Preprint tests models that predict which pages answer engines cite, and releases one

Researchers Elisha Bajemon and Andre-Louis Rochet posted an arXiv preprint on Oct 28 that scores page-ranking models on how well they predict the share of an answer each candidate page receives as citations. The test used GEO-Bench queries answered by ten answer engines that were all handed the same candidate sources in their prompts; where several engines answered one query, their citation shares agreed with each other at a within-group Spearman correlation (a rank-agreement score) of 0.704.

Per the paper, a hosted decision model called Jev and two open 27B-parameter models, all used without training, were the strongest scorers at 0.575 to 0.578. A Qwen3-Reranker-4B model fine-tuned on other GEO-Bench queries, released as GEO Citation Ranker 4B, reached 0.542 to 0.549 across three training runs, and the authors say its difference from the 27B models and from Jev is not distinguishable from zero on 229 queries. The authors say they are releasing the model, the benchmark, every scorer’s predictions and the code. The paper is a preprint and has not been peer-reviewed.