How to Find Related Papers With Connected Papers
Connected Papers takes one seed paper and draws a graph of related work, positioning each node by similarity rather than by citation count. It answers a question a search engine cannot: what else sits in this neighborhood, including papers that do not cite or get cited by the seed.
That last part is the useful bit. Citation chasing finds ancestors. Connected Papers finds siblings.
What does the graph actually show
Nodes are papers; edges are similarity based on citation relationships and content. Papers by the same authors sit together. Older nodes cluster in one direction and newer ones in another. If you hover a node, you get its year and citation count.
The axes approximate prior works on one side and derivative works on the other, which lets you answer two questions at a glance: what led here, and what followed.
How do you pick a good seed paper
The seed defines the neighborhood. A survey paper gives a wide but shallow map. A specific method paper gives a narrow, deep map. For a new field, start with a survey and then re-seed on the most important method you find.
Wrong seeds are usually tangential-paper mistakes: you seed with something that merely mentions the topic in an introduction and get a graph of an unrelated literature.
| Seed type | Graph shape |
|---|---|
| Survey | Wide, many clusters |
| Seminal method paper | Deep, dense center |
| Recent paper | Sparse prior side |
| Tangential paper | Unrelated clusters |
If the graph looks like noise, the seed is wrong. Re-seed before you spend an hour clicking nodes.
How do you turn a graph into a reading list
Hover the most central nodes and the recent high-citation nodes. Central means adjacent to many others, which usually means foundational. Recent high-citation means the current consensus.
Export the list to a citation manager, then read in this order: one survey, two seminal papers, two recent papers. Five papers is enough to hold a conversation in most fields.
The graph also shows isolated clusters, which are often adjacent fields where a method was applied differently. Those are the source of cross-domain ideas.
What are the limits
It depends on having citation metadata, so very new papers and non-indexed venues are missing or poorly connected. It does not tell you whether a paper is good, only that it is nearby. And it has no full-text search or summarization, so you still read the papers.
For the reading itself, a PDF reader with citations such as ChatPDF handles the next step: ask the paper what its contribution is and where the numbers are.
Who should skip it
Engineering work where the relevant knowledge is a documentation page rather than a paper, and any field where the literature is mostly preprints that are not well indexed. Also not a substitute for a systematic review protocol, which needs reproducible searches, not a graph.
Quick Summary
- Connected Papers graphs related work from a single seed paper.
- Seed with a survey for breadth, a method paper for depth.
- Prior works and derivatives sit on opposite sides of the layout.
- Read five papers: a survey, two seminal, two recent.
- It ranks proximity, not quality, and misses unindexed preprints.
