.

Try it at no extra charge

If you have a Google account, you can try the new workflow right away. The low barrier to adoption is a practical strength, too.


05
Frontier

Sometimes it is overkill

It is not all-purpose. For people who want to manage and curate their sources themselves, the autonomous searching behavior can, in some cases, feel harder to handle instead. For simple one-shot Q&A use, the agentic back-and-forth may be overkill.

Even so, it is a fact that the long-standing split of "NotebookLM when I want citations, ChatGPT for web research" now has more solid reasons to lean toward NotebookLM. For people who want to entrust the whole sequence—investigate, confirm, and cite—to a single tool, this revamp becomes a move that cannot be ignored.

AI Navigate — Daily Update · 2026.06.11

NotebookLM Evolves Into a Research Agent That Searches and Runs Code | AI Navigate