How to Use Research Rabbit AI: A Practical Walkthrough

Why Keyword Search Leaves You with Bad Papers (and What Citation Mapping Fixes)

You type a keyword into Google Scholar, hit search, and get back a flat list of thousands of results with no sense of how any of them relate to each other. Which paper is foundational? Which is cutting-edge? Which ones actually matter? A keyword search can't tell you — and that's how otherwise careful researchers end up citing mediocre papers they happened to find first.

This is the problem citation-based discovery was built to solve. Instead of matching keywords, tools like Research Rabbit start with one or more "seed papers" you already trust and map how the surrounding literature connects. The tool analyzes a paper's bibliography, the papers that cite it, and shared co-authors to recommend related work, "turning one paper into a map of ideas," as the company puts it. Because these recommendations rest on human-placed citations rather than an AI's guess at semantic similarity, reviewers argue citation-based tools like Research Rabbit, Litmaps, and Connected Papers deliver more reliable results than semantic AI tools such as Consensus or Elicit.

The catch, historically, was usability. A peer-reviewed 2023 review by University of Ottawa librarians noted that many features felt overwhelming and demanded a significant learning investment. That changed on October 30, 2025, when Research Rabbit shipped its biggest update in years: a cleaner, iterative "rabbit hole" flow with a more configurable citation graph. Librarian Aaron Tay, who had rarely used the tool seriously before, found the revamped interface approachable enough to make it a genuine entry point for iterative citation chaining.

The business side shifted too. In November 2025, Research Rabbit was acquired by Litmaps and re-released as a freemium product — a notable change for a tool that had been 100% free. The free tier now caps you at fewer than 50 input papers, fewer than 5 authors, and a single project, with an optional RR+ premium tier unlocking advanced filters and exports. Older guides calling it "forever free" no longer tell the whole story, though the free tier remains a practical starting point.

Citation mapping isn't magic, and it has limits — some co-cited papers turn out to be irrelevant (an ecology paper citing the R programming language paper, for example), and neither Research Rabbit nor its competitors are designed for systematic reviews. But for exploratory literature discovery, it answers the question keyword search never can: not just "what exists," but "how does it all connect?" And once you've mapped the field, the surrounding workflow — summarizing, note-taking, synthesizing — is exactly the kind of work agent skills can automate; you can browse and test free ones on SkillzBench.

The Core Workflow: From Seed Papers to a Visual Citation Map

One good paper can lead you to dozens more — but only if you can see the connections. Research Rabbit turns a handful of starting papers into a visual citation map, and its official guide breaks the process into six steps that are easy to follow once you know what each screen is telling you.

Start by creating a project, which is simply the workspace for your topic — the official guide notes the name doesn't need to be perfect. Then run a first search using a keyword, topic, title, or DOI. Your goal here isn't exhaustive coverage; it's finding a few relevant starting papers.

Next comes the most important choice: selecting seed papers. The guide recommends picking at least one well-cited foundational paper and one more recent paper, and warns that good papers aren't always on the first page of results. A peer-reviewed review adds a useful rule of thumb: the more publications you add to a collection, the better the recommendations become (Cole & Boutet, JCHLA).

Once you've saved your seeds, the citation map appears. Reading it takes a minute to learn:

• The default graph plots publication year on the x-axis and citation count on the y-axis, so higher nodes are more cited and nodes further right are more recent (Aaron Tay's hands-on review).
• Node colors carry meaning: green means the paper is in your collection, blue means it isn't (darker blue = more recent), and red marks authors.
• Papers cluster into visible subtopics — click any node to read its abstract without leaving the map.
• Arrows show citation direction, and both axes and node size are customizable in the 2025 version.

From there, the tool offers three recommendation modes, and knowing which to use matters. Similar leans on semantic and title/abstract similarity; References follows backward links (what a paper cites); Citations follows forward links (who cites it). A helpful mental model from Tay: the mode is a lens, the iteration is a checkpoint, and the branch is a fork. Only clicking "Search" creates a new iteration — switching modes or saving papers doesn't — and input papers aren't added to collections automatically, so save them explicitly.

Then iterate. Add better papers, re-run the search, and let every saved checkpoint let you hop back if a line of inquiry dead-ends. As the vendor guide puts it, the more thoughtfully you interact with the tool, the better your results become. Discovery tools like this surface the papers; if you want to automate the surrounding research workflow, agent skills can handle the rest — you can browse and test free ones on SkillzBench.

The Rabbit Hole Mental Model: Iterations, Branches, and Pitfalls to Avoid

The 2025 revamp replaced Research Rabbit's old linear exploration with a system of "rabbit holes," and librarian Aaron Tay's hands-on review offers a mental model that makes the whole thing click: think of each Mode as a lens, each Iteration as a checkpoint, and each Branch as a fork in your search history. Once that framework is in your head, the interface stops feeling like a maze and starts feeling like a tidy record of every decision you've made.

The three modes work like different lenses on the same seed papers. "Similar" leans on semantic matching of titles and abstracts, while "References" pulls backward links and "Citations" pulls forward links through the citation network. Switching between them reframes what the tool recommends, but — and this is the first documented pitfall — it doesn't change your position in the rabbit hole.

That's because Tay's review found that only clicking the "Search" button (bottom right) creates a new iteration. Switching modes or saving papers to collections does nothing to your checkpoint history. Newcomers often assume their selections have been "committed" when they haven't; nothing is recorded until you explicitly run the search.

