How to Use Google NotebookLM, Stress-Tested on My Own Articles
NotebookLM answers only from the sources you give it. I fed it three of my own articles and spent an afternoon trying to make it wander, to show how it really works and where it fits.
Eddie Ochieng
August 11, 2026· Updated Sep 17, 2026
Most AI assistants answer from everything they were trained on, which is powerful and, occasionally, a problem, because they will confidently fill gaps with things that are not true. Google NotebookLM does the opposite. It answers only from the sources you hand it, and it cites them. That single constraint is the entire point of the tool, so rather than describe it, I tested it the hard way. I loaded three of my own AI-tool comparisons and spent an afternoon trying to make it break its own rule. This was September 2026, on the free tier.
How I tested this
I added three articles I wrote myself as the only sources, then ran a series of questions designed to catch it out. Some it could answer, some it could not, and one told it to ignore the sources entirely. Every answer below is real and unedited.
Start by Giving It Your Sources
NotebookLM is only as good as what you feed it, so the first step matters most. I gave it three comparisons, ChatGPT vs Gemini vs Claude, GitHub Copilot vs Cursor vs Cody, and Perplexity vs ChatGPT vs Google. You can paste text, upload PDFs, or add links, and the more focused the sources, the sharper the answers. Vague, sprawling input gets you vague, sprawling output.
Ask, and Always Check the Citations
With the sources in, I started broad and asked what themes ran across all three pieces. It did not just answer, it pulled a distinct theme from each article and told me which sources each one came from.
Several core themes run across the sources in your notebook. 1. Empirical, real-world testing over marketing hype (appears in all 3 sources). 2. Model convergence and “no single winner”: core capabilities have matured to the point where top tools perform similarly, shifting the decision from raw power to workflow fit (all 3 sources). 3. Viability of free tiers versus paid subscriptions (all 3 sources). 4. The necessity of human verification (all 3 sources).
Every claim carried a citation chip that jumped straight back to the sentence it came from. Then I tested precision with a single checkable fact, one that appears in only one of the three articles.
The cheapest paid AI plan mentioned across the sources is Gemini AI Plus, which costs $4.99 per month. For comparison, the second-cheapest option mentioned is ChatGPT Go at $5 per month.
Correct, and it cited the exact article. When an answer lives in one of your sources, NotebookLM finds it and shows its work.
The Real Test, Does It Stay on Your Sources
This is what NotebookLM is actually for, and where I pushed hardest. First, something my articles do not mention at all.
The population of Kenya is not mentioned in your sources. Your sources focus strictly on evaluating and comparing AI tools, including general assistants, AI coding tools, and AI search engines.
No guess, no filler, just an honest "not in here." Then the real trap. I told it to abandon the sources and give me its own opinion on the best tool.
I cannot offer personal opinions or draw on outside knowledge, so I cannot name a single best tool, and according to your sources, attempting to crown one “winner” in the abstract is the wrong approach. The ideal choice depends entirely on the specific task and workflow.
It held the line completely and redirected me to what the sources actually say. That discipline, refusing to wander even when instructed to, is the whole product. Asked to quote a specific line about privacy, it returned a real sentence from the right article rather than a paraphrase.
Do not point a personal-tier assistant at proprietary code. (Source: GitHub Copilot vs Cursor vs Cody)
Use the Built-In Tools
NotebookLM can do more than answer questions. From the same sources it will build a study guide, an FAQ, a briefing document, flashcards, a timeline, even a mind map, each drawn only from your material. I generated the study guide, the FAQ and the briefing doc. The briefing doc was the most useful of the three, a tight executive summary of the through-lines across all my pieces, the kind of thing that would take me an hour to write by hand.
The Audio Overview, and a Word on What It Can See
The headline feature is the audio overview, which turns your sources into a spoken discussion between two AI hosts. I listened to the whole thing. It was accurate to my articles, a little more emphatic in tone than I would ever be, and, importantly, it invented nothing. Better still, it surfaced connections between my own pieces that I had not consciously made.
One thing caught me out. It spoke about my methodology and even a little of my biography, which made me wonder for a second whether it had read my whole website. It had not, and it cannot. NotebookLM only ever sees the sources you give it. Because each of my articles carries my byline and bio alongside it, that detail was sitting inside the text I pasted. It is a neat proof of the very rule the rest of this test kept demonstrating. NotebookLM does not go beyond your sources, for better and for worse.
Where It Fits, and Where It Does Not
NotebookLM is a research specialist. It is superb when you already have the material, papers, a contract, lecture notes, your own drafts, and you need to understand, cross-reference or summarise it without the model drifting into invented facts. It is the wrong tool when you want open-ended help or knowledge from beyond your documents. For that, reach for a normal assistant. Think of it as the tool for making sense of what you have, not for fetching what you do not.
The habit that matters
The one habit that makes it sing: feed it good, focused sources, and click a citation before you trust an answer. NotebookLM will not invent, but it can only ever be as good as what you give it.
This was tested in September 2026. For assistants that answer from everything rather than only your sources, see ChatGPT vs Gemini vs Claude. And for the through-lines this exercise surfaced across my own reviews, see five things every AI tool test keeps proving.
FAQ
Is Google NotebookLM free?+
Yes. I ran this entire test on the free tier. Higher limits and some extras come with Google’s paid AI plans, but the core tool, sources, chat, citations and the audio overview, is free.
Does NotebookLM make things up?+
In my test, no. Asked about something not in my sources it said so, and told to ignore the sources and give an opinion it refused. It only answers from what you give it, which is the whole point.
Can NotebookLM read my whole website or account?+
No. It only sees the specific sources you add to a notebook. If it seems to "know" something extra, that text is inside one of your sources, as my own bio turned out to be.
What is NotebookLM best for?+
Understanding material you already have, study notes, papers, contracts, your own drafts. It is a research and comprehension specialist, not a general assistant for open-ended questions.
