A/B testing is the primary decision-making mechanism in large-scale e-commerce platforms. However, reliable experimentation requires more than random user assignment: heterogeneous user populations, low-sensitivity metrics, and interference from concurrent experiments can lead to misleading conclusions and costly product decisions. In this paper, we describe the statistical methodology layer developed by Trendyol's data science teams to improve both the trustworthiness and iteration speed of online experiments at scale. We first apply persona-based stratified assignment to balance behavioral distributions across experimental groups and reduce variance induced by population heterogeneity. During test execution, we continuously monitor statistical power at the metric level and expose power alongside statistical significance, enabling practitioners to detect underpowered outcomes and reduce false positives even when metrics appear significant. Finally, we enforce orthogonal assignment principles across concurrent experiments to minimize cross-experiment interference and enable safe parallel experimentation. Rather than focusing on experimentation infrastructure, we highlight compact and operational statistical design choices and formalized decision rules that strengthen experiment reliability in production environments.