NotebookLM quiz question quality: what the source-grounding fixes, and what it cannot.

Mostly yes, and for a specific reason. NotebookLM generates quiz questions grounded only in the sources you upload, not from a model's general knowledge of the topic, and when you get a question wrong it shows an explanation with a citation pointing back to the line in your material. That is the single biggest cause of wrong answer keys in AI question generators, and NotebookLM closes it. The honest caveat is the one Google states itself: the quiz is only as good as the material you upload. Feed it a sloppy slide and it will key a sloppy answer, faithfully.

Multiple-choice questions. That is the quiz feature. NotebookLM also has a separate Learning Guide chat mode, added in its April 2026 update, that asks probing open-ended questions in conversation, but that is a chat experience, not a gradeable quiz. If you want a free-response question, a clinical-vignette-style case question, or an image-occlusion card, the quiz feature does not produce them. It produces a stem and four options.

Yes, within limits. The customization panel lets you set difficulty to easy, medium, or hard, set the question count to fewer, standard, or more, and type a prompt to focus the quiz on a topic. Setting it to hard genuinely produces tougher questions. What difficulty does not change is the format: a hard question is still a harder multiple-choice question. You are turning up the difficulty of recognition, not switching to application.

No. NotebookLM's April 23, 2026 update added the ability to save progress, mark cards as got it or missed it, replay only the questions you missed, and shuffle card order. Shuffling the order stops you from memorizing a position. It does not change the wording of the question stem. By the second or third retake of the same quiz you can answer from the first few words of the stem rather than from the material, which is the point at which a retake stops teaching you anything.

The free tier allows 10 quizzes per day. That is generous for casual use and a real constraint if you are converting a full day of lectures and want to drill each deck several times. The paid tiers raise the cap substantially. The per-day quiz count, not the per-quiz question count, is the limit most students hit first.

It is good for the memorize-the-term layer: definitions, dates, mechanisms, anything where recognition is the skill. It is weaker for the layer most clinical exams actually test, which is whether you can apply a fact inside a scenario you have not seen. A multiple-choice recognition question and a vignette-style case question are different cognitive tasks, and a quiz feature that only produces the first one leaves the second one untrained. For a program where the exam is mostly application, that gap matters.

On the dimension you can measure, question-type coverage, Studyly produces four formats from one source where NotebookLM's quiz produces one. On overall quality, Studyly scored 81.3 on a held-out three-document eval run by Jungle, the company behind Studyly, graded on factual correctness, clarity, distractor quality, and question-type coverage. NotebookLM was not a tool in that eval, so there is no comparable number for it, and you should read 81.3 as Studyly's own measurement rather than an independent audit. The clearest, checkable difference is format coverage, not a head-to-head score.