What academic research actually says about AI citation factors

Search-industry blog posts have circulated GEO advice for two years now, quotations help, statistics help, structured data helps, backed mostly by aggregator posts citing each other rather than any named study. That is starting to change. Four papers published or accepted in 2026 ran controlled experiments or large-scale tracking studies specifically on what makes generative AI engines cite a source, and their actual text, not a summary of a summary, is the basis for what follows.

The largest controlled experiment: position and relevance win#

The most rigorous of the four is a factorial study by Vishwakarma, Kumar and Jamidar, accepted to ACM SIGIR 2026. It built a two-document retrieval-augmented generation testbed: inject exactly two candidate sources into a model’s context, vary exactly one content factor between them, and measure which source the first citation marker in the answer points to. Across six LLMs the team ran 252,000 trials covering eighteen content factors, using brand anonymization and counterbalanced source order to isolate content effects from simple position bias.

The result undercuts a lot of GEO folklore: topical relevance and list position were the biggest drivers of being cited first, not any single piece of on-page dressing. Explicit price information and a recent timestamp helped consistently but by a smaller margin. Completeness and trust cues added smaller gains still, and formatting-only edits, the kind of change a checklist tool flags, had little measurable effect at all.

Where citations actually land, and how failures get fixed#

A separate, much larger tracking study by Pratyush Kumar, run on the Ranqo platform, analyzed 102,025 AI prompt responses across 102 brands on five engines between March and May 2026. Its most striking number is not a percentage lift from adding a quote, it is where citations go at all: only 2.9% of citations pointed to a brand’s own domain, while 75.2% pointed to other companies in the same space, corporate sources overall taking roughly 78% of all citations and the rest split among YouTube, Reddit, editorial media and Wikipedia. Ranked best-of listicles were the single most-cited content format, at about 21% of citations. The same study found visibility scales sharply with brand stature: global household names appeared in 73% of relevant first-run answers, established mid-market brands in 44%, and niche or small brands in just 11%.

A third paper, by Tian, Chen, Tang, Liu and Jia, treats citation as something to diagnose rather than just measure. Their AgentGEO framework classifies why a specific page failed to be cited, then applies a targeted repair instead of a generic rewrite. Tested on a 12,240-query benchmark built from 204 real webpages, AgentGEO reached a 79.52% citation rate against a 68.80% baseline, a relative improvement of more than 40%, while touching only 5% of a page’s content against 25% for baseline methods. Its failure analysis is the more useful finding for publishers: 62.2% of failures traced to the content simply not matching what the query needed, 27.1% to broader content-quality issues, and only 10.1% to technical integrity problems like malformed markup or blocked crawling. Formatting is rarely the bottleneck; substance is.

A widely repeated figure that does not check out#

Search results for this topic keep surfacing a specific claim: that 68% of AI citations come from third-party sources and only 32% from brand-owned websites. That figure does not appear in any of the four papers checked for this piece, including Kumar’s tracking study, the one paper that actually measures this split at scale. Kumar’s real number is more extreme in the same direction, just 2.9% of citations land on a brand’s own domain, not 32%, which suggests the circulating 68/32 figure is either from an unverifiable source or a garbled version of a different study’s numbers. It is not used here.

A critical survey covering the field from 2023 to 2026 also traces one specific number back to its actual source. Search snippets describe “252,000 trials across six LLMs and eighteen factors” without naming a paper; the survey attributes that exact figure to Vishwakarma et al.’s SIGIR 2026 paper, which this piece fetched and confirmed directly. The same survey separately reports, citing the original 2024 GEO paper by Aggarwal et al., a roughly 41% relative gain in position-adjusted word count from adding quotations, a figure this piece did not independently verify against Aggarwal et al.’s own text and so treats as one step more removed than the four primary findings above.

Why we care#

For publishers, the concrete, verified takeaway is narrower than the SEO-blog version of GEO advice. Position in the retrieved set and topical relevance to the actual query dominate every factorial result checked here; no study reviewed found a formatting trick that outperformed simply being the most relevant, best-positioned source for the question being asked. Best-of listicle formats earning the largest single share of citations in Kumar’s data is worth acting on: a page structured as a ranked comparison has a real, measured citation advantage over a generic landing page. And AgentGEO’s failure breakdown is a useful diagnostic lens even without the tool itself: if a page is not getting cited, check whether it actually answers the query being asked before reaching for schema markup or citation-bait phrasing.

What is not yet proven is anything close to a stable formula. Every number above comes from a specific set of LLMs tested on specific dates in 2026, none of it peer-reviewed, and the field is young enough that a widely repeated statistic, like the 68/32 citation split, can circulate for months without tracing to any real study at all. Treat any GEO percentage a vendor quotes the same way this piece treated one: ask which paper it is from, then go read that paper.