Structured content wins more LLM citations, audit confirms
When AI engines choose among equally valid sources, structured formatting shifts citation credit. A causal audit puts a number on the gap.
Key takeaways
- Structured formatting earns +0.50 citations per LLM answer versus prose, in a controlled causal experiment.
- The gain comes from concentration: total citations stay flat, and competitors lose no credit.
- Being retrieved but uncited is a real, measurable outcome. Formatting determines which documents get named.
- The effect is confirmed for agentic, multi-turn search; single-turn RAG contexts remain untested.
- For multilaterals, industrial groups, and policy institutions, reformatting existing content is the immediate priority.
Formatting is not cosmetic. When two sources support the same claim equally well, a generative search engine will cite the one that uses structured presentation roughly half a citation more per answer, according to CITECHOICE, a causal audit published on arXiv.
That figure needs unpacking, because its mechanism is counterintuitive. The gain does not come from an engine admitting more sources overall, nor from a competitor losing its slot. Total citations per answer stay flat. Competitor credit is unaffected. What structured formatting does is concentrate credit: the same document earns a larger share of an unchanged citation pool. The experiment ran 103 blinded, human-confirmed document pairs, each pair drawn from authentic multi-turn agentic search transcripts, each pair verified to support the same specific fact. A hash-verified 2-by-2 replay crossed pair order against jointly generated structured and prose renderings of one target document. The result was +0.50 citations per answer (95% confidence interval: +0.20 to +0.84; Holm-adjusted p = .033). That is a controlled, causal estimate, not an observational correlation.
The mechanism that content teams are missing
Most discussion of LLM citation patterns has treated the problem as one of source admission: does the model retrieve your document at all? CITECHOICE reframes it as citation allocation within a retrieval set. If the model has already retrieved five documents that all support the same claim, which one gets named in the answer? The finding is that structured rendering shifts that allocation in the target's favour, without displacing competitors in any measurable way. The credit pool is fixed; structured content claims a larger portion of it.
The practical implication is specific. A document that is retrieved but rendered in dense prose may yield zero explicit citations even while its information shapes the answer. A document with equivalent evidential content, reformatted with headers, lists, or structured data fields, earns a citation more often. For a brand trying to establish visible authority in LLM answers, the distinction between being read and being named is decisive. Invisible influence does not build brand recall and does not appear in whatever AI-referral traffic currently remains.
This matters most for organisations whose credibility depends on being cited correctly, not merely consulted. Consider a multilateral institution publishing a policy brief on disaster-risk financing: if a structured rendering of that document earns half a citation more per relevant query, across thousands of agentic sessions, the cumulative visibility gap between a well-formatted and a poorly formatted publication is substantial. The same applies to a global industrial group whose sustainability disclosures are retrieved routinely but credited inconsistently. In both cases, the loss is not access but attribution.
What the audit does and does not prove
The design is more rigorous than most work in this space. Using same-call document pairs, meaning both documents are retrieved in the same session and evaluated on the same fact, eliminates confounds from retrieval differences. Blinded human review of factual support prevents the audit from smuggling in quality differences through the back door. The hash verification confirms the replay conditions were stable.