Keyword search has a structural blind spot. It finds papers that use the same words you do. But the paper that would change your framing may describe the same idea with a completely different vocabulary — a different field’s term for the same mechanism, or the name the idea had before yours became standard. No amount of clever query writing surfaces it, because the words are not there.
Literature mapping fixes this by following a different signal: citations. Authors cite the work they built on, regardless of what words they use. So if you walk the citation graph outward from a paper you already trust, you reach relevant work through a path that does not depend on shared vocabulary.
The two directions
A literature map has exactly two moves, and it is worth being precise about them because they answer different questions.
Backward (references). Look at what a paper cites. This answers “what did this build on?” — it takes you toward the foundational work, and it is bounded: a paper has a finite reference list that never grows.
Forward (citations). Look at what has cited a paper since. This answers “what happened next?” — it takes you toward current work, including critiques and failed replications. It is unbounded and grows over time, which is why a map you built two years ago is already stale in the forward direction.
Most people do the backward pass instinctively and skip the forward one. The forward pass is usually where the surprises are, because that is where you find out that the result you were about to build on was qualified, extended, or contradicted.
The method
1. Pick seeds deliberately
Start with two to five papers you are confident are central. Good seeds are:
- Recent and well-cited — gives you a rich backward pass.
- A survey or review, if a good one exists — its reference list is a curated map someone else already drew.
- Deliberately varied — if all your seeds come from one lab or one venue, your map inherits that community’s boundaries and you will mistake a subfield for the field.
That last point does most of the work. Two seeds from different communities produce a far better map than five from the same one.
2. Do one full backward pass
For each seed, go through the reference list and mark anything that looks load-bearing — not everything, just what the paper actually depends on. You are looking for names that appear across several of your seeds. A paper cited by four of your five seeds is almost certainly something you need to read, even if the title means nothing to you.
This co-citation signal is the single most useful thing a map gives you, and it costs nothing to compute: it is just “which papers keep showing up”.
3. Do one full forward pass
For each seed, look at what has cited it. Sort by citation count to find the influential follow-ups, then sort by date to find what is happening now. Both matter, and they give different lists.
Read the titles first and only open what looks relevant. The forward list for a well-known paper can run to thousands of entries; you are skimming for shape, not reading.
4. Expand from the new nodes, then stop
Take the papers that appeared in step 2 or 3 and repeat once. One or two rounds is usually enough. Beyond that, you are mostly re-finding papers you already have.
The stopping rule is concrete: when a full round produces nothing you had not already seen, you are done. In the systematic-review literature this is called reaching saturation, and it is a real signal rather than a feeling. If you are still finding new central papers on round three, your seeds were probably too narrow — go back and add a seed from a different community.
5. Write down why each paper is on the map
This is the step people skip, and it is the one that makes the map useful three months later. For each paper you keep, note in one line why it is there: “the original result”, “shows it fails for small samples”, “the survey everyone cites”. Without this, you end up with a folder of PDFs and no memory of what connected them.
What a good map looks like
You are not aiming for a pretty diagram. A finished map should let you answer four questions without opening anything:
- What are the two or three foundational papers everything traces back to?
- What are the distinct clusters or schools of thought, and how do they disagree?
- What is the most recent credible work?
- Where is the gap you are going to write into?
If your map cannot answer question 4, it is not finished — usually because you have only done the backward pass.
Tools
You can do all of this manually with Google Scholar’s “Cited by” link and a spreadsheet. It works. It is just slow, and the manual version tends to lose the “why is this here” annotations.
Several tools automate the graph walk. Connected Papers builds a similarity graph around one or more origin papers — it arranges papers by co-citation and bibliographic coupling rather than drawing a literal citation tree. ResearchRabbit and Litmaps are not two independent choices: Litmaps acquired ResearchRabbit in 2025 and both now run under the same company, so picking between them is choosing an interface, not a vendor. ResearchRabbit turns the same walk into a growing collection — earlier work, later work, similar work — with up to 50 seed articles on its free tier. Litmaps keeps a saved map and can monitor it, emailing you when newly published work connects to it; on its free plan that email is a monthly summary. Inciteful is good at multi-seed network analysis.
Litlas approaches it from the library side: you search across multiple sources, open the citation graph around any paper to walk backwards and forwards, and save what matters into a library where each paper carries your own note — which is where step 5 lives. Citations export as BibTeX or RIS when you start writing. The free plan is five searches and five graph views a day — enough to open a small map, not enough to run several rounds of expansion in one sitting.
Two honest caveats about tools in general, including ours. Citation data is incomplete — coverage differs by field and by source, and very recent papers are always under-counted because the citations have not accumulated yet. And a tool can show you the graph, but deciding which node matters is still your judgement. The map is an instrument, not an answer.
The common failure
The most common way literature mapping goes wrong is stopping after the backward pass, because it feels complete — you followed every reference, so surely you are done. But the backward pass can only ever show you the past. If you build on a result without checking what happened to it afterwards, you find out at review time.
Do the forward pass. It takes twenty minutes and it is the half that protects you.