03 · What You Need to Know
How AI-Assisted Citation Chaining Works and Where It Falls Short
What Is Citation Chaining?
Citation chaining is a literature discovery method that follows citation relationships among publications. It begins with one or more known relevant studies, often called seed references, and uses their bibliographies or citing records to identify additional potentially relevant publications.
The process is also commonly called citation searching or citation snowballing, although terminology and specific procedures vary among disciplines and methodological frameworks.
Unlike a keyword search, which retrieves records through search terms and indexed fields, citation chaining follows relationships established through scholarly references.
This can be particularly useful when relevant studies use different terminology or when a research area spans several disciplines.
The TARCiS statement, developed to guide citation searching in systematic reviews, addresses the terminology, conduct, and reporting of citation searching. It recognizes citation searching as a potentially valuable supplementary approach rather than an automatic substitute for database searching.
What Is the Difference Between Forward and Backward Citation Chaining?
Backward citation chaining
Examines the reference list of a selected publication to identify earlier or otherwise previously published works that the paper cites.
Forward citation chaining
Identifies publications that cite the selected paper, allowing researchers to investigate subsequent scholarly engagement with its findings or ideas.
For example, suppose Paper B cites Paper A, and Paper C subsequently cites Paper B.
Starting from Paper B, backward citation searching may identify Paper A. Forward citation searching may identify Paper C.
These are direct citation relationships. Paper A and Paper C may be intellectually connected through Paper B without directly citing each other.
This distinction matters because an AI system may recognize an intellectual relationship and incorrectly present it as a direct citation.
What Can Generative AI Actually Do During Citation Chaining?
The answer depends on whether the system has access to actual bibliographic and citation information.
Extract references from accessible documents. If a publication's reference list is available, AI may help identify and organize the works it cites. Extraction errors remain possible, especially when references are incomplete or inconsistently formatted.
Retrieve citing publications. A system connected to a citation index or scholarly metadata service may identify records that cite a selected study.
Organize citation networks. AI can help group retrieved publications by research topic, theoretical framework, methodological approach, or apparent relevance.
Support screening. Given actual bibliographic records, AI may help identify which publications warrant closer examination against the research question.
Interpret citation contexts. When the relevant full texts are accessible, AI may summarize how later authors discuss a cited publication, subject to verification against the original passages.
These activities can reduce repetitive work, but none automatically establishes that every relevant citation relationship has been identified.
Why Is Finding Related Papers Not the Same as Citation Chaining?
This is perhaps the most consequential distinction.
Semantic scholarly search identifies publications that appear conceptually related. Citation chaining identifies publications connected through actual references.
A study investigating teacher trust in artificial intelligence may be semantically similar to another study on algorithmic trust without citing it.
Conversely, a paper may cite a publication for a statistical procedure even though their substantive research topics are unrelated.
| Approach |
Relationship used |
What the result establishes |
| Semantic discovery |
Estimated conceptual similarity |
The publications may address related ideas |
| Backward citation searching |
References contained in the seed publication |
The seed publication cites the identified works |
| Forward citation searching |
References in later citing publications |
The identified works cite the seed publication |
| Bibliographic coupling |
Shared references between publications |
The publications cite some of the same sources |
| Co-citation |
Two publications are cited together by another work |
They share a citation relationship through citing publications |
These relationships can complement one another, but they answer different questions. Researchers should not describe AI-generated related-paper recommendations as verified forward or backward citation searches.
Which Sources Can Support AI-Assisted Citation Chaining?
Forward citation searching generally requires a source that records citation relationships. Examples include Scopus, Web of Science, Google Scholar, OpenAlex, and OpenCitations, although their coverage and access arrangements differ.
Backward citation searching may be conducted by examining the reference lists of original publications or through indexed cited-reference metadata.
Some AI-enabled research tools use these or other scholarly information sources. Researchers should establish which citation data the tool actually accesses.
A system's ability to recommend related publications does not establish that it has queried a citation index.
The narrower task of identifying papers that cite a particular study requires verification of the citing relationship before the results can support a forward citation search.
Why Can Citation Chaining Miss Relevant Publications?
Citation chaining is not inherently comprehensive.
Its effectiveness depends partly on the seed references selected. If the starting papers represent only one theoretical tradition or disciplinary community, their citation networks may overlook relevant research from other traditions.
Forward citation coverage also depends on the citation index. Missing records, incomplete reference metadata, indexing delays, and publication types outside the database's coverage can affect results.
Backward searching may be limited by inaccessible reference lists, incomplete citations, or the decision to examine only selected references.
There is also a temporal limitation. Newly published studies may not yet have accumulated citations, making them difficult to discover through forward citation relationships.
AI cannot resolve these limitations simply by producing more recommendations. A larger network visualization or longer list of related papers does not establish comprehensive coverage.
Can AI Perform Multiple Rounds of Citation Chaining?
Potentially. A retrieval-enabled system may begin with a seed publication, identify its references and citing works, then repeat the process using newly discovered publications.
This is sometimes described as iterative citation searching or snowballing.
However, repeated expansion can produce a rapidly growing network containing many irrelevant records.
Researchers need stopping rules or other procedures appropriate to their review design. They should also distinguish records discovered during different rounds and preserve the citation relationships used to identify them.
An AI system that silently selects which branches of a citation network to follow may introduce selection decisions that are difficult to inspect.
For a systematic review, this lack of transparency can be more consequential than the time saved through automation.
Can AI Decide Which Citation Links Are Worth Following?
AI may help prioritize candidate publications based on titles, abstracts, or other available information. This can be useful when a citation search retrieves a large number of records.
However, screening should follow the review's eligibility criteria rather than an unexplained relevance ranking.
For example, a highly cited paper may be historically influential but irrelevant to the population or intervention specified in a systematic review.
Likewise, a less prominent publication may be eligible because it directly addresses the review question.
The distinction between research importance and relevance remains essential when deciding which citation relationships deserve further investigation.
What Does Reproducible Citation Chaining Require?
A defensible citation search should identify the seed references, citation sources, search dates, and procedures used to retrieve and screen records.
Where multiple iterations are conducted, researchers should describe how new seed references were selected and when searching stopped.
The TARCiS statement provides specific guidance on citation searching in systematic reviews. PRISMA-S also addresses reporting of supplementary search methods.
Researchers should preserve citation-search outputs where feasible, recognizing that citation indexes may change over time.
If AI assists with selecting seeds, retrieving records, screening results, or interpreting citation contexts, its contribution should be documented appropriately.
The broader issue of whether an AI-assisted search is reproducible enough for a systematic review depends on the transparency and validation of the actual process, not merely the availability of saved prompts.
When Is AI-Assisted Citation Chaining Particularly Useful?
It can be useful when researchers have already identified a small set of relevant publications and want to explore how the surrounding literature developed.
It may also help when terminology is inconsistent, making conventional keyword searching less effective.
For theoretical reviews, citation networks can help trace the development of ideas. For systematic reviews, citation searching may identify additional eligible studies missed by database queries.
Nevertheless, the appropriate role of citation chaining depends on the research question and review methodology. It should not automatically replace structured database searching or other evidence-identification methods.
