Selecting which recommendation algorithm variant to advance to online experimentation is a critical decision in industry practice. Manual evaluation is subjective and time-consuming, while offline metrics such as nDCG often fail to correlate with real-world customer preferences. We present SAER, a two-stage framework (pointwise filtering and pairwise comparison) that uses Large Language Models as judges to evaluate e-commerce recommendation quality. For pointwise evaluation, SAER achieves higher agreement with human consensus (κw = 0.58) than human annotators achieve with each other (κw = 0.38). For pairwise evaluation, SAER mitigates position bias (88.0% consistency) and demonstrates strong adversarial robustness (91.0% detection rate). In a large-scale online case study serving over 10M impressions, the algorithm SAER preferred offline also achