Study finds a 42.3 point citation gap from rank falls to zero: Reveneau

Study finds a 42.3 point citation gap from rank falls to zero

A new arXiv preprint from Sriram Selvam and Anneswa Ghosh audits an agentic search engine and reports that raw citation rates gap of 42.3 points between rank 1 and rank 5 falls to 0.0 points in a held-out reorder, and that rewriting a page with headings and lists redistributes 0.50 more citations per answer to that page without raising the total.

Why it mattersFor GEO work, position in a retrieval call is far less important than how a page is written, but the effect of structure is redistribution of credit among pages already cited rather than getting a new page cited for the first time.

Search Engine Journal reports on 17 September that Sriram Selvam and Anneswa Ghosh have posted an arXiv preprint titled "CITECHOICE: A Causal Audit of How Document Presentation Redistributes Citation Credit in Agentic Search". The paper, dated 14 September and not yet peer-reviewed, audits one agentic search product: a GPT-5.4 agent using Exa as its retrieval provider, replayed offline with no live web edits.

What the paper actually tested

The authors took 129 real multi-turn transcripts, and from those pulled 113 pairs of documents inside the same search call that independently supported the same pre-specified fact. A blinded human reviewer confirmed 103 pairs. For each pair, they replayed the saved conversation four ways: putting one page above or below the other, and showing that page either as plain prose or rewritten with headings and lists. Everything else in the transcript was kept the same, and only the final answer was regenerated. Both prose and structured versions were AI rewrites of the same source, produced by Grok 4.3, with a separate Grok pass checking that the facts still matched.

The three numbers a working GEO team should remember

Raw position looks huge, controlled position is zero. In the saved transcripts, pages sitting in the first Exa result were cited 85.1% of the time and pages in the fifth were cited 42.8%, a raw gap of 42.3 percentage points. When the researchers moved the same page up inside its pair, the citation chance rose by 7.9 percentage points, and that increase was not statistically significant after adjustment. On a second set of 56 pairs where only the order was switched, the estimate was 0.0 percentage points with a 95% confidence interval from -5.4 to +5.4.

Structure raises the count for the page that has it, and does not raise the total. Rewriting a page with headings and lists gave it, on average, 0.50 more citation markers per answer, with a 95% CI from +0.20 to +0.84 and a Holm-adjusted p-value of 0.033. The total number of citations per answer did not go up, and the number on the other page barely changed. The authors read this as citation credit being redistributed toward the structured page rather than more pages being included overall.

Whether a page is cited at all varies from run to run. The pre-planned test for the structure effect on incidence, meaning does the page get cited at all, was +4.5 percentage points with a 95% CI from -1.4 to +10.4 and was inconclusive. Rerunning 120 responses on the same inputs changed the binary decision for 15% of cases.

What this means in practice

Search Engine Journal quotes the paper as calling this "an attribution-sensitivity warning". A raw benchmark that says pages at rank 1 get cited twice as often as pages at rank 5 is measuring page quality alongside position, since retrieval usually orders by relevance. On a like-for-like reorder inside one pair, the ordering effect on this one agent was zero.

The structure result gives a page owner a clearer goal: moving prose into headings and lists that are easy to scan shifts citation credit toward that page without increasing the number of pages cited overall. A rival page with the same facts and plainer formatting will get less of the credit on the same answer.

Source

Reporting: AI Citation Test Finds Source Order Matters Less Than It Looks, by Matt G. Southern at Search Engine Journal, 17 September 2026. Primary source: CITECHOICE: A Causal Audit of How Document Presentation Redistributes Citation Credit in Agentic Search, by Sriram Selvam and Anneswa Ghosh, arXiv preprint 2609.15164, 14 September 2026.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.