Nodus vs NotebookLM: two different models for AI-assisted research

Google renamed NotebookLM to Gemini Notebook in 2026, though almost nobody calls it that yet. Under either name it's the tool that convinced a lot of researchers that AI could be trusted with their sources at all. You pick the documents. The model answers from those documents. Citations take you back to the passage. After two years of chatbots confidently inventing references, that was a relief.

I built Nodus from the same conviction. AI gets useful for research when it works on your actual sources and when you can walk any sentence back to the page it came from. So the interesting question isn't which of the two has a better model or a longer feature list. It's what each one thinks you're left holding at the end.

NotebookLM is very good at helping you interrogate a set of sources. Nodus is built for the other half of the problem: keeping what you learned from those sources, as a structure that survives the next two years of the project. With ten papers you'll barely notice the difference. With four hundred, it's the whole thing.

A shared premise, two different research models

Put the feature lists side by side and the two look like near-twins. Both take collections of documents. Both use AI to read them. Both cite. Both do things you'd otherwise do by hand: comparing, summarising, synthesising.

The difference shows up when you ask what the basic unit is.

In NotebookLM, it's the notebook. You gather material and then run operations across it: questions, summaries, comparisons, themes, format conversions. That works because the boundary is legible. You know what's in the notebook, and so does the model. Nothing wanders in from the open web unless you invite it.

Nodus treats that collection as the raw material, not the finished object. A corpus holds documents, but it also holds the claims, passages, themes, authors, concepts and relations pulled out of them. Those don't vanish when you close the answer that produced them. They stay, they're searchable later, and they pick up new relations when new material arrives.

So: NotebookLM optimises the conversation with a set of sources. Nodus tries to turn those sources into a structure you can keep building on. Both are reasonable answers to different problems.

You feel the split immediately, in week one of a project. NotebookLM's real advantage is how little stands between importing something and thinking with it. Drop in articles, documents, web pages, videos. Start asking. Because the answers are pinned to the material you chose and the citations go back to real passages, it's a genuinely different experience from asking the same question of a general chatbot.

Say you're starting on an unfamiliar field with twenty papers. Before you read any of them properly you want the vocabulary, the recurring fights, who disagrees with whom, which three are worth a careful afternoon. NotebookLM is excellent at exactly that.

Google has also pushed well past question answering. The same set of sources can be turned into reports, study material, mind maps, audio explanations. That makes a notebook useful for teaching something as well as understanding it. Calling the product "chat with your PDFs" undersells it badly. Its real trick is turning a bounded pile of sources into a place you can think.

The trouble starts when the pile stops being bounded, and stops being temporary.

A thesis doesn't proceed in a straight line. A source you nearly discarded in month three turns out to anchor chapter four. Two arguments you filed separately turn out to be the same argument. A document you read last year quietly contradicts an interpretation you've since built on. A single archival find can force you to re-read six secondary works.

The problem isn't understanding each document. It's holding the relations between them across time.

Here's the shape of it. A historian reads a book in the first months of a project and takes away one account of economic modernisation. Eight months later, another author offers a partly incompatible explanation. Later still, an archival series complicates both. Eventually all three belong to a chapter that didn't exist as an idea when the first book was read.

An AI can rediscover that triangle, if you happen to ask the right question. But rediscovering a relation is not the same as having it be part of the project. That's the gap Nodus is built into. When an idea gets identified in a source, it persists as an object attached to its author, its work, the passage that supports it and the page that passage sits on. Later analysis connects it to other ideas, marks support or contradiction, files it inside a debate, or ties it to a theme that meant nothing when the source first arrived.

The practical failure this prevents is mundane and everybody has lived it. You remember that someone, somewhere in the pile, said the opposite of what you've just written. You can picture the paragraph. You cannot remember the author, and the phrase you'd search for is not the phrase they used. So you spend an hour of a Tuesday grepping your own library for a memory. Multiply that by three years and it stops being an annoyance and becomes a tax on the whole project. A corpus that already holds the contradiction as a typed relation hands it to you instead, along with the passage and the page.

None of that automates interpretation. It just stops so much of the structure from dissolving back into the PDFs every time you close a chat window.

From grounded answers to persistent research knowledge

Nodus works at several levels at once: documents, then sections, then passages, then ideas, then relations. Each answers a different kind of question. Whole documents tell you which works bear on a topic. Sections narrow that to a chapter or a results section. Passages give you the text you can actually quote. Ideas let you line up claims from authors who never cite each other. Relations show you patterns that stay invisible while every document sits in its own box.

Early versions of Nodus leaned hard on the idea level. It pulled claims, connected them across the corpus, and kept the original passage attached as evidence. That worked, and it also lost things. An idea stripped from its chapter can be technically true and still misleading.

