NotebookLM (Gemini Notebook) use cases for many sources

NotebookLM — Google renamed it Gemini Notebook on July 16, 2026, same product — answers questions about the material you give it, and cites the passage each answer came from. It does not answer from outside your sources, which is the point: no source, no answer.

Its Studio panel builds the rest from those same sources: an Audio Overview (a discussion between AI hosts), a report in study-guide, FAQ or briefing form, a Mind Map, Flashcards, a Quiz (Google’s list, checked 2026-08-21).

So “what can I do with NotebookLM” has a dull answer — drop in a PDF and ask, covered in how to use NotebookLM — and an interesting one.

The interesting cases are the ones where the corpus is not one document but many items — eighty lectures, three years of a group chat, two hundred papers, a year of council hearings. That is where a cited answer beats a search box, and where two walls appear: Add source takes items one at a time, and the NotebookLM source limit is 50 per notebook on the free plan, 100 on Plus, 300 on Pro.

Seven cases below — the corpus, how it gets in, what to ask, the catch.

1. A lecture playlist or a course channel

Who: a student, or anyone taking a course. The corpus: 40 to 80 lecture videos on one channel or in one playlist.

NotebookLM takes public YouTube URLs and imports the caption text. Natively that is copy, switch tab, paste, wait — once per lecture. source-lm puts an Add to notebook button into YouTube’s own header: a playlist page adds everything loaded on it in one click, a channel its newest N videos, 50 by default. A channel run, end to end, is in the walkthrough; the product detail is on the YouTube page.

Ask:

• “Which topics did the lecturer say would be on the exam? Name the lecture for each.”
• “Where do two lectures give different accounts of the same thing?”
• “Which term is defined twice, and do the two definitions agree?”

The catch: every video is one source and nothing bundles them — 80 lectures are 80 sources. Transcript only: what the lecturer writes but never says is not in the notebook, and a video without captions does not import. The study routine is on the students page; the lecturer’s version, one class notebook per course, is NotebookLM for teachers.

2. A years-long Telegram group

Who: a member or moderator of a group that has been running for years. The corpus: tens of thousands of messages nobody can read.

Telegram Desktop and the native macOS client (12.10 and newer) both export a chat — JSON or HTML from Desktop, HTML from the Mac app. source-lm reads either and packs the messages, in order, into Markdown files budgeted at 400,000 words each — a margin under NotebookLM’s 500,000-word per-source cap, measured on the rendered Markdown. A message is never split across two files. In a real run, a 19.78 MB result.json came out as four sources covering about three years: the walkthrough, and the short version on the Telegram page.

Ask:

• “What was said about registering a business here, in order, and what changed over time?”
• “Which questions came up repeatedly and never got a clear answer?”
• “Which providers were recommended more than once, and by whom?”

The catch: text only. A photo or voice message with no caption becomes a placeholder; links and bold come through as plain words, without the URL; sender and timestamp live in a Metadata block under each message, not on its line.

3. A Telegram channel’s back catalogue

Who: anyone who follows a channel closely enough to argue with it. The corpus: every post a channel has published, going back years.

A channel exports from the same menu as a chat, in the same two formats, and packs the same way — a back catalogue lands as a handful of sources. Turning a Telegram channel into a knowledge base is the step-by-step; it works on private channels and needs no bot and no API key — the export comes out of the client you are already signed into.

Ask:

• “Summarise what this channel published in Q2, by theme, with dates.”
• “Where has the author’s position on [topic] changed? Quote the before and after.”
• “Which claims are repeated most often, and what evidence is ever given for them?”

The catch: an HTML export drops polls and service messages entirely, and a forwarded post announces itself with a Forwarded from X: line inside the body rather than in metadata. Fine for reading; not a source for counting posts.

4. A literature review of 200 papers

Who: a grad student, or anyone starting a review. The corpus: a Zotero collection of 200 references.

Two hundred PDFs do not fit a 50-source notebook, and they are the wrong first move anyway. Export the collection as CSL JSON — titles, abstracts, authors, venues, DOIs. That is text, and 200 abstracts are a fraction of one source’s word budget, so the reading list packs into a handful of files. Then ask which ten papers matter and add those PDFs through NotebookLM’s own Add source; source-lm does not upload PDFs. The full method.

Ask:

• “Group these papers by method and list the three most cited in each group.”
• “Which ten should I read in full for a review on [topic], and why each?”

The catch: re-running on a grown Zotero library is not reliable — the incremental cursor keys on a date field and CSL JSON’s issued is not recognised as one, so a second export means a fresh notebook or deleting the old files. Abstracts only: attachments and notes stay in Zotero.

