Follow citations with a record of every seed, direction and decision

1. Start with a seed whose identity is known

Choose a publication because it is relevant to the defined question, not simply because it has a large citation count. Record its exact title, authors, year, version and verified identifier where one exists. Open the actual publication to establish what it says and which reference list belongs to that version.

A seed can be eligible for a review, useful background or a methods source. Name that role explicitly. Using a paper to find other work does not mean you have included it as evidence. If a correction, updated version or retraction notice changes its interpretation, retain the link and assess the implications before repeating its claims.

2. Keep the direction of each connection visible

Backward chaining examines the works cited by the seed. Forward chaining looks for works that cite the seed. Google Scholar documents a Cited by link and a separate Related articles function. The latter finds similar documents; it is not proof of a citation relationship. Label similarity-based discovery separately in the log.

For each candidate, record how it was found and verify its identity against a primary record. If you need to know why it cites the seed, read the relevant passage. A citing paper may challenge, qualify or merely mention an earlier work. Counting the connection cannot establish endorsement, replication or support for the same conclusion.

Reference: Google Scholar: search help

3. Select a documented route and record its limits

A seed’s own reference list provides a direct backward route. For forward searching, Google Scholar supports citation export through its Cite function. Web of Science documents cited-reference searching, including reference variants, and explains that access can depend on the institution’s subscribed years and products. Use the interface and access actually available to the team.

These functions are described in current primary help pages checked on 15 September 2026. They have not been independently trialled here. Record the platform, collection, search fields, date, result selection and export format. An inaccessible result or missing link needs an unresolved status, not an invented citation or a claim that the underlying publication does not exist.

Reference: Google Scholar: search help · Web of Science: cited reference search

On small screens, scroll the table horizontally, or focus it and use the arrow keys.

4. Build a fictional chaining log with honest totals

Imagine a training seed called S, a wholly invented paper about implementation of hospital discharge documentation. In a fictional log dated 12 September 2026, its supplied PDF reference list yields eight candidate records. A simulated forward-search worksheet for Google Scholar yields six. Two exact records occur in both sets, so the combined set contains eight + six − two = twelve distinct records.

Three of those twelve are already in the team’s existing library, leaving nine new candidates. Suppose screening excludes four against the predeclared scope, sends two for full-text retrieval and leaves three with unresolved identity. Four + two + three = nine. No additional eligible study has yet been established, and the numbers say nothing about records outside the supplied exercise.

5. Resolve identities before discarding duplicates

Keep the untouched source files and a record connecting each imported item to its seed and direction. Compare exact identifiers and citation details before consolidating records. A spelling variation may be the same work, while a protocol, conference abstract and full report can be separate publications associated with one study. Preserve those relationships when they matter to the review.

When two exact records are consolidated, retain both discovery routes. This makes it possible to explain where a candidate came from without counting it twice. Do not erase the original export or silently change an eligibility decision. Give ambiguous matches an owner and a next action, such as checking the publisher record or obtaining the missing citation fields.

6. State the stopping rule and what remains unknown

Before expanding the search, define which new eligible seeds will be chained, which directions and sources will be used, and how an update will be handled. If the team stops after one generation because of a time constraint, record that decision and the remaining work. Reaching a planned stopping point does not prove that all relevant literature was found.

Report the executed method, dates, exports and eligibility decisions alongside the original database strategy. Keep background discoveries separate from included evidence and avoid using citation counts as quality scores. In the fictional exercise, nine new candidates are a workload count; they are not nine included studies or nine confirmations of the seed’s conclusions. The next useful action is to resolve the pending identities and obtain the specified full texts.

Questions for a practical final check

• Seed identity, version and role recorded.
• Citation direction distinguished from similarity.
• Platform, collection, date and export preserved.
• Exact duplicates retain all discovery routes.
• Candidate counts separated from included studies.
• Stopping decision and unresolved records reported.