Citation chasing and snowballing

A citation is a claim by a domain expert that two pieces of work bear on each other. A keyword match is a claim that two documents share a string. The first is a far better predictor of relevance, and it is free — someone already did the judging.

It also routes around the vocabulary problem entirely. If a paper in your field cites a result from a field with completely different terminology, that citation is a bridge no query of yours would have crossed.

The practical consequence: in most searches, two or three rounds of snowballing from a small set of known-good papers yields more relevant work per hour than any further refinement of a database query.

1. Start from a seed set of three to five papers you are confident are on target. Fewer than three and the sample is idiosyncratic; more than five and the first round is unmanageable.
2. Open each reference list. Screen on title first, then on the sentence in the body that cites it — the citing sentence tells you what the seed paper thought that reference established, which is more informative than the title.
3. Mark each reference as relevant, uncertain, or out. Do not read yet; screen.
4. Add the relevant ones to the pool and repeat on them.

The citing sentence is the trick most people skip. A reference list is a flat list of titles; the body text around each citation is a labelled edge telling you why the link exists.

Backward snowballing has one structural limit: it can only reach work published before the seed paper. It will never find last year results, however relevant. That is what the forward pass is for.

Forward snowballing asks the opposite question: who has cited this since? It is the only search method that reliably surfaces recent work, and it is how you discover that your seed paper has been corrected, extended, contradicted or retracted.

1. Look the seed paper up in an index that exposes citing works — Google Scholar cited by, Semantic Scholar, Scopus, Web of Science, OpenAlex or Crossref Event Data.
2. Sort by date rather than by citation count. You are looking for what is new, and citation-count ranking guarantees you see the old.
3. Screen on title and on the citing context where the index exposes it.
4. Repeat on each newly relevant paper.

Coverage differs sharply between indexes, and the difference is not marginal. Run the forward pass in at least two, because a citation present in one and absent from another is common rather than exceptional.

Run backward and forward together on each round rather than exhausting one direction first. A round is: take every paper added last round, snowball it both ways, screen, add.

Stop when a full round adds nothing new. Two or three rounds is typical for a focused question. If round four is still producing relevant work, that usually means the seed set was too narrow — the search is finding a neighbouring literature it should have started from.

• Work nobody has cited yet. A paper from three months ago has no forward citations and will appear in no reference list. Only a database query or venue scan finds it.
• Isolated literatures. If a relevant body of work has no citation path to your seed set — common across disciplinary boundaries — snowballing cannot reach it by construction.
• Uncited grey literature. Reports, standards and theses are cited far less consistently than papers, so they are under-represented in any citation graph.
• Whatever the seed set is biased toward. Snowballing amplifies the seed. A seed set drawn from one school of thought returns that school of thought.

This is why snowballing is a component of a search and not the whole of one. It is the strongest single path and it has a blind spot that only querying and venue scanning cover.