I’ve never really cared that AI can draft my emails, optimize my schedule, write a quick script for me, churn out a slide deck, and so on. Yes, I’ve used it for all of these tasks before, and I’ll definitely continue to do so. However, none of it has ever made AI feel like it actually belonged in my workflow. Instead, it just felt like a faster way to do things I already know how to do. What ended up changing my perspective and convincing me that AI could improve my productivity within my workflow was mandating that the AI start with the documents I provided before it reached for anything else.
NotebookLM was the first tool that actually worked that way, and it’s the reason why I started actively building AI into how I work. I stuck with NotebookLM exclusively for a long time, until I saw how much more I could do once I moved that same grounded setup into Claude Projects. That’s exactly what I did, and most of my research has lived there ever since. Or at least, it did, until NotebookLM’s Gemini 3.5 update pulled half of it back.
NotebookLM recently got a massive update
It’s been busy
A few days ago, NotebookLM got what Google itself is calling its biggest update to the tool yet. I’ve covered the full update in detail in a separate article, but the short version is that the tool now runs on Gemini 3.5, and each notebook now ships with a sandboxed “secure cloud computer.” This cloud computer, which is powered by Google Antigravity, enables it to write and run actual code.
While this last bit is something that confused me a tad too when it launched, and it still sometimes does (I don’t see why a research assistant needs to code), it does mean that NotebookLM can now do tasks beyond just reading through your files and explaining what they do. It can now act on them by writing and running code to analyze data, run scripts, graph trends, build spreadsheets (among other outputs), and so on.
While I’ll talk about the main reason I switched from NotebookLM to Claude Projects in the section below, one reason I started leaning more toward the latter was the outputs I could create with it. Credit where credit is due, NotebookLM has a great (and increasing) selection of outputs you can generate, like Audio Overviews, Mind Maps, Video Overviews, Slide Decks, Infographics, and Reports. All of these lived in the Studio panel as one-click, templated quick generators, though. Most had customization options, but they were still fairly rigid.
With this update, though, you can now ask for outputs directly in the chat interface. You can ask for data visualizations and charts, PDFs, Word docs, Excel sheets, PowerPoints, and CSV/JSON files. Rather than dropping a templated file on you, NotebookLM plans the output, writes the code to build it, and runs it, with the finished (and fully editable) file landing in your Studio panel. Ultimately, this meant that NotebookLM’s inability to produce the polished, editable deliverables I needed was no longer a reason to leave it for Claude Projects.
I tested Gemini Notebooks and Claude Projects side by side, and one didn’t make the cut
Same idea, different answers
NotebookLM’s reasoning is finally visible
No more taking its word for it
When a tool claims to be a research assistant, I think it’s very important that I can see how it actually reaches its conclusions rather than just what it lands on. For the longest time, NotebookLM didn’t offer that. It would take my sources, do its thing, and hand me an answer. While that answer was always grounded and cited, I still had to take the path it took to get there on faith. With this update, that finally changes. One of the most tangible benefits of this update is that you can follow its reasoning directly in the chat, watching it work through a task step by step instead of just dropping a finished answer on you. When it’s generating an answer, you can simply expand its reasoning and watch it work through the task in real time.
Given that I use NotebookLM for a lot of cross-synthesis, I lean on this feature when I don’t quite agree with the conclusion it’s reached, since being able to trace its steps makes it far easier to spot where it went wrong than just taking its word for it. With NotebookLM now running on Gemini 3.5, I’ve also noticed it’s a lot better at reasoning through denser material than it used to be. I’ve found it much quicker, and the responses I get are more in-depth and nuanced.
Before this update, there were a few times when NotebookLM would show me the “NotebookLM can’t answer this question” or similar errors (I don’t remember the exact wording), and I haven’t run into one since the update. Whether that’s the new model or just luck, I can’t say for certain, but the responses I’m getting now feel noticeably more capable than what I was used to.
NotebookLM still can’t pull from outside sources on its own
Claude Projects gives me the grounding and the open web
Before I go any further, though, let me link back to why I’ve been such a huge fan of NotebookLM’s structure for so long and why I’ve long preferred it for anything research-related. NotebookLM’s built around being source-grounded. Ultimately, unlike a tool that pulls from the open web and occasionally makes things up along the way, NotebookLM only ever worked with the sources I actually gave it. Everything it told me could be traced straight back to something I’d uploaded, citation and all. That closed, you-control-the-inputs setup is exactly what made it feel trustworthy in a way most AI tools just don’t, and it’s the whole reason it never felt like just another chatbot to me.
So when I say I moved to Claude Projects, I want to be clear that I didn’t move away from that grounding. Instead, I moved to something that does the exact same thing, just with more room around it. Claude Projects works on the same principle: you give it your sources; it leans on them as its base, and what it hands back is anchored in your material with citations, rather than pulled from elsewhere. That RAG-style, your-notes-first setup is what I’ve always praised NotebookLM for, and it’s alive and well in Claude Projects, too.
The engineer behind NotebookLM’s best features showed me his exact setup, and it changes everything
198 articles later, I finally got to ask the team how they actually use it.
The difference is what happens at the edges. NotebookLM, even after this update, still can’t natively pull from outside sources mid-answer the way I’d want it to. Yes, it’s now better at finding new sources, and if you choose to, you can add them to your notebook, technically ending up with outside material in the mix. However, that’s a deliberate, you-curate-it step, not a tool reaching past your documents on its own when a question genuinely needs it.
With Claude Projects, my notes are still the first thing it reaches for, but it can also draw on what it knows about them when that’s actually useful, without me having to go hunt down and manually feed it every single source first.
The other thing that keeps me in Claude Projects is structural. A Project isn’t one endless thread. I can spin up a bunch of separate conversations that all draw on the same underlying knowledge base, which means I’m not stuck scrolling through one ever-growing chat to find where I asked about something three topics ago. Because of the way I actually research, which involves lots of small, parallel threads hanging off the same pile of sources, that separation matters more than I expected it to.
The best setup uses both
So that’s the half of my research that stayed put. NotebookLM closed the output gap, made its reasoning visible, and got noticeably sharper with Gemini 3.5 under the hood. However, the grounded-plus-reach combination and the freedom to splinter my work into separate conversations off a single knowledge base are still something only Claude Projects gives me.
The best approach I’ve found here is to work with both interchangeably. Thankfully, there are some great community-built MCP servers that bridge the two, so you can wire NotebookLM into a setup like Claude’s and move between them without rebuilding your source base each time.
