
NotebookLM Sources Exporter: Save Every Linked Doc From A Notebook
NotebookLM has no export button. Here is the sources exporter pipeline that pulls every PDF, audio overview, transcript, and note into a local archive.
Why NotebookLM has no export button
NotebookLM is the only product in Google's lineup that treats your source material as a sealed workspace. You drop in PDFs, YouTube links, Google Docs, and pasted text, and the model answers only against that bounded set. The citations point back to your own files, which is the whole reason people trust it. The problem is leaving. Google has not shipped a sources exporter, a bulk download, or a public API, so the moment you want your material anywhere else it stays locked behind the workspace chrome.
That matters more than it looks. A notebook with forty PDFs, eight audio overviews, and a few hundred saved notes is a research asset, and losing access to it means re-uploading the whole stack to rebuild the same context somewhere else. The team ships new features monthly, and every refactor risks shifting where sources live inside the DOM. A reliable exporter is less about convenience and more about owning the archive you built over months of work.
The two shapes of a NotebookLM export
A full export has two halves, and conflating them is the mistake that breaks most scripts. The first half is the sources themselves, meaning the original uploaded PDFs, the pasted text blocks, the audio overviews Google generated for you, and the YouTube transcripts it fetched. The second half is the derived material, meaning your notes, your chat log, and the Studio outputs like mind maps and briefing docs. Each half needs a different route out.
Sources leave through the browser cache and the Drive picker. Derived material leaves through the DOM, because Google renders it client side and never writes it to a public endpoint. If your script treats both the same way you will end up with half the archive and no idea what is missing. Our earlier guide on the reading side of the same workflow covers the output shape, so this post is about the inbound capture instead.
A sources exporter that holds up
The reliable route is a short pipeline, not a single click. Open the notebook in Chrome with DevTools docked. In the Network panel, filter by the fife and googleusercontent hosts while you click each source in turn. Every uploaded PDF streams back from a signed URL, and every audio overview resolves to an .ogg or .m4a on fife.googleusercontent.com. Right click, save as, and name the file after the source title so your archive stays searchable later.
For YouTube sources, grab the video ID from the chip, run it through yt-dlp --write-sub --skip-download, and store the auto captions as a sibling .vtt. Google already pulled that transcript to feed the model, and keeping a local copy means your archive matches what NotebookLM read. Audio overviews have a transcript toggle inside the workspace itself, and the rendered transcript pastes as clean markdown. Save it next to the audio file under the same stem.
A short shell loop keeps the folder layout sane. One directory per notebook, three subfolders named sources, audio, and notes, and a top level manifest.md with the notebook title, the export date in ISO form, and a one line note on what the source set covers. The pattern I described in building a personal knowledge base from AI conversations applies almost verbatim, with the notebook acting as the folder root. Keep filenames lowercase and dash separated so grep still works six months later.
Keeping notes, chat, and Studio outputs linked
The derived material is the half that rots fastest, because Google never promised stability for the UI classes you need to select against. The cleanest capture is a browser extension that copies the visible Shadow DOM as markdown. Install one of the general purpose ones, load the notebook, and run the capture once per panel: notes, chat log, mind map outline, and briefing doc. Each gets its own .md under notes with the panel name as the filename.
For the chat log specifically, scroll to the top of the thread before you capture. The panel virtualises older messages, and grabbing from the middle will silently cut the start of the conversation. If the thread is long enough that even scrolling misses rows, the same trick covered in exporting long AI chat histories to local files works here, because the underlying rendering primitives are shared across Google's chat surfaces. Save the chat log with a short header that names the source set it referenced, which keeps it readable when the notebook evolves later.
Studio outputs are the nicest surprise in the export. A briefing doc rendered against your source set is a cleaner research artifact than any single note, and it pastes as well formed markdown with its own inline citations. Treat each Studio output as a standalone .md under notes, and link them from the manifest. When you compare what NotebookLM produced against what the general models produced, the workflow in comparing ChatGPT, Claude, and Gemini answers handles the three column read without extra tooling.
Reading the archive back cleanly
An exported notebook is only useful if you read it the way you read a real document. The sources sit as plain PDFs, the notes and chat sit as markdown, and the manifest is the glue. Open the manifest in a reader that renders inline links and keeps anchors clickable, and the whole folder behaves like a small research site. Prism MD is tuned for exactly this shape, with coral accented anchors, a sticky outline on the left, and KaTeX math for the notes that quote formulas back verbatim from the sources.
Three small habits keep the archive alive rather than decaying into a flat folder of dead files. They cost almost nothing per export and compound every time you reopen the notebook later. The list below is short on purpose so it sticks through the first few manual exports. After a month you will not need to look at it again.
- Overwrite the same canonical folder each time you refresh the notebook instead of versioning filenames.
- Stamp the manifest with the export date in ISO form at the top, not the bottom.
- Keep the sources folder read only on disk so a stray edit never desynchronises the archive from the live notebook.
Read every NotebookLM export in a reader built for cited research
Free to start — no credit card.
FAQ
Is there an official NotebookLM API or exporter?
No public API has shipped at the time of writing. Google has added mobile apps and Workspace hooks, but the sources panel still has no export button and no bulk download. The browser based route is the only reliable path, which is why owning a short script matters more here than for products with first party exports.
Do audio overviews export as audio or as text?
Both, if you want them. The overview plays from a signed URL you can save directly, and the workspace itself exposes a transcript toggle that pastes as markdown. Keeping the audio and the transcript together under the same stem is cheap and makes the archive searchable.
Will the citations still resolve after I export?
The Source 3 style chips break the moment you leave the workspace, but a small find and replace during export turns them into real markdown anchors like [S3 p.12](#source-3). After that the navigation survives in any reader that honours anchor links. The page numbers remain accurate as long as the original PDFs are untouched.
How often should I refresh the export?
Every time the source set changes meaningfully, which for an active notebook usually lands around once a week. Overwrite the same folder rather than versioning, and lean on the ISO date stamp in the manifest to track drift. The point of the archive is a fresh reading copy, not a museum of old states.
Related reading
Ready to read your own AI documents?
Open ChatGPT, Claude, Gemini, or any markdown file in the reader built for the way models write.
- ✓Renders code, math & Mermaid out of the box
- ✓Works offline once you've opened a doc
- ✓Free forever for personal reading


