SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents
How product recommendations change when one candidate’s source is rewritten to influence the system.
Accepted to EMNLP 2026.
Film credits
Music: “Effervescence” by Scott Buckley, used under CC BY 4.0; edited and mixed for this film. The narration is a synthetic voice, not a person.
The assistant window and the web pages are drawn to show how a benchmark case works; the quoted sentences come from the released case, and its product is fictional. The ranking is one recorded run (DeepSeek-V4-Flash, one case), not a measurement across the benchmark. Results: SafeGEO, 2026-09-04, on 600 requests across 6 product categories and four AI systems. “Rewritten honestly” is the clean condition, the truthful rewrite; “rewritten to win” is the shaped condition, the average of the paper’s eight realistic rewrites. The check against evidence is the evidence breakdown, evaluated on the same shaped instances. Each instance changes one candidate-controlled text source.
Research resources.
Benchmark domains.
Explore the requests, compare clean and shaped sources, and see how we evaluate a recommendation against the person’s needs.