# A 4B Model Predicts Which Pages Answer Engines Cite

> Researchers introduce GEO Citation Ranker 4B, trained to predict citation share across ten answer engines using the same candidate sources.

Oossa · 2026-10-30 · https://oossa.com/en/a-4b-model-predicts-which-pages-answer-engines-cite

A team publishing on arXiv introduced GEO Citation Ranker 4B, a small model trained to predict which pages answer engines will cite. They tested scorers on GEO-Bench queries answered by ten engines, each given the same candidate sources.

The researchers report that the 4B model’s results were not statistically distinguishable from those of larger 27B models or hosted model Jev across 229 queries. They also released the model, benchmark, predictions and code.

## The facts

- The ten engines agreed with one another at a within-group Spearman correlation of 0.704.
- GEO Citation Ranker 4B scored 0.542–0.549 across three training runs.

## Why it matters

Teams studying how to get pages cited by answer engines can try a released model and benchmark, though the reported results do not establish how well it predicts citations outside this test setup.

## Sources & references

1. [Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval](https://arxiv.org/abs/2610.38090) – arXiv cs.IR, 2026-10-30

Last updated: 2026-10-30