NotebookLM: Working From Your Own Sources

The president of a condo board receives the minutes of three annual meetings, a quote for a facade renovation, and the condominium rules. He has to find out over a weekend whether last year's vote allows the work to go ahead. In an ordinary chatbot, he would paste random chunks and get a confident-sounding answer. In NotebookLM, he uploads the five documents, and every sentence of the answer points to a specific passage.

NotebookLM, Google's tool, doesn't answer from its general memory but from the sources you give it: documents, web pages, pasted text, depending on what your account allows. That's its main appeal, and it's also the source of misunderstandings.

One notebook, your sources, and nothing else

You create a notebook per subject, add your sources, then chat. A few habits simplify everything.

• One notebook per project. Mixing the condo's meeting minutes with your Italian class notes is pointless: the answers get muddled and you lose the readability of the citations.
• Readable sources. A PDF scanned crooked or a table saved as an image may be extracted badly. If an answer looks strange, open the source and check whether the text was read properly.
• Clear file names. "AGM Minutes 2025" helps the tool and you find your way around. "scan0043" doesn't.
• Variable limits. The number of sources and the size of each file depend on your plan and can change: check in the app before building an organization around a number.

In the sources panel, you can also check or uncheck documents. For a specific question, leave only the ones that matter active: the answer is sharper and comparing two texts becomes possible.

Ask questions that force citations

The flaw of a question that's too broad is that it invites a smooth summary. The best questions ask for an answer that can be located.

Based on the meeting minutes only: what vote was taken on the facade renovation, by what majority, and on what date?

For each element, quote the passage from the document. If two sets of minutes contradict each other or the information isn't in them, say so clearly instead of choosing.

The last part changes the answer. Without it, faced with two versions, the tool picks one without telling you.

To compare, name the sources in the question:

Compare the quote from the company [NAME] with the work approved at the meeting. Make a three-column table: item, what was approved, what appears in the quote. Mark with a "?" each line you can't reconcile, and don't fill in the gaps for me.

This instruction is very useful for any consistency check: a contract against a specification, a course syllabus against a curriculum standard, an invoice against a quote.

The extras: notes, summaries, study

You can save a good answer as a note, with its citations, then use it in another notebook or as a new source. NotebookLM can also produce summaries in other forms, such as an audio version or a mind map, depending on the features of the moment. For a student preparing for an exam, the most rewarding remains the quiz:

Based on chapters 3 and 4 of the course, ask me ten review questions, one at a time, from the simplest to the hardest. Wait for my answer before moving to the next one. After each answer, correct me by quoting the passage of the course that gives the right answer.

It's a good use because the correction rests on your course and not on what the AI believes it knows about the subject.

What can still go wrong

A citation tells you where the tool went looking. It doesn't guarantee it understood correctly. It can summarize a paragraph by making it say more than it does, mix two passages, or read a figure from the wrong column. For the points that carry weight (an amount, a voting majority, a deadline), click the citation and read the original sentence. It takes ten seconds, and it's the method of checking answers.

Second limit: the tool only knows what your sources contain. If the answer depends on a legal text that isn't in them, it won't have it, and it may replace it with generalities. For a legal point like the one in the condo, the result is for preparing your questions to a professional, not for deciding in their place.

Third point: what you upload leaves your computer. Before adding a document containing personal data (lists of condo owners, student files, client contracts), read Google's terms for your type of account and remove what isn't useful. Our article on privacy gives the landmarks.

NotebookLM or a chatbot with an attached file?

If you have a single question about a single document, a chatbot with the file attached is enough, and you'll find advice in the article on long documents. NotebookLM becomes more interesting when you come back several times to the same set of documents, want to compare sources against each other, or insist on seeing the passage behind each claim. For cases where you don't own the sources and are searching the web, a tool like Perplexity answers a different need. On limiting inventions in general, see reducing hallucinations.

To test this weekend

Take a set of three documents you already know by heart, for example three months of your team's meeting reports. Upload them, ask a question whose answer you know, and check the citations. You'll know whether the tool reads your files properly before entrusting it with a question whose answer you don't know.