ChatGPT citation rates vary sharply by topic, study finds
The model's linking behaviour rewards travel content and ignores education. For multilaterals and financial institutions, that is a structural visibility problem.
Key takeaways
- ChatGPT cites sources far more often on travel queries than on education or policy topics.
- Citation behaviour follows perceived verification need: fresh, specific queries pull links; stable conceptual questions do not.
- Multilaterals, standards bodies, and financial institutions concentrate publishing in low-citation topic zones.
- Brands can improve citation surface by producing dated, specific, verifiable content rather than purely explanatory material.
- Citation rate is not a fixed property of an industry; it shifts with how queries are framed.
Travel questions get cited. Education questions do not. That asymmetry, reported by Search Engine Journal, is one of the more consequential findings to emerge from recent analysis of ChatGPT's linking behaviour, and it has direct implications for any brand whose authority lives in a low-citation category.
The pattern is not random. ChatGPT's citation logic appears to follow perceived verification need. When a user asks about a flight connection or a hotel in Lisbon, the model reaches for external sources, presumably because freshness and specificity matter and the model knows its training data ages. When a user asks how photosynthesis works or what caused the First World War, the model answers from its own weights and links to nothing. The question type shapes the citation behaviour more than the quality of any individual source does.
What the citation gap actually means
For brands in high-citation categories, this is an opening. If ChatGPT regularly links out on travel queries, a travel brand with well-structured, crawlable, authoritative content has a genuine shot at appearing in LLM-generated answers. The competition for that slot is real but finite. The path from "good content" to "cited in ChatGPT" is shorter here than almost anywhere else in the topic map.
The inverse is where the strategic risk concentrates. A financial services firm, a multilateral institution publishing development research, or an industrial group releasing technical standards guidance all face a harder problem: if their primary subject matter sits in a low-citation zone, the model will absorb their intellectual output and produce answers that owe a debt to that work without acknowledging it. The World Bank can publish a rigorous paper on financial inclusion; ChatGPT can synthesise its findings for millions of users and cite nothing. The institution gets zero visibility credit.
This is not a speculative concern. It is the default behaviour the citation-rate data describes. Education and policy-adjacent topics, precisely the domains where multilaterals and philanthropic institutions concentrate their publishing, appear to be among the categories where ChatGPT answers from internal knowledge rather than linking out. For organisations whose mandate includes influence, that is a structural problem, not a content quality problem.
The mechanism worth watching
Citation rates also vary by how a question is framed, not just by topic. A query that implies a need for current, verifiable information ("best hotels in Osaka this autumn") pulls citations. A query that implies a stable answer ("explain compound interest") does not. B2B brands can test this by examining whether their core queries are framed as lookup tasks or as explanation tasks. If the answer is predominantly explanatory, citation probability drops regardless of content quality.
There is a practical response. Brands in low-citation verticals can reframe their content around queries that carry implicit verification signals: recent data, specific figures, named sources, dated findings. A pension fund's market commentary is more likely to be cited if it answers "what happened to emerging market bonds in Q1 2025" than if it explains how bond duration works. The distinction is between content that answers a stable conceptual question and content that resolves a specific, time-bound information need. ChatGPT appears to treat only the latter as requiring an external pointer.
The deeper implication is structural. Citation rates are not evenly distributed across the knowledge economy, and the topics that drive authority for major institutions, standards bodies, multilaterals, and policy organisations tend to cluster in the low-citation zone. Those organisations cannot wait for the model's behaviour to change. They need to build citation surface where it exists, which means producing more of the specific, dated, verifiable content that prompts the model to reach outward, even if that sits alongside rather than instead of their core analytical work.
The brands that treat citation rate as a fixed property of their industry will cede that surface to whoever decides it is not.