Nodus 5 added the level above. It builds document profiles, represents sections, and keeps lexical and semantic indexes at document level, then uses those to decide where a finer search should go. Retrieval now runs in both directions. A broad question starts at documents and descends towards passages. A conceptual question starts in the network of ideas and climbs back to the documents underneath.

That's Nodus catching up to something NotebookLM has always been better at: treating a book as a book rather than as a bag of fragments. What happens afterwards is where the two part company. Once a passage or an idea has been identified, Nodus keeps it.

Persistence is only worth anything if the thing being persisted stays tied to evidence.

Source grounding is most of why researchers took NotebookLM seriously in the first place. The system answers, the citations let you check, and you're no longer in the position of reading a plausible paragraph with no way to find out where it came from. Nodus takes the same rule and pushes it past the chat window. A stored idea keeps its passage, and its page where the page is known. When that idea later turns up in a debate, an argument map or a research report, the route back to the text is still there.

This is also where I've been deliberately conservative about generated summaries. A document profile can tell the system that a particular book deserves a closer look, or point retrieval at the right chapter. It is not allowed to stand in for the book as evidence. For anything that matters, the system goes back to the passage. Summaries guide retrieval; passages support claims. In academic work that line matters more than it sounds like it does. A generated description of what a book argues is useful. It is not a citation, and treating it as one is how people end up quoting a book that says the opposite.

Grounding is one half of the difference. The other half is what's still there tomorrow.

NotebookLM doesn't forget everything when a chat ends, and it would be lazy to say it does. Notebooks persist, sources stay, old work is there when you come back. The real question is what form the knowledge takes. In NotebookLM, what persists is the notebook, its sources, and the interactions you ran over them. In Nodus, results of analysis can become objects in their own right. An interpretation becomes an idea. A disagreement becomes a typed relation between two ideas. Several of those relations become a debate. Places where the evidence thins out repeatedly become candidates for a research gap. A concept gets synthesised in the Dictionary from evidence scattered across dozens of authors.

The point of all that machinery is a corpus that gets richer as the project goes on rather than heavier. A book imported in year one can become relevant again in year three without you having to remember it existed. Its ideas and its evidence are still sitting in the same workspace where the new material is being analysed. Over a long project that continuity is worth more than any single impressive answer.

Zotero, discovery and the boundaries of the workspace

If your bibliography already lives in Zotero, all of this stops being abstract.

Nodus doesn't want to replace Zotero. Zotero is a better reference manager than anything I'm going to write, and it's where your metadata, your collections and your PDFs already are. Nodus reads that library and builds an analytical layer on top of it. Zotero keeps managing references. Nodus manages what those references contribute to the argument.

From a connected library you get semantic search, author analysis, idea extraction, debates, argument maps, gap analysis, Deep Research and the writing workflow, without moving a single item out of Zotero. There's also a standalone Zotero plugin that stays inside Zotero: it indexes PDF, EPUB and HTML attachments and gives you semantic and keyword search there, returning passages with page-level citations where the pages are known.

NotebookLM will happily read documents that also happen to be in Zotero. It just has no notion that Zotero exists. For anyone with a large library they've curated for years, that's a real difference rather than a philosophical one.

Both products have also moved outward, from organising what you have to finding what you don't.

NotebookLM's research features look outward through the web and the rest of Google. That fits the product: find something, look at it, pull it into the notebook. Nodus Compass goes at discovery from the bibliographic side instead. It searches across OpenAlex, Crossref, OpenAIRE, Semantic Scholar, HAL, DOAB, OAPEN, Dialnet, OpenEdition, SciELO, Unpaywall and OpenCitations at once, then normalises and deduplicates what comes back before showing it to you. Authorship, identifiers, publication type, provenance, citation data and open-access status survive the trip, and anything you keep goes into the same Library the rest of the workflow uses.

Which one you want depends on what you're hunting. Web search wins when the relevant thing could be anywhere: an institutional report, a project site, technical documentation, a news archive. Bibliographic search wins when scholarly identity is the point, when you need the DOI, the publication type, the citation graph, the open-access status, or the difference between a book, a chapter, an article and a thesis. Most projects want both at different moments, and there's no prize for picking one.

Underneath the discovery models sits a plainer disagreement about where the workspace itself should live.

NotebookLM is a Google service, with the advantages that implies: it syncs, it's on every device, it talks to the rest of Google, and you administer nothing. Nodus is a local-first desktop application. The vault databases, the shared Library and the main search indexes sit on your own machine, and the desktop app doesn't ask you to make an account. When an AI task uses an external provider, the context for that task goes to the provider you chose. Local models handle the work you'd rather not send anywhere.

