What I really use NotebookLM for

I've been using NotebookLM for months, which since 16 July 2026 has been called Gemini Notebook. It hasn't made me study ten times faster, but it has saved me hours of reading documentation blind. Here I explain what it's genuinely useful for, what it does better than a chat, where it falls short and how I combine it with Claude.

A few months ago a client handed me the documentation for their management software: about 14 PDFs, more than 600 pages of manuals, release notes and a technical appendix apparently written by someone who hated humanity. I had to understand how it exported billing data to connect it to another tool, and I had two days.

I uploaded it all to a notebook and started asking: where the export was configured, what formats it supported, what happened with custom fields. Every answer came with its citation, and I went to that bit of the PDF to check it. In about three hours I knew what I needed and which two chapters I had to read in full. Doing it the hard way, I reckon it would have taken me a day and a half.

Now the counterpoint. The "study ten times faster with this trick" videos promise that the tool understands for you. It doesn't. What it saved me was searching: knowing where everything was. What it didn't save me was thinking, or calmly reading the two important chapters. I had to do that myself, as always.

NotebookLM is Google's tool for working with your own documents. You upload PDFs, web pages, videos or notes, and the assistant answers, summarises and explains using only those sources. Every answer shows which part of your documents it comes from, so you can go and check.

That's the big difference from an ordinary chat. A chat answers with everything it learnt in training: it explains well, but it can invent things or wander off topic. NotebookLM anchors itself to what you give it. That cuts down the inventions a lot, though it doesn't eliminate them entirely; further down I tell you about a few I've run into.

The name change throws a lot of people. Since 16 July 2026 it's been called Gemini Notebook, but it's the same product and your existing notebooks are still there; I didn't have to move anything (checked in September 2026). And don't confuse it with the Gemini chat: the chat answers with everything it knows, and Gemini Notebook works only with your sources. Two different products with a similar surname.

It's my most frequent use, by a long way. Whenever I get landed with documentation I don't know, I open a notebook. The latest example: the API documentation for a payment gateway a client wanted to integrate into their shop, about 200 pages of guides and reference.

My way of working is always the same. First I ask for a general summary to see what's there. Then I ask specific questions, and here's the important part: I don't trust any answer without opening the citation. With the payment gateway, the summary said partial refunds were supported; the citation led to a paragraph saying they were supported only in certain countries. The answer wasn't false, but it had swallowed the nuance that mattered to me.

Compared with reading it the hard way, I save roughly half the time in the getting-my-bearings phase. Compared with asking a general chat, I gain confidence: the chat would have answered with what it knew about that gateway in its version from two years ago, and the documentation was this year's.

Before a meeting with a new client, I usually set up a notebook with whatever I have. Recently, for a meeting with a renovation company that wanted to automate its quotes, I uploaded its website, two PDF catalogues and the email in which they explained the problem. I asked it to summarise what they sold, what seemed to worry them and what questions I should ask. I arrived at the meeting knowing what it was about, and the first half hour wasn't spent on them explaining what was already on their website.

What I never upload: clients' personal data, contracts with names and figures, invoices, passwords or anything the client hasn't already published or authorised me to use that way. My rule is simple: if I wouldn't put it in a presentation in front of another client, I don't upload it. When I need to work with something sensitive, I either anonymise it first or don't use the tool.

This is where it has changed the way I work the most, and not because of the summaries. To study with AI, what works is using it as a patient teacher and not as a homework machine.

My real example: when I had to learn to use Proxmox, the program I manage virtual machines with on some servers. I uploaded the official documentation and a couple of guides and, instead of asking for summaries, I asked it for three things. To explain each concept as if I were starting from scratch and then with an example. To quiz me at the end of each topic. And, what helped me most, to let me explain the topic and have it look for mistakes in my explanation. It caught two concepts I thought I understood and had the wrong way round.

The two-voice audio overviews, the ones that sound like a podcast about your documents, I've used mainly in the car, to go over something I'd already studied. For learning something new they don't work for me: I get distracted by the chummy tone and the content slips past me. For revision, fine.

If you're wondering which is the best AI for studying, my answer is that it depends on the material. For general questions, any chat. For studying from your own sources, a tool like this one, because working on your documents cuts down the inventions a lot.

If the answer isn't in the sources, it's not its territory. Sounds obvious, but I forget it more than I'd like. Once I asked it about a configuration option that wasn't in the uploaded documents and it gave me a reasonable answer, but one built from pieces of other options. The citation led to a paragraph about something else. That's why I always check.

Another failure I've seen: with poor-quality scanned documents, some tables came out jumbled and the numbers in one column got mixed up with those in another. And with several versions of the same manual in the notebook, it sometimes mixed what the old version said with the new one. Since then I only put in the current version.

What it doesn't do, and shouldn't: decide what's important. The summary highlights what appears a lot, not what matters to you. That's your call.

Practical limits: with very large notebooks, over forty sources, I've noticed the answers become more general and cite less. I don't know if it's an official limit; it's what I've found.

My workflow is simple: Gemini Notebook to understand a specific pile of documents, and Claude to produce and act on what I've understood.

An example from start to finish: with the payment gateway from before, I first used the notebook to understand the payment and refund flow and pull out a list of the points that affected the client. Then I gave Claude that list, with the citations, and asked it to help me write the integration. When Claude proposed something that didn't add up, I went back to the notebook to check what the documentation said. Each does what it does best: one sticks to the sources and the other builds.

I use Gemini Notebook alone when all I need is to understand something or prepare a meeting. I use Claude alone when there's no specific pile of documents behind the task, which is most of my daily work.

If you want to try it, this is what I'd tell you:

1. One notebook per topic. Mixing a contract with a syllabus and a program's documentation is a recipe for strange answers.
2. Good, focused sources. Better five good documents than forty half-relevant ones, and only the current version.
3. Questions that force a citation. "According to these documents, where is … explained?" works better than "how do you …?".
4. Check the citations. Always, especially when the answer suits you.
5. No sensitive data. If you wouldn't show it, don't upload it.

And a final note: if you had NotebookLM notebooks or links, according to Google the old links keep working under the new name.

What's true in the promise of studying ten times faster? Finding things, yes, much faster. Understanding them, no: that still costs the same, though with a patient teacher beside you it costs a little less.