I stopped dumping raw sources into NotebookLM, and it got so much better

Published Aug 24, 2026, 12:30 PM EDT Mahnoor Faisal is a tech journalist covering AI and productivity tools with bylines at XDA, SlashGear, MakeUseOf, Laptop Mag, and Android Police. She's been writing professionally since she was sixteen, and has since penned hundreds of articles. This includes in-depth coverage of AI tools like NotebookLM to breaking news across the AI space. Her passion for technology started when she received her first iPod Touch (4th generation) on her 8th birthday, and she's been deep in the tech world ever since. Currently pursuing a degree in computer science, Mahnoor brings both a journalist's eye and a technical foundation to her coverage of how AI is reshaping the way we work and learn. With most AI tools, the difference between a good response and a subpar one depends largely on the prompt you feed it. This is why "prompt engineering" was such a massive thing in the early days of ChatGPT, since it was what differentiated an average response from a useful one. With NotebookLM, though, I've found that the prompt is only half the equation. Given NotebookLM is built around helping users work with the sources they provide the tool rather than scouring the web for information, what you feed it matters just as much (if not more) as what you ask it. This is exactly why I stopped dumping raw sources into NotebookLM and expecting it to do all the heavy lifting. Everything starts with the sources you add Your notebook is only as good as what goes in With NotebookLM (now called Gemini Notebook), what you get out of it is almost entirely dependent on the sources you add to your notebook. This is because the tool grounds the AI assistant in those sources and uses them to generate responses with inline citations, rather than relying on its internal training data or general web knowledge. When you ask a question that isn't addressed in your sources, NotebookLM won't go ahead and give you an answer for the sake of giving you an answer. It'll also not pull in outside information just to fill in the gaps. Instead, it'll tell you that it can't answer based on the sources you've provided. This is exactly why the quality of those sources matters so much in the first place. If you feed it a source that's entirely inaccurate to begin with, the information NotebookLM fetches from it will naturally reflect those inaccuracies. While the information it'll give you will still have clickable citations, and you'll be able to use them to trace every claim back to where it came from, a citation only tells you where the information came from. It doesn't guarantee that the source itself was accurate in the first place. The onus for that is entirely on you. That means vetting my sources before adding them to a notebook has become just as important as anything I do with them afterward. While a lot of NotebookLM users recognize the importance of a source being high-quality in terms of accuracy, most don't realize that how that source is structured and presented can matter just as much. Accuracy is only half of what makes a good source A good source needs more than facts Something I realized after months of using NotebookLM for both studying and work is that a source can be perfectly reliable and still not make for a particularly good NotebookLM source if important parts of it never make it into the notebook in the first place. At this point, chances are you're wondering how that can even happen when you're supposedly giving NotebookLM the source itself. The answer comes down to how NotebookLM actually imports different types of sources. Web links are probably the clearest example of this. When you paste a webpage URL into NotebookLM, it doesn't import the entire page exactly as you're seeing it in your browser. According to Google's documentation, it only scrapes the text content of the HTML page. Images, embedded videos, and nested webpages aren't imported at all. Paywalled webpages aren't supported either. Additionally, you also get a lot of unnecessary text along with the actual article. If I pasted a link to the very XDA article you're reading right now, for example, NotebookLM wouldn't necessarily get just the article body. Depending on how the page is structured, it can also pull in things like navigation text, related articles, footer copy, and other elements surrounding the piece. None of that is necessarily wrong, but it is noise when all I really want NotebookLM to work with is the article itself and the context provided by its images. The same idea extends beyond normal webpages. If you add a YouTube URL, NotebookLM imports only the video's text transcript, not the video itself. So if the person speaking says something like "as you can see here" while demonstrating the most important part on screen, that visual context isn't automatically carried over with the transcript. This is why I've stopped treating "NotebookLM supports this source type" as meaning "I should throw it in as-is." Before adding something now, I think about what NotebookLM is actually going to receive from it. If some of the most useful information lives in a chart, diagram, image, or other context that won't survive the import, I'll usually turn that into something more self-contained first instead of assuming NotebookLM will piece it together for me. I started preparing my sources before uploading them A little prep goes a long way Once I realized the above, I stopped treating the upload button as the very first step of the process. Now, I usually spend a little time preparing a source before it ever makes its way into my notebook. What that looks like depends entirely on the kind of material I'm working with. I split giant sources into smaller chunks For instance, if it's a particularly long source like a textbook or a semester's worth of lecture material within a single PDF, I'll usually split it into smaller, more logical chunks before uploading it. A textbook might become one source per chapter, while a massive lecture pack might be divided by week or topic. While Google allows individual sources of up to 500,000 words, I've found that the tool doesn't work nearly as well for me when I give it one incredibly dense source covering dozens of different concepts. Breaking that same material into smaller chunks makes it much easier to target the exact chapter, lecture, or topic I want NotebookLM to work with instead. I clean up messy PDFs before uploading them I've also started cleaning up messy PDFs before uploading them. For instance, if I'm working with documents that have unusual layouts, poorly extracted text, complicated tables, or simply formatting that won't work well once I add it to my notebook, I'll often convert the material into Markdown. Rather than leaving NotebookLM with a wall of extracted text, I can preserve things like headings, subheadings, examples, and the overall hierarchy of the document. I can also add a short description of an important diagram or table if I know that visual context would otherwise be lost. I fix transcripts before making them permanent sources Lecture recordings and videos are another place where this makes a noticeable difference. When you upload an audio file to NotebookLM, Google says the file is transcribed and that transcription becomes the source. YouTube sources similarly rely on the video's transcript rather than the actual video. That means any mistakes or messiness in the transcript become part of the material NotebookLM has to work with. Anyone who has dealt with automatic transcription knows what that can look like — technical terms get mangled, sentences run into each other, and a lecturer correcting themselves halfway through a thought can make the final transcript unnecessarily confusing. So, if a transcript is something I'm going to rely on heavily, I'll clean it up first. I fix obvious transcription errors, remove filler where it adds nothing, break the text into sensible paragraphs, and add headings when the topic changes. I don't summarize the lecture or remove substantive information; I'm just making the original material easier to navigate. I'll also sometimes feed the transcript into an AI assistant like ChatGPT or Claude and ask it to clean up the formatting and obvious transcription errors without summarizing or changing any of the actual information. This saves me from having to go through a lengthy transcript line by line myself. I give NotebookLM some context first Sometimes the issue isn't that a source is messy at all. It's that NotebookLM doesn't have the background knowledge about my particular project or course that I do. This is where I like adding a small source of my own before everything else. For studying, that might be my syllabus, a glossary of course-specific terminology, or a simple document explaining which textbook chapters correspond to which weeks of the semester. For work, it might be a short source explaining the project, terminology, or how the other sources relate to one another. Ultimately, the trick here is just to do a tad bit of prep work before you upload anything to NotebookLM. You don't need to completely rewrite or summarize every source first, but taking a few minutes to clean up obvious mess, split overly dense material, or add missing context can make the notebook much easier to work with!