NotebookLM's Auto-Categorization: A Smarter

There's a particular kind of digital graveyard most knowledge workers know well. It's the folder—or the app, or the notebook—where you dutifully saved everything and then never went back. Good intentions, terrible follow-through. Not because you're disorganized by nature, but because the tool made retrieval harder than just Googling it again.

Google's NotebookLM has been trying to be the antidote to that problem since its launch: a research tool that lets you have a genuine conversation with your own documents, grounded in sources you actually trust rather than the open web's chaos. As of April 24th, it shipped what might be its most practically significant update yet—automatic source categorization, now rolled out to 100% of users on all plan tiers.

The feature is, on its surface, straightforward. Load five or more sources into a notebook—PDFs, Google Docs, YouTube links, websites, audio files—and the AI reads them all, groups related material together, assigns labels, and flags sources that span multiple topics with multiple tags. What was previously a flat, undifferentiated scroll of icons becomes something closer to an organized filing system that built itself while you weren't looking.

SEO educator Julian Goldie covered the update in a recent video, and his diagnosis of why it matters cuts to something real: "The reason most people stop using AI tools isn't that they don't work. It's friction. You upload 30 sources, you can't find anything. You give up, you go back to Google."

That's not hyperbole—it's a well-documented pattern in tool adoption research. The activation energy required to retrieve information often exceeds the activation energy to just search fresh. NotebookLM's original design was already better than most at answering questions from your own documents, but navigating to the right answer still required you to know roughly where to look. Auto-categorization attacks that specific bottleneck.

What the update actually does

The mechanics are worth understanding precisely, because the feature's value lives in the details.

When sources are auto-labeled, NotebookLM isn't just sorting by file type or date—it's reading the content and inferring thematic relationships. A single source can belong to multiple categories simultaneously, which is how information actually works. A customer interview transcript might surface under "pricing objections," "onboarding feedback," and "competitive mentions" without you having to manually duplicate it. The AI proposes the structure; you retain the right to override, rename, reassign, or add emoji (yes, emoji labels are supported, and yes, they make scanning genuinely faster).

Goldie demonstrates this with a few concrete business scenarios that are instructive precisely because they're mundane. A consultant uploading client deliverables—brand guides, past campaign reports, survey data, team interview transcripts—now gets those materials sorted into coherent clusters: brand voice, past performance, customer feedback, internal notes. When a client calls asking what their team said about pricing in a survey, the answer is one click away rather than a fifteen-tab excavation.

The update also quietly fixed bulk sharing in the same week—you can now paste an entire list of email addresses to share a notebook with a team, rather than adding collaborators one address at a time. Small change, genuinely annoying problem solved.

The grounding argument

Here's where NotebookLM's pitch gets philosophically interesting, and also where it most distinguishes itself from general-purpose AI assistants.

Most AI tools are, to put it plainly, creative with facts. They synthesize answers from training data that you can't inspect, can't cite, and can't verify in real time. NotebookLM operates differently: it only draws answers from the sources you've uploaded, and every response carries inline citations that link back to the original text. Click the citation, read the source, verify the claim. Goldie puts the practical stakes bluntly: "Your credibility took years to build. One made-up number can break it."

This grounding model isn't new to this update—it's been NotebookLM's core design principle from the start. But it becomes far more useful when you can actually navigate a large source library. Grounding only matters if you can find the right ground, as Goldie notes. Auto-categorization makes the grounded approach viable at scale, which is a meaningful shift for anyone who was hitting the practical limits of managing dozens of documents in a flat list.

There's a reason NotebookLM has reportedly surpassed Gemini itself in Google Trends popularity. For researchers, writers, consultants, and knowledge workers who need answers they can stand behind, a tool that refuses to hallucinate and tells you exactly where every claim came from has obvious appeal—even if the interface occasionally requires patience.

What the update doesn't solve

It's worth being clear-eyed about the limits here, because the update is being received with the kind of enthusiasm that tends to outrun what a single feature can actually deliver.

Auto-categorization is only as good as the sources you give it. Garbage in, organized garbage out. If your uploaded documents are inconsistent, redundant, or poorly scoped, the AI will dutifully cluster your mess into labeled piles of mess. The organizational intelligence is real, but it doesn't compensate for the prior work of deciding what's worth including in the first place. That curation judgment—which documents matter, which are noise—still belongs entirely to you.

There's also a question of how the feature performs at the higher source limits. The free plan allows 50 sources per notebook; the Pro plan bumps that to 300; an "Ultra" tier goes to 600. Auto-categorizing 50 sources is one thing. Whether the labeling remains coherent and useful at 300 or 600—where thematic overlap becomes far more complex—is a different question that will require real-world testing at scale.

And then there's the model tier distinction. Pro and Ultra users get Gemini 2.5 Pro under the hood (Goldie references "Gemini 3.1 Pro," though Google's current public naming convention differs—worth verifying against Google's own documentation before assuming). Free users get a capable but less powerful model. The auto-categorization works on the free plan, which is genuinely useful for evaluation. But the ceiling of what NotebookLM can reason about and synthesize will be higher for paying users, and that gap may widen as Google continues developing the platform.

The "second brain" question

The phrase "second brain" has been circulating in productivity circles for years—Tiago Forte built a methodology and a bestselling book around the concept. The idea is alluring: an external system that holds your knowledge so your biological brain doesn't have to carry it all. Every few years, a new tool claims to finally deliver on this promise.

NotebookLM is a more credible candidate than most, specifically because of its grounding constraint. A second brain that makes things up isn't a second brain—it's a liability. One that surfaces only what you've explicitly given it, with visible citations, is at least trustworthy within its scope.

What auto-categorization adds is something like executive function for that second brain: the ability to recognize relationships between ideas and impose structure without requiring you to do the taxonomic work manually. That's a meaningful capability. Whether it's "everything changed" territory or "useful incremental improvement" territory probably depends on how much you were already using NotebookLM and how many sources you're managing.

Goldie's framing is deliberately maximalist—"source organization isn't a cute feature. It's the difference between a notebook you actually use and a graveyard of uploads you forget about"—and he's not entirely wrong. The feature addresses a real failure mode. But the graveyard problem has many causes, and file organization is only one of them.

The update is live, it's free to test, and the use case is concrete enough that most people can evaluate it quickly. Five documents, a few seconds of processing, and you'll know whether the AI's sense of categories matches yours. That's a low enough bar to clear that the argument for just trying it is basically unanswerable.

Whether "second brain" ever stops being a metaphor and starts being a description—that's a longer question, and NotebookLM is only one data point in it.

By Marcus Chen-Ramirez, Senior Technology Correspondent