I’ve been testing NotebookLM for a few weeks now, specifically focusing on its automated prompts. The “suggest related questions” feature pops up regularly, but I’m questioning its actual utility versus just adding clutter.
Sometimes the suggestions are obvious rephrasings of what I just asked. Other times, they leap to tangential topics that aren’t relevant to my current source material. This feels like it could mislead newer users into a shallow query loop instead of building a deeper line of inquiry.
I want to know if others have found a practical use case for it. Does it genuinely help you discover angles you hadn’t considered when working with your documents? Or does it mostly serve as a distraction, encouraging you to chase generic questions instead of your own focused analysis?
Share your workflow examples if you have them. I’m particularly interested in whether this feature improves over time as you add more sources, or if it remains a static, noise-generating function.
—AF
—AF
I've found it depends entirely on the quality of your source material. With a messy, unfocused doc, the suggestions are all over the place. But I loaded a tight set of campaign performance reports last week and the suggested questions actually surfaced a correlation between email send time and mobile open rates I hadn't thought to check. That was a legit "aha" moment for me.
So maybe the utility isn't in the feature itself, but in how well you've curated the sources it's reading? I've stopped expecting it to build my analysis for me. Now I treat it like a junior analyst throwing out ideas - most are noise, but occasionally there's a gem.
Has anyone noticed if it gets better as you interact with it, or is it purely source-dependent?
—b
That's a really good way to frame it - like a junior analyst throwing out ideas. I've had a similar experience with Obsidian's related notes plugin, which feels like the same concept. It was useless until I spent a weekend cleaning up my tags and links. Then it started making decent connections.
So I'm leaning towards it being source-dependent, but I wonder if there's a feedback loop? Like, if you ignore the bad suggestions and only engage with the good ones, does the model learn your style a bit, or is it a fresh start each session?
Self-host or die trying.
The source-dependent argument makes sense, but I'm skeptical about the feedback loop. These features are usually stateless session widgets, not persistent learning agents. You're curating your own thinking, not training a model. It's an illusion of interactivity.
I've seen this pattern before, like with those "recommended next steps" in cloud consoles. They're canned responses based on keyword matching or simple embeddings from your current context. If you click one, it doesn't make the next batch smarter, it just runs a new query with a slightly shifted context.
Treating it like a junior analyst is generous. It's more like a very literal intern that only reads the headings and bold text. If your docs are messy, it's confused. If they're clean, it can do a basic lookup. The moment you expect it to learn your style, you're over-engineering what is essentially a fancy autocomplete.
monoliths are not evil
Agreed, the feedback loop point is spot on. I've never seen one of these features get smarter in real time. It's a simple retrieval, not a training session.
I do think the "junior analyst" comparison still works, though. Even a literal intern can be useful if you give them the exact right file and tell them exactly what to look at. The problem is when we expect them to figure out what's important on their own. That's when the illusion breaks.