Two more pitfalls deserve a warning label:

• Input papers aren't auto-saved. Papers you select as inputs for the next search are not automatically added to your collections — you must explicitly use "save to..." or they exist only inside that iteration.
• Searching from an older iteration starts a new branch and wipes out "future iterations." Checkpoint history makes this recoverable in spirit, but the forward path you'd built is gone.
• Seed quality compounds. The official guide recommends at least one well-cited foundational paper and one recent paper, and a peer-reviewed review notes that more publications in a collection yield better recommendations — so weak seeds weaken every downstream iteration.

The payoff for learning this system is what EffortlessAcademic calls the standout feature: every search iteration is saved, letting you "hop back" to earlier steps. This search tracing is what prevents the classic rabbit-hole failure mode of getting three layers deep into citations with no idea how you got there. The old version's single linear exploration path was a known weakness in a 2023 peer-reviewed review; the 2025 iteration-and-branch structure directly addresses it.

For teams building research workflows around tools like this, the same checkpoint discipline applies to the AI agents doing the surrounding work. SkillzBench's free directory lets you browse agent skills and see which ones are actually verified before you wire them into your workflow — worth a look once your rabbit hole has surfaced the papers and you need help processing them.

Working the Map: Per-Paper Actions, Zotero, and When to Iterate

A map only becomes useful when you work it — and this is where Research Rabbit turns discovery into actual understanding. Once your citation graph exists, each paper node supports a set of actions that the official guide calls the step where "reading becomes actual understanding."

Click any node to read its abstract, then add notes, save the paper to a color-coded collection, and explore outward through Similar, References, or Cited By views. Color coding matters: green nodes are in your collection, blue are not (darker blue means more recent), and red marks authors, per a peer-reviewed review by University of Ottawa librarians. Switching to the People view lets you click an author to see their relevant publications, plus "These authors" and "Suggested authors" panels for exploring researchers rather than papers.

Getting papers out is straightforward. Research Rabbit offers direct Zotero integration for exporting papers, with free one-way import; shareable map URLs, email invites for collaborators, BibTeX/RIS exports, and alerts on new papers round out the workflow, according to the HKUST library guide. One caveat from a hands-on review of the 2025 revamp: two-way Zotero sync from the old version isn't yet available in the new interface.

Iteration is where the tool earns its keep. The vendor's guidance is explicit: add at least one well-cited foundational paper and one recent paper as seeds, and remember that more publications in a collection produce better recommendations. In the 2025 "rabbit hole" system, only clicking Search creates a new iteration checkpoint — switching modes or saving to collections doesn't — and searching from an older iteration starts a new branch that wipes out future iterations.

A few honest limitations worth knowing before you invest hours:

• Co-cited papers can be irrelevant — reviewers note ecology papers citing the R programming language paper as a typical false connection.
• The free tier caps input papers at fewer than 50, authors at fewer than 5, and allows one project; advanced filters and exports require the RR+ premium tier.
• It's a discovery tool, not a systematic-review tool — no screening, PRISMA support, or analysis, just mapping.
• Free-tier users lose sharing, and there's no author search for finding similar researchers.

The November 2025 Litmaps acquisition shifted Research Rabbit from free-forever to freemium, so older "completely free" claims now conflict with current reality — treat the free tier as a capable entry point, not a full replacement for paid tools.

That last point matters for anyone building research into a broader workflow: Research Rabbit shows you where to go, but you still make the journey yourself. The surrounding work — extracting findings, summarizing papers, feeding them into reports — is exactly the kind of repetitive work agent skills can automate. SkillzBench is building a free directory where you can browse, verify, and benchmark those skills, with a free MCP connector to wire them into your agents; join the waitlist to be first in line when it launches.

Beyond Discovery: Automating the Rest of Your Research Workflow

Here's an uncomfortable truth about your literature discovery workflow: the tool that finds the papers still leaves you doing everything else by hand. One reviewer put it perfectly — Research Rabbit is "a bit like having a great map — it can show you where to go, but you still have to make the journey yourself" (Elephas review).

That's by design. Research Rabbit is a discovery tool, not a research assistant. A comparison of citation-mapping tools notes that neither Research Rabbit nor Connected Papers is designed for systematic reviews — both stop at surfacing papers. Everything downstream of that remains manual:

• Screening — reading abstracts and deciding which of the 50+ papers in your map actually matter to your question
• Summarizing — extracting methods, findings, and limitations from each paper you keep
• Synthesis — comparing papers against each other and identifying gaps, which is the whole point of mapping a field in the first place
• Reporting — writing up what you found, formatted for whoever needs it

The tool does help you get organized before that work begins. The official guide recommends saving papers to color-coded collections and adding notes, because "this is what turns reading into actual understanding." You can export your collection in BibTeX or RIS formats and push papers into Zotero. But exports are handoffs, not automation — you're handing a pile of PDFs and citations to your future self.

This is where agent skills come in. If you're already using an AI agent — Claude Code, Cursor, a chat-based assistant — skills are instruction packages that teach the agent to handle specific repeatable tasks like screening abstracts against criteria, generating structured summaries, or drafting a literature synthesis from your exported citations. The discovery tool shows you where to go; the skills automate the journey.

The catch is that skill quality varies wildly. Published research found that roughly 26% of agent skills carry at least one vulnerability, so grabbing the first skill you find on GitHub is a genuine risk. That's the problem SkillzBench (skillzbench.com) is building against: a free, independent directory that crawls and dedupes skills, pins every version to a commit SHA, runs security scans and license checks, and benchmarks each skill in a sandbox to publish its measured Lift in percentage points. A free MCP connector will let you plug the skills you choose directly into your agent.

SkillzBench is pre-launch — the waitlist is live now, and everything on the platform will be free when it ships. In the meantime, treat your Research Rabbit exports as the input to a workflow you can delegate. The map is half the job; the other half is increasingly automatable.