Study models AI citation-seeking as a repeated game, proposes a reward fix

A new paper models the escalating cycle of publishers rewriting content to chase generative-engine citations — which can degrade document quality and introduce unsupported claims — as a repeated Stackelberg game between content providers and platforms, per the arXiv preprint. The authors propose “verifiable-content rewards,” a mechanism that credits rewrites adding checkable factual substance rather than only penalizing manipulative ones, reporting it beat the strongest baseline defense by an average of 12.1 percentage points across three benchmarks.