As someone who tests and writes about productivity tools for a living, NotebookLM is hands-down the tool I find myself using most. It’s packed with just the right features that don’t feel like AI purely for the sake of AI, and it’s clear that it’s designed to solve a specific problem — retrieving information with minimal (close to none) hallucinations.

It’s meant to help you interact with the information you upload, rather than pull from its training data or search the web in real time and mix that into your answers. As impressive as the tool is, that doesn’t mean it’s without its problems. I’ve been very vocal about the issues that plague NotebookLM, and while the team has addressed a fair number of them, there’s still one major drawback that holds it back significantly: its source limit.

NotebookLM has different source limits depending on your account

Not all tiers are created equal

NotebookLM operates on a notebook-based system, where each notebook essentially functions as a self-contained database. Every notebook you create is independent from the others and completely siloed from anything you’ve added elsewhere.

Within each notebook, the documents, website URLs, YouTube videos, pasted text, audio files, images, etc., that you upload are all treated as sources. These are the materials NotebookLM relies on within that specific notebook to generate answers and produce all of its outputs, including Slide Decks, Audio Overviews, Mind Maps, and more.

So, the sources you populate your NotebookLM notebooks with are key. They determine the quality, accuracy, and depth of everything the tool generates. If you add low-quality or overly dense sources to your notebooks, you’re naturally going to get weaker, less helpful outputs. The AI can only work with what you feed it, so the clearer, cleaner, and more structured your sources are, the better NotebookLM performs.

NotebookLM currently has four tiers you can subscribe to: Standard, Plus, Pro, and Ultra. Each tier lets you create a different number of notebooks and sets different limit caps. For instance, on the free tier, you can create 100 notebooks per user, while on the Pro and Ultra tiers, that number goes up to 500 notebooks.

Similarly, on the free tier, you can generate only 3 Audio Overviews per day, while on the Ultra tier, you can generate 200 per day — that’s about 66 times more!

Beyond limits on Studio outputs and the number of notebooks you can create, each tier also determines how many sources you can add to a single notebook. The caps break down as follows: 50 sources on Standard, 100 on Plus, 300 on Pro, and 600 on Ultra.

The tool begins to struggle long before you hit the cap

You don’t need 600 sources to break it

Given that NotebookLM recently announced the $200 Ultra plan, which lets you add up to 600 sources per notebook, the number of sources you can add isn’t really the core issue anymore.

Even if you’re not willing to spend $200 per month for the Ultra plan, there are ways to work around the source limits in terms of quantity. You can merge multiple PDFs, combine text files, consolidate notebooks, and even convert long web pages into a single document to maximize what counts as a single source.

I even created a Python script that lets me merge large batches of documents automatically, allowing me to squeeze as much content as possible into a single source slot (given I don’t currently have the Ultra plan).

That said, even if you manage to work around the quantity problem, you quickly run into a different issue altogether: quality. One thing I’ve noticed when using NotebookLM is that its accuracy drops as you approach the source limit.

The technology that NotebookLM uses is called RAG (Retrieval Augmented Generation), which allows LLMs to reference an authoritative knowledge base outside their training data before generating a response.

The way RAG works is that a user’s query is converted to a vector representation and then compared against the vector embeddings of your source documents to find the most semantically similar chunks of text.

These "top matches" are the only parts the AI actually reads to answer your question; it doesn’t re-read your entire notebook every time. This is where the quality issue arises. As you stuff your notebook with more documents (or massive merged files), you exponentially increase the size of the haystack.

The AI then starts missing obvious data because there’s too much hay to search through. I’ve had notebooks with over 200 sources that haven’t had any major issues with search, and notebooks with only three sources that struggled to give accurate answers.

This limitation is particularly evident when you upload Google Sheets, when a spreadsheet has multiple tabs, or when it contains large, dense tables with lots of data. The AI can easily get confused about which rows or columns are relevant, often pulling incomplete or inaccurate chunks as a result.

Even a single complex spreadsheet can overwhelm the retrieval process, causing answers to miss key details or misinterpret the data entirely. This is why I’ve had to resort to workarounds like converting each page of a spreadsheet to a separate PDF or CSV. This way, each section essentially becomes its own source.

I wish I didn’t have to choose between quantity and quality

This effectively means that NotebookLM's source limit forces you to choose between quality and quantity. There's currently no way to have both. The best approach is to create separate, focused notebooks for different topics or projects.

That way, each notebook stays manageable, the sources remain relevant, and the AI can deliver accurate, high-quality outputs without being overwhelmed.