GEO content scores barely predict AI citations, study finds

A new arXiv preprint validates a deterministic content score built as a cheap proxy for AI-citation likelihood, using adversarial falsification tests — negative control, dose-response, bounded amplification — rather than assuming the proxy holds. Re-testing the field’s original 2023 citation-lever effect sizes on ten modern AI engine families found none of the levers moved citations, and recalibrating to current data strips the score of its lever-responsive components entirely. On a 500-source adversarial benchmark, amplifying the surviving levers gained an attacker at most 6 points, and the score’s own citation-predictive power was weak — a within-query Spearman correlation of 0.11 — positioning it as a content-quality filter rather than a citation predictor, per the authors.