Citation Chaining and Snowballing a Literature Search

Citation chaining (also called snowballing) is the practice of finding new relevant papers by following citation links from a paper you already know is relevant, rather than by running another keyword search. It works in two directions: backward, through a paper’s own reference list, and forward, through the later papers that have cited it. Used together and repeated across newly found papers, the two directions “snowball” outward from a small starting set into a much fuller picture of a literature — and they routinely surface work that a keyword search misses entirely, because a genuinely relevant paper may use different terminology, sit in an adjacent discipline’s vocabulary, or predate the specific term you searched for.

Backward Chaining: Following a Paper’s Reference List

Backward chaining (sometimes called backward snowballing or reference-list checking) means reading through the bibliography of a paper you already trust and pulling out the citations that look relevant to your question. It is the fastest way to find the intellectual foundations underneath a paper — the studies, methods, and definitions it built on.

• No special tool required. Every paper’s own reference list is the source. If you already have the paper in a reference manager, its imported bibliography is often faster to scan than the PDF’s own reference section.
• Works best on well-cited review articles and highly-relevant primary studies — a thin or narrowly-scoped paper’s reference list will chain you into a thin set of leads.
• It moves you backward in time only. Backward chaining alone cannot surface anything published after your seed paper — that gap is exactly what forward chaining exists to close.

Forward Chaining: Finding Who Has Cited This Paper Since

Forward chaining (forward snowballing, or “cited-by” searching) runs the opposite direction: starting from a known relevant paper, find everything published since that has cited it. This is how you catch a field’s more recent developments, replications, critiques, and applications of an older foundational study — work a backward-only search would never reach.

Forward chaining requires a citation index, because nothing on the seed paper itself tells you who later cited it. The three practical options, and what actually differs between them:

• Scopus (Elsevier) shows a “Cited by” count on every document record, linking straight through to the citing documents. It’s a subscription database, and its citation counts reflect only the literature Scopus itself indexes.
• Web of Science (Clarivate) shows a “Times Cited” count on every record within its Core Collection, with the same kind of link through to citing records. Also subscription-only, and — like Scopus — its count reflects its own indexing scope, not the literature as a whole.
• Google Scholar puts a free, no-login “Cited by” link on every result. Its crawl-based index is broader than either subscription database (it picks up preprints, theses, and repository copies they may not), which is exactly why forward counts differ across all three tools for the same paper — see our guide to using Google Scholar for a literature review for what it’s good and not good at as a primary search tool.

Because coverage differs, running forward chaining through only one of these will undercount. A paper that looks lightly cited in Scopus may show a longer citing list in Google Scholar simply because Scopus never indexed some of the venues that cited it.

Visual, Graph-Based Chaining: Connected Papers and Research Rabbit

Two purpose-built tools turn chaining into an explorable graph instead of a list, which is genuinely useful for orienting yourself in an unfamiliar literature quickly — but they don’t do the same thing underneath, and conflating them leads to under-searching.

• Connected Papers builds a one-shot visual graph around a seed paper, with “Prior Works” and “Derivative Works” panels alongside it. Both panels are curated subsets drawn from the same underlying similarity computation (co-citation and bibliographic coupling) — not an independent, exhaustive backward/forward citation trace. Prior Works surfaces commonly-cited-by-the-set ancestors; Derivative Works surfaces commonly-citing-the-set descendants. That’s a real, useful signal, but it is filtered by similarity, not a complete reference list or a complete citing-papers list.
• Research Rabbit instead offers dedicated “Earlier Work” and “Later Work” modes built to function as a genuine exhaustive backward/forward citation trace from a selected paper or collection — closer in spirit to running Scopus/Web of Science chaining manually, but inside a visual, collection-based interface.

See our Connected Papers alternatives comparison for how these and other citation-mapping tools stack up on cost and coverage. Worth keeping separate in your own head from citation network analysis — co-citation and bibliographic-coupling mapping used to understand how a field’s literature is structured, which is a research-assessment technique, not a way to find more papers for your own review.

