Performative decision-making arises in many real-world problems where decisions influence the populations that generate future data. A major line of existing research focuses on learning performatively stable solutions. We study average-reward performative optimization, where the learner observes an evolving population through sample batches of fixed size and learns from a single continuing trajectory with unknown dynamics and without resets. We propose a policy-gradient method combining randomized trajectory lengths and gradient tracking. Our analysis uses Poisson equations to control initialization and return-truncation bias, even when population-state distributions do not converge over time. Under regularity assumptions, a uniform Poisson inverse bound, and an exact gradient-domination condition, the algorithm achieves \(\widetilde{O}(\sqrt{T})\) expected regret against the best fixed policy in the chosen class.