한국어

Is NotebookLM Good for Studying? What Makes It Stick

At a glance
  • NotebookLM, renamed Gemini Notebook in July 2026, answers from the sources you upload and shows citations.
  • Given a guideline as a source, it staged 86% and 70% of test cases correctly, far above models given the same text in a prompt.
  • Listening to and reading the same text led to about the same comprehension (g = 0.07). No study has tested Audio Overviews themselves.
  • Students who recalled a text with it closed beat students who mapped it with the text open, one week later (67% vs 45%).
  • Use it to understand, then quiz yourself before checking answers, and retake the quiz a few days later.

NotebookLM is good for studying when you need to understand your own material, because it answers from the sources you upload and cites them. Whether you remember that material later depends less on which feature you use than on how you use it. Listening to an Audio Overview or browsing a mind map is a form of review. What sticks is closing the sources and recalling first, which you can do with its flashcards and quizzes if you answer before you look.

What is NotebookLM, and what does it do for studying?

It is a tool that answers from your own sources and turns them into study material. Since July 2026 it has been called Gemini Notebook. Google says it is the same product, and existing notebooks stay where they were.

According to Google's page for students, you upload notes, lecture recordings and other class materials, then ask questions in chat and get answers with citations. From the same sources it can make Audio Overviews (a podcast-style talk between two AI hosts), Video Overviews, mind maps, slide decks, reports, flashcards and quizzes. Quizzes come with explanations for wrong answers.

Seen through memory, these features fall into two groups. Some show your material again in a new shape: Audio and Video Overviews, mind maps, slides and summaries. Others ask you for answers, which makes you pull it out of your own head if you answer before looking: quizzes and flashcards. The rest of this note is about what each group does.

Can you trust NotebookLM's answers?

In two studies it was far more accurate than chatbots given the same text in a prompt, but don't trust it blindly. The authors credit the way it looks up the relevant passages, not just having the text, and some answers were still wrong.

A team at the University of Yamanashi in Japan gave NotebookLM a summary of Japan's lung cancer staging guideline and asked it to stage 100 fictional cases from CT findings (Tozuka et al., 2025). NotebookLM got 86% right. GPT-4o got 39% with the same guideline pasted into the prompt and 25% without it. The same group's pancreatic cancer study (Johno et al., 2026) made a fairer comparison. Gemini 2.0 Flash, the model inside NotebookLM at the time, got 38% with the guideline pasted in and 35% without it, so pasting barely helped. NotebookLM got 70%.

NotebookLM beat pasting the same guideline into a prompt

  • Guideline pasted in prompt
  • NotebookLM (guideline as source)
  • No guideline
0%20%40%60%80%100%Guideline pasted in prompt · Lung cancer · 39%39%NotebookLM (guideline as source) · Lung cancer · 86%86%No guideline · Lung cancer · 25%25%Lung cancerGuideline pasted in prompt · Pancreatic cancer · 38%38%NotebookLM (guideline as source) · Pancreatic cancer · 70%70%No guideline · Pancreatic cancer · 35%35%Pancreatic cancerSame model,same guideline
Share of 100 fictional cases staged correctly (%). Lung counted the TNM stage, while pancreatic also required invasion and resectability. The comparison model was GPT-4o for lung cancer and Gemini 2.0 Flash, the model inside NotebookLM at the time, for pancreatic cancer.Source: Tozuka et al. (2025) abstract; Johno et al. (2026) abstract
See the numbers
Guideline pasted in promptNotebookLM (guideline as source)No guideline
Lung cancer39%86%25%
Pancreatic cancer38%70%35%

In both studies, the passages NotebookLM pointed to as its evidence were correct 95% and 92% of the time. That matters for a student, because the citation lets you check the original. Still, these were controlled tests with one guideline and made-up cases, and 14% of lung cases and 30% of pancreatic cases were still staged wrong. If it is going to be on the exam, follow the citation and read the source.

Do NotebookLM Audio Overviews help you learn?

No study has tested Audio Overviews. Listening to a text gave about the same comprehension as reading it, and one classroom study found podcast listeners scored lower. Use listening as review, like rereading.

Clinton-Lisell (2022) pooled 46 studies with 4,687 people that compared reading and listening to the same text. Overall, the difference was g = 0.07, which was not reliable. Reading came out ahead when readers set their own pace (g = 0.13) and on questions that needed inference rather than recall of what was stated (g = 0.36), with a smaller edge on general comprehension (g = 0.15). These studies played the same text aloud. An Audio Overview is different, an AI conversation that summarizes your sources.