5. A support-ticket or CRM export

Who: a founder or support lead sitting on a year of conversations. The corpus: a JSON export of tickets, conversations or deals.

This is the JSON path in its most general form. source-lm finds the array of records without configuration, works out which field is the title, which the body and which the date, and puts the rest under a Metadata heading — ids, tags, assignees and statuses stay attached to the text and come back in citations. Records are packed in order under the same 400,000-word budget, one record never split across files; an oversized record gets a file to itself, whole. The files are built in the popup, in memory, and sent to the notebook’s own origin — never to a server of ours, there is none.

Ask:

• “Which complaints repeat, and roughly how often?”
• “What did we tell customers about refunds, and did the answer change?”
• “Which feature requests came from more than one customer? Quote each.”

The catch: JSON in, not CSV and not a database dump. Re-running next quarter picks up only new records when the date field is recognised as the cursor — check the Preview file names before assuming.

6. A newsroom’s public-meeting channel

Who: a reporter covering a council, a court or an agency. The corpus: a year of hearings and press conferences on the body’s own YouTube channel, plus the pages where it posts documents.

Public bodies stream everything and index nothing. The channel button collects the newest N videos in one click, and the Link tab takes a list of URLs, one per line, for the agenda and minutes pages. Questions that meant scrubbing through video get citations instead.

Ask:

• “Every time the transport budget was discussed, with the date and who spoke.”
• “What has the commissioner said about [policy] across the year? Quote each instance.”
• “Which questions were asked at these meetings and never answered?”

The catch: NotebookLM reads YouTube’s caption track, not the video — treat an answer as a pointer to a timestamp and quote from the recording. Videos uploaded in the last 72 hours may not import, and a year of weekly meetings is more than 50 sources: the whole free notebook.

7. A reading list, one page at a time

Who: anyone whose corpus accumulates rather than arrives. The corpus: bookmarks, documentation, a handful of long reads.

The Link tab adds the page you are on, a URL you paste, or a whole list of URLs one per line. For a page NotebookLM cannot fetch, Add page as .md captures the text of the tab instead, and a right-click on a selection adds just that passage. One page, link or selection is always free and unmetered — the paid boundary is bulk, not features.

Ask:

• “What do these pages disagree about?”
• “Collect every recommendation about [topic] with the page it came from.”

The catch: a web source is the text of that one HTML page. No images, no embedded video, no nested pages, and paywalled pages are not supported.

Where the limits bite

Some of these corpora shrink and some do not — that is the planning problem. Records — chat messages, papers, tickets — pack into few, fat sources under the word budget. Videos, PDFs and web pages do not: one item is one source, and nothing merges them. So a 200-paper library is a handful of sources while an 80-lecture course is 80, on a plan that stops at 50.

The options are the same three: pack what can be packed, split by notebook (the free plan gives 100), or pay Google for a higher cap. Every per-plan number, dated and linked to Google’s own page, is in NotebookLM (Gemini Notebook) limits.

FAQ

What is NotebookLM good for? Asking questions of material you already have, and getting answers that cite it. It is strongest when the material is too large to reread — a course, an archive, a reading list — and weakest as a general-knowledge chatbot, because it answers only from the sources in the notebook.

Can it handle a whole YouTube channel or a whole Telegram chat? The content, yes; the adding is the work. A chat export packs into a few sources, so three years of messages fit comfortably. A channel is one source per video, so it fits only if the video count fits your plan’s cap. Either way, source-lm adds them in one click instead of one at a time.

Is there a limit on how many sources a notebook takes? Yes: 50 per notebook on the free plan, 100 on Plus, 300 on Pro, 500 or 600 on Ultra, and 500,000 words or 200 MB per source on every plan (Google’s numbers, checked 2026-08-21). The per-source size does not change when you pay; the count does.

Is NotebookLM free? There is a free tier — Gemini Notebook Standard — with 100 notebooks, 50 sources each, and a usage budget for chat and generated outputs that refreshes every 5 hours up to a weekly limit. It is free within an allowance, not unlimited. source-lm’s own free tier is 5 bulk actions a calendar month, with single videos, pages and links unmetered; after that it is $29, once.

Do I need an extension for any of this? No. NotebookLM accepts every source type above through its own Add source dialog, one item at a time. The extension exists for the moment “one item at a time” stops being reasonable — a playlist, an export, a reading list of fifty links.

Not affiliated with Google. NotebookLM and Gemini are Google trademarks; this is an independent extension that automates a signed-in session.