Snowball literature search: backward and forward citation chasing

The short version of citation chasing is familiar: one relevant paper, ten titles from its bibliography, done. That can be enough for a short essay. In a longer thesis it often leaves the methods chapter silent about how the reference base was built, and the citations can sit inside one closed group. This guide covers both directions, a start set that is not a single research group, and a record a supervisor can follow.

What snowballing is

Snowballing (also called citation chasing) starts from one or more papers and moves through citations. The name is a snowball picking up snow as it rolls: each round adds more texts.

Claes Wohlin wrote the procedure for systematic literature studies in software engineering (Wohlin 2014). Two directions:

• Backward snowballing. You work through the reference list and decide which cited texts belong to your question. You move to older work the anchor builds on.
• Forward snowballing. You look up who cited the anchor. Wohlin uses Google Scholar for this. You move to later work that extends, supports, or contradicts it.

This is not snowball sampling. That term belongs to recruiting participants through referrals. Here the units are documents, not people.

Knowing only the backward direction is the usual gap. Papers published after the anchor are not in its reference list.

When to use it

Snowballing is a search strategy, not a substitute for reading. Wohlin treats it as a possible first strategy for a systematic literature study once a start set exists. For a bachelor’s or master’s thesis, the part that transfers is the same: a justified start set, both directions, and a written trace. A keyword search in a subject database is still the usual way to build that start set. For a comparison of databases, see Web of Science, Scopus, and PubMed.

Four situations where citation chasing adds something a keyword list does not:

• A narrow topic. A keyword search can return a long list of which only a few items actually bear on the question. Moving outward from those few often helps more than another synonym string.
• More than one field. Wohlin notes that different communities name the same things differently. A search locked to one vocabulary misses the other. Citations follow the reference, not the keyword.
• A review as an anchor. A review has already screened the field. You do not adopt its list wholesale. You apply your own criteria.
• Later work. If you want to use an older foundational text, forward snowballing shows who cited it afterwards. Whether the text was retracted is a separate check: spotting retracted papers.

If the topic is broad, one citation chain is not enough. Find the start set with keywords first, which is what Wohlin did in Google Scholar.

Backward snowballing, step by step

Backward snowballing needs the reference list of a paper you have decided to include. Wohlin requires that inclusion be decided on the full text before you snowball from that paper. Otherwise you later have to remove everything that hung off a paper you should have excluded.

Choose the start set

Wohlin describes properties of a good start set, not a fixed count. The size depends on how broad the area is. A narrow question needs fewer papers than a wide one. The set should not be too small, and it should cover different communities, publishers, years, and authors. Otherwise you stay inside clusters that cite each other and never mention the neighboring debate. Keywords come from the research question, including synonyms. If the search is too broad, he suggests taking relevant papers that are highly cited in the area. He does not give a citation threshold that applies to every field. Counts are not comparable across disciplines.

Useful anchors:

1. A review that still matches the debate you need. In a fast-moving field that is often a recent review. In history, law, much qualitative research, and theory, the review that matters may be older.
2. Papers that are highly cited in that field, when a keyword search is too wide. Highly cited means highly cited there, not a universal minimum.
3. A dissertation on the same question. Its literature chapter can range wider than the reference list of a single empirical article. How many sources a dissertation, a master’s thesis, or a bachelor’s thesis should carry depends on the discipline, the level, and the supervisor. There is no fixed range. You still judge every item against your own question.

A single empirical article from one research group is a narrow start. Its references can stay inside that conversation.

Screen the reference list

Wohlin screens in passes, not as a gut decision on each title:

1. Drop items that fail your basic criteria: language, publication type, and a year limit only if the question needs one.
2. Drop items you already screened in an earlier round.
3. Decide the rest from the title, from how the anchor uses them, and, when that is not enough, from the full text.

Record every decision: include or exclude, and why. You will need that reason in the methods chapter.

Four questions per source

1. Does it bear on the question? Not merely: it mentions the topic.
2. Can you see what it did or argued? That applies to quantitative, qualitative, and theoretical work. Peer review is one signal, not a requirement in every field. Grey literature needs a reason.
3. Is the age right for the question? In a fast-moving field, check whether later work has superseded the text. In history, law, much qualitative research, and theory, an older text is often the object of study or still authoritative. There is no universal year cutoff. If you set one, record it as a decision for this question.
4. Do you have the full text? Without it you cannot check the passage yourself. Citing it only through another paper is a secondary citation.

Forward snowballing: who cited it?

Forward snowballing needs a citation index. Three that provide one:

• Google Scholar. Wohlin screens citing papers there. It is free and broad. The list includes more than peer-reviewed journal articles. In his replication he excluded non-peer-reviewed candidates. That was his criterion for that study, not a rule for yours.
• Web of Science and Scopus. They also list citing documents and can filter them more tightly, for example by year and document type. Access usually comes through the university library, not a personal subscription.
• Semantic Scholar. It lists citations and references and is free. It is another index, not proof that the search is complete. Tools that turn citation data into a graph are compared in the AI research tools overview.

In practice: fix the criteria, screen the citing titles, and include only what you can defend from the full text. When the question is whether the anchor still holds, start with later citing papers. In a slow-moving field, go further back. “The last three to five years” is not a general rule.

What to write down

Without a record, the search looks like a random pick. A supervisor then cannot see why these particular texts are in the reference base. That applies to a bachelor’s thesis as soon as you claim the search was systematic.

Wohlin’s replication records the search path, the date, the candidates, and the inclusion and exclusion decisions. The minimum is:

The names and counts in the table are a filled-in pattern. They are not publications, and they are not advice about how large a bibliography should be.

A systematic review often shows that flow under a reporting guideline such as PRISMA: identified, screened, included. A normal thesis usually needs only a short paragraph in the methods chapter with the same facts.

Common mistakes

• Backward only. Later papers that cite the anchor are missing.
• One cluster. Wohlin warns about groups that do not cite each other. Use more than one anchor, from different groups, years, and publication venues.
• No record. The sentence “literature was found by snowballing” describes no search path.
• A closed citation group. If you keep landing on the same names, you have found a group, not necessarily the field. Look for texts that disagree with that group.
• The citing list copied across. Google Scholar also lists preprints, theses, and reports. What of that belongs in your thesis is a matter of your criteria, not of how the service sorts the list.

AI tools and invented references

A chatbot that writes prose can invent a reference that is not in any catalogue. What that looks like is described in hallucinated sources. Tools that query a citation index return records from that index. Their summary of a paper can still be wrong. A comparison of those tools is in the AI research tools overview. None of them decides for you whether a source belongs in the thesis.

What the search does not check

Snowballing produces candidates. It does not check whether a sentence in your text matches the passage in the source. No citation index and no graph does that.

Example: This is what a reference list with fabricated and mismatched entries looks like in the check. Check a sample citation →

Comparing a claim with the uploaded full text is a different step. How that check is structured is on the bibliography checker page.

Checklist before submission

• Both backward and forward snowballing were run.
• The start set is not a single paper from one group. How many papers it needs depends on how broad the question is.
• You can say whether the state of the debate in your field is covered, including later work that supports or contradicts the anchors. There is no fixed age cutoff, such as 24 or 12 months.
• Search path, dates, indexes, criteria, and counts are in the methods chapter.
• Every included source has a reason.
• Journal articles were checked against the signs in spotting predatory journals.

A supervisor can then see how the reference base was built. Whether each in-text citation matches the source is a separate check, and it is not part of snowballing.