I want to be precise about what that does and doesn't buy you, because this is the claim most likely to be oversold. Nodus is not automatically offline, and choosing an online model does not somehow avoid sending your text to it. If you pick Gemini, Claude or DeepSeek, the context reaches that service exactly as you'd expect. What's different is the workspace itself: it isn't hosted by me, and it doesn't stop working if I do. For plenty of people that's a shrug. For anyone with unpublished material, a private archive, an embargo, or a project they intend to still have in six years, it isn't.

Different strengths, complementary roles

None of this adds up to a ranking, and I'd distrust anyone who told you it did.

NotebookLM has built an unusually strong set of tools for turning sources into something you can hand to another person. Audio Overviews are the famous one, but the direction matters more than the example: the same collection becomes study material, a visual map, an explanation, a briefing. For teaching, revising, presenting or getting a colleague up to speed quickly, it's genuinely hard to beat. Nodus doesn't try to compete there and I don't plan to start. Its outputs point inward, at the project: Deep Research reports, the writing workshop, argument maps, idea graphs, coverage analysis, the Dictionary. If what you want this afternoon is a polished audio explainer of twelve papers, use NotebookLM. That's a real strength, not a missing checkbox.

Nodus starts to win when the question changes from what can these documents tell me now to how do I keep what I learn from them over the life of the project.

Research accumulates slowly and untidily. Arguments shift, chapters get reordered, sources acquire meanings they didn't have when you filed them, and questions surface that you couldn't have anticipated. Keeping the documents is the easy part. The hard part is keeping the structure you built around them, which is exactly the part that normally lives in your head and in a folder of notes you'll never read again.

So Nodus tries to make that structure explicit. An idea found today can join a relation found next year. A contradiction between two authors can become a node in an argument map. A concept can draw evidence from forty sources. A report written in month thirty can cite a passage marked in month four. The value isn't in any one output. It's in the continuity between hundreds of small acts of reading, searching, comparing and writing.

Which is why the useful comparison isn't "which is better". They optimise different stages. Reach for NotebookLM when you want to understand a bounded collection fast, ask grounded questions with no setup, get oriented in unfamiliar material, search the web, or turn a corpus into something explainable. Reach for Nodus when the corpus is large and long-lived, when Zotero stays the reference layer, when ideas and evidence need to persist as objects, when claims need connecting across authors, and when you need a traceable path from a finished sentence back to the original text.

	NotebookLM (Gemini Notebook)	Nodus
Basic unit	The notebook and its sources	Documents, sections, passages, ideas and typed relations
What survives a session	Sources, notebook, past interactions	All of the above, plus ideas, relations, debates, gaps and dictionary entries
Where it runs	Google's cloud, no setup, syncs everywhere	Your own machine, no account, external models optional
Reference manager	Reads files that happen to be in Zotero	Reads the Zotero library itself, read-only, plus a plugin inside Zotero
Finding new material	Web and the wider Google ecosystem	Compass, across twelve bibliographic and open-research sources
Strongest output	Audio Overviews, study material, explanations, mind maps	Deep Research reports, argument maps, idea graphs, coverage analysis, cited writing
Best at	Understanding a bounded collection quickly	Holding a growing corpus together over years
Price	Free tier, paid tiers through Google	Free and open source, AGPL-3.0-only

Nothing in that table is a knockout blow in either direction, which is the point. Read down the "best at" row and you have the whole argument.

Said plainly: NotebookLM is a place to interact with sources. Nodus is a place to build a research structure out of them. Those overlap. They aren't the same job.

Which is also why running both is not a compromise. Use NotebookLM for fast exploration of material you may not keep, for testing questions, for web research, for turning something into an explanation. Leave Zotero as the bibliographic record. Let Nodus hold the durable layer, where accepted sources, evidence, concepts and relations pile up over years. Zotero manages the bibliography, NotebookLM accelerates exploration, Nodus keeps the structure. There's no methodological virtue in forcing every stage of academic work through one application.

Seen together, the two are answers to the same starting problem, diverging on the second question. NotebookLM asks how a collection of sources can be made easier to understand, question and transform. Nodus asks how the knowledge you pull out of them stays connected and usable as the project grows. The first question leads to grounded conversation, synthesis and audio. The second leads to persistent ideas, evidence, relations and argument maps.

For a bounded collection you want to understand this week, NotebookLM is hard to beat. For four hundred sources that have to become a coherent network of arguments and evidence over four years, Nodus is built around a different kind of continuity. Which one you need depends less on the feature lists than on what you want the corpus to be at the end.

Continue with Nodus for academic research, see how Nodus works with Zotero, or learn more about the open-source and local-first project.