Reading and listening differed little, most on inference questions

00.20.40.6Reading minus listening · Overall · 0.070.07OverallReading minus listening · Inference · 0.360.36InferenceReading minus listening · Self-paced reading · 0.130.13Self-pacedreadingInference g = 0.36
Comprehension difference between reading and listening to the same text (Hedges' g). Above 0 means reading did better. The overall difference was not reliable (p = .23), and on literal questions there was none (g = -0.01).Source: Clinton-Lisell (2022) abstract
See the numbers
Reading minus listening
Overall0.07
Inference0.36
Self-paced reading0.13

A classroom test of podcasts went further against listening. Daniel and Woody (2010) compared students who listened to a podcast of a primary source with students who read it as text. The podcast group scored lower on the quiz. Students preferred podcasts at first, and changed their minds right after the quiz.

So use Audio Overviews for a first pass or for another run on the bus. Then take the earbuds out and say three things you just heard, out loud. That small step turns listening into practice.

Do NotebookLM mind maps help you remember connections?

Recalling with the text closed left far more after a week than drawing a map with the text open.

Karpicke and Blunt (2011) split 80 college students across four ways of studying a science text. In the two compared here (40 students), one group drew a concept map while looking at the text. The other closed it, wrote down what they could recall, reread it, and recalled again. Both spent exactly the same time. On a test one week later, the recall group got 67% right and the mapping group 45% (d = 1.50). In a second experiment, recall won even when the final test was to draw a concept map, and across both kinds of final test, 101 of 120 students (84%) did better on the text they had recalled.

Recalling with the text closed beat mapping it a week later

0%20%40%60%80%100%Share correct · Map with text open · 45%45%Map withtext openShare correct · Recall with text closed · 67%67%Recall withtext closedd = 1.50
Share correct on a short-answer test one week later (%), 40 students in two of four groups. Time spent was the same, and the score combines verbatim and inference questions.Source: Karpicke & Blunt (2011), Experiment 1 results
See the numbers
Share correct
Map with text open45%
Recall with text closed67%

Most students didn't see this coming. In the second experiment, 90 of 120 (75%) believed mapping would work as well as recall or better. The authors added that maps are not the problem. A map drawn from memory, without the text, is itself retrieval practice.

A NotebookLM mind map is even more passive than the maps in this study, because you look at one rather than draw it. That is our inference from the experiment, not a test of NotebookLM's maps. If you want to use them, flip the order. Draw your own map from memory on a blank page first, then open NotebookLM's map and fill in the branches you missed.

What you remember comes from the moments you pull material out without looking, not from the time you spend looking at it again.

How should you use NotebookLM quizzes and flashcards?

Answer before you look, and retake them days later instead of once. These two features ask for answers, and they work as recall when you answer before you look.

Three habits help. First, say your answer in your head before you read the options. A meta-analysis of the testing effect found larger benefits from tests where you recall the answer than from tests where you recognize it (Rowland, 2014). You can also ask the chat for short-answer questions from your sources. Second, when you miss one, read the explanation and follow the citation, because AI-written questions can be wrong too. Third, take the same quiz again a day later and a week later. There is more on review timing in Spaced repetition schedule and on the pitfalls of AI-written questions in ChatGPT practice tests.

FeatureWhat it does for memoryTry this
Chat with citationsUnderstand and checkAsk, then open the cited passage
Audio OverviewReview by listeningAfterwards, say three points aloud
Mind mapReview by lookingDraw yours first, then compare
Quizzes and flashcardsRecall, if you answer firstAnswer first, retake days later

Something to try today

You don't need NotebookLM for this. Close today's chapter and spend five minutes writing everything you remember on a blank page. Then open the material and mark what you missed or got wrong in another color. If you use NotebookLM, make a quiz next and focus it on what you marked. Why reading a summary and stopping there sticks less is in Using AI to summarize readings.

In Aloom

The point of this note is that recalling first sticks better than looking again. Aloom turns that recall into a daily habit. You ask AI and keep only the answers you want. Kept answers join what you already know, and each day you recall 1, 3 or 5 of them, whichever you chose, as a question before you see the answer. The more you recall, the darker the node's ink. If NotebookLM helps you understand your sources, Aloom helps you keep what you chose from them for a long time. See how Aloom works

Frequently asked questions

Is NotebookLM now Gemini Notebook?

Yes. Google's page for students says NotebookLM became Gemini Notebook in July 2026. It is the same product, and notebooks you already made are still there.

Is NotebookLM better than ChatGPT for studying?

If you study from fixed material, NotebookLM answers from that material and shows where each answer came from, which makes checking easy. In the studies above, NotebookLM beat models given the same text pasted into a prompt (one of them GPT-4o). For broad questions beyond your material, a general chatbot can be handier. Either way, what you remember depends on whether you close the answer and recall it later.

Can NotebookLM make flashcards and quizzes?

Yes. It makes flashcards and quizzes from your sources, and quizzes explain wrong answers. They help most when you answer before you look and come back to them over several days.

References