How Can Brands Navigate the Complexities of AI Citation Behavior?

What did the new study change about AI citation behavior?

It changed the confidence with which anyone can point to search-result position and call it the reason a page gets cited. A preprint titled CITECHOICE, posted to arXiv on September 14, 2026 by researchers Sriram Selvam and Anneswa Ghosh, ran a causal audit of how an AI search agent (GPT-5.4 paired with the Exa search provider) allocates citations when two sources support the same fact.

The team pulled 129 everyday-query transcripts and isolated 113 document pairs where both sources independently verified the same pre-specified claim. They then replayed each transcript twice, swapping the order of the pair and rewriting one target source as either structured (stats, bullets, a short table) or plain prose, while holding everything else fixed. That design lets the study separate correlation from cause, which is the part most GEO commentary skips.

The raw numbers looked dramatic before controls: sources in the top position were cited in 85.1% of transcripts, versus 42.8% for sources ranked fifth, a 42.3-point gap, according to AtoZ SEO's summary of the working paper. Once the researchers controlled for other variables, that gap fell to 7.9 percentage points in one test and to a statistically flat 0.0 points (95% confidence interval, -5.4 to +5.4) in a second. Position, it turns out, was mostly standing in for formatting.

Rewriting a source in structured form (the CITECHOICE authors call it 'structured rendering') raised the target's citation count by +0.50 citations per answer, with a 95% confidence interval of +0.20 to +0.84 and a Holm-adjusted p-value of .033. Crucially, that gain did not come by stealing credit from a competing source. It concentrated citation credit onto the structured page without reducing total citations across the answer. A third finding is a caution for anyone running single-prompt spot checks: citation decisions flipped in 15% of identical reruns of the same query, and the researchers estimate roughly 45% of the variation they saw came from model randomness rather than anything about the page itself.

Who does this change affect?

Two groups feel this most directly. B2B SaaS and tech growth teams that have been told to 'win position one' on comparison and 'best tool' prompts now have evidence that position is a weaker lever than they assumed, and that the structured asset on the page (a comparison table, a named stat, a dated data point) is doing more of the real work.

Local, multi-location, and service businesses face a related problem from the other direction. A business that ranks well for '[service] in [city]' but presents that information as a wall of prose is leaving citation credit on the table that a structured competitor could pick up, even from a lower-ranked page.

It also affects anyone selling GEO tactics as position-gaming. If a vendor's pitch rests on 'we get you to position one and the citations follow,' this study is the first controlled evidence that the correlation behind that pitch is weaker than the industry has assumed. That does not mean position is worthless. It means position without a structured, sourced claim on the page is doing less than most GEO decks imply.

How should brands respond to the new citation research?

Respond by moving budget and attention from rank-chasing to asset-building, and by testing claims more than once before acting on them. Three concrete moves follow directly from the study's findings.

First, put a real citable asset on every page competing for a claim: a dated statistic, an original data point, or a comparison table, each with a named, linked source. The +0.50 citation gain in CITECHOICE came specifically from structured rendering of a verifiable fact, not from prose that merely mentions the fact.

Second, stop treating a single query and a single answer as proof a page is or is not being cited. With a 15% flip rate on identical reruns, one test tells you little. Run the same prompt several times, across more than one AI engine (ChatGPT, Perplexity, Gemini, and AI Overviews all allocate citations differently), before drawing a conclusion about a page's citation performance.

Third, keep rank as one input, not the strategy. A page still has to be found and considered before it can be cited, so ranking and indexing work still matters. What changes is the story you tell a client or a stakeholder about why a page got cited: the honest answer, per this study, leans toward the structure and sourcing of the content on the page rather than its position in a results list.

What should brands watch for before overreacting?

CITECHOICE is a preprint that has not yet been peer reviewed, and the authors tested one model and one search provider, GPT-5.4 with Exa, replayed offline instead of against live, editable webpages. The findings are a strong, controlled signal, not a settled law for every AI engine.

The caution cuts both ways. It is a mistake to conclude position never matters, and it is equally a mistake to conclude formatting is a trick that reliably wins a citation. The honest framing, consistent with the study's own recommendation, is to test in aggregate, across repeated runs and more than one provider, instead of acting on a single query result in either direction.

Sources

Source: arXiv (Selvam, S. and Ghosh, A.)

CITECHOICE study abstract, methodology, and the +0.50 citations per answer structured-rendering finding with confidence intervals arXiv (Selvam, S. and Ghosh, A.), 2026-09-14

Source: AtoZ SEO

Raw, uncontrolled citation rates by search-result position (85.1% vs 42.8%) and the 15% rerun flip rate reported from the working paper AtoZ SEO, 2026-09-17

Source: arXiv (Aggarwal, P. et al., GEO: Generative Engine Optimization, KDD 2024)

Foundational finding that optimizing content for generative engines, including adding statistics and citations, can boost visibility in AI answers by up to 40% arXiv (Aggarwal, P. et al., GEO: Generative Engine Optimization, KDD 2024), 2024-06-28