A paper posted to arXiv's information retrieval section introduces E-GEO, a testbed for generative engine optimization (GEO) in e-commerce: rewriting product content so AI answer engines and conversational shopping agents surface it. The authors say current GEO practices are largely ad hoc and their impact understudied.
E-GEO contains 13,746 realistic, multi-sentence consumer product queries, each paired with 10 retrieved Amazon listings. The authors say the queries capture intent, constraints, preferences and shopping contexts that existing datasets miss. They used it to study five generative engines, seven LLM rewriters and fifteen rewriting heuristics.
The authors also built a prompt meta-optimization algorithm that they say significantly improves over heuristic baselines. Prompts optimized independently from different heuristics converged to a stable pattern, which they say suggests a shared, transferable GEO strategy. After red-teaming the system with static and optimization-based attacks, they report that under a simple in-prompt defense in their setting, GEO gains reflect genuine content improvement.