OctoLink GEO

A SIGIR Study Names Four Citation Gatekeepers in a Two-Page Test

Author Editor

A SIGIR 2026 paper reports 252,000 trials and says topic match, price, a recent timestamp, and list position were unanimous citation gatekeepers.

GEO Citations SIGIR

Direct answer

The paper says four factors were unanimous across six models, with odds ratios above 100: topic mismatch, missing price, an old timestamp against a recent one, and second position against first.

A five-page paper in SIGIR 2026, held 20–24 July in Melbourne, studies competitive citation inside a two-document test. Rahul Vishwakarma, Shushant Kumar, and Ratnesh Jamidar inject exactly two candidate pages into the model context and record which URL appears first in the answer. Each pair differs in one of 18 factors. Brands and publishers are replaced with fictional names. Order is swapped so position is not confused with content. They say the run is 252,000 trials across Gemini 2.5 Flash, Claude 3.5 Sonnet, Kimi K2 Thinking, GPT-5 Nano, GPT-5 Mini, and GPT-5.2.

The results section says 11 of the 18 factors reached significance in at least four models. Four of those were unanimous, with odds ratios the table reports as greater than 100: being off topic, omitting the price, carrying an old timestamp against a 2026 timestamp, and sitting in position 2 instead of position 1. The authors call these gatekeepers. Failing one, they write, can wipe out the odds of being cited first even if the rest of the page is strong.

What the design does not measure

The method never calls a search engine. The tool response always lists the same two variants. About 86.4 percent of successful answers contained exactly one distinct URL. The paper's own limitation is that production retrieval often returns five to ten pages, so a two-page forced choice is not a live answer engine.

Source: SIGIR 2026, What Gets Cited.

FAQ

How large was the test?
The authors say they ran 252,000 trials across six models, from 1,440 scenarios that differ in exactly one of 18 factors.
Is this a live Google ranking?
No. The method injects exactly two documents as the only tool results. The authors say they never call a search engine.