Combining Chaining With a Database Search — and Documenting It

Chaining is a supplement to a systematic database search, not a replacement for one. A single, carefully-scoped seed set biases what you’ll ever chain into — you only find what’s connected to what you started with. The standard practice, reflected in the Cochrane Handbook‘s guidance on supplementary search methods, is to run a proper database search first (see our guide to Boolean search operators for literature searching and PubMed’s advanced search builder), then chain from the studies that search returns as a way of catching what a keyword strategy structurally can’t.

If the search is feeding into a systematic or scoping review, this distinction has to be reportable, not just practiced. PRISMA 2020‘s flow diagram separates records identified through database and register searching from records identified through other methods — citation chaining belongs in that second stream. Track, as you go, how many records each chaining pass contributed and how many survived screening, the same way you’d track a database search’s yield; reconstructing that after the fact is far harder than logging it as you work. The same “other methods” logic applies to hand-searching and grey literature — see our guide on searching grey literature for a systematic review for the parallel case.

When to Stop: Recognizing Saturation

Chaining doesn’t have a natural endpoint the way a database search does (a database search ends when the query has been run) — left alone, backward and forward links can be chased indefinitely. The practical stopping rule used across evidence-synthesis methodology is saturation: keep chaining from newly-added relevant papers until a full round adds no new relevant papers, then stop. In practice that means:

• Chain from your original seed set first, add whatever is relevant, then chain from those new additions too — snowballing is iterative, not a single pass.
• Stop a given branch once it stops returning anything new and relevant, rather than an arbitrary number of “rounds.” A branch that saturates after one round and a branch that takes three both end the same way: no new relevant records.
• For a scoping or rapid review with a tighter timeline, it’s defensible to cap chaining depth deliberately (e.g., one round from the seed set only) — but say so explicitly in your methods section rather than letting an unstated time limit look like saturation.

Common Pitfalls

• Relying on a single tool’s forward-citation count as if it were definitive. Scopus, Web of Science, and Google Scholar will each report a different number for the same paper, because each indexes a different slice of the literature — see the coverage note above.
• Treating Connected Papers’ Prior/Derivative Works as a complete reference or citing-paper list. They’re similarity-filtered curated subsets, not an exhaustive trace — use Research Rabbit’s Earlier/Later Work modes or a citation database directly when you need genuine exhaustiveness.
• Chasing a very recently published seed paper forward and finding almost nothing. That’s usually indexing lag, not evidence the paper is uncited — recheck after a few months, or lean more heavily on backward chaining and database searching for very new work.
• Not logging chaining yield as you go, which turns “how many records came from citation searching” into a reconstruction problem right when you’re writing the PRISMA flow diagram.

Frequently Asked Questions

What’s the actual difference between backward and forward citation chaining?
Backward chaining reads a paper’s own reference list to find earlier related work; forward chaining uses a citation index to find later papers that cited it. Together they cover both directions in time from a single seed paper.

Can citation chaining replace a database search for a systematic review?
No. Reviewer guidance treats it as a supplementary method that catches what a database search structurally misses, run alongside a properly scoped database/register search, not instead of one.

Which tool should I use for forward chaining — Scopus, Web of Science, or Google Scholar?
Ideally more than one, since each indexes a different slice of the literature and will return a different citing-paper count for the same source. Use whichever subscription database your institution provides for the systematic count, and treat Google Scholar as a free supplement that often catches preprints and repository copies the others miss.

Are Connected Papers and Research Rabbit interchangeable for citation chaining?
Not quite. Connected Papers’ Prior/Derivative Works panels are similarity-filtered subsets of a one-shot graph; Research Rabbit’s Earlier/Later Work modes are built to function as an exhaustive backward/forward trace. Pick based on whether you need fast visual orientation or a complete count.

How do I report citation chaining in a PRISMA flow diagram?
As a separate identification stream from database/register searching — PRISMA 2020’s flow diagram distinguishes the two. Log how many records each chaining pass added and how many survived screening as you go, rather than reconstructing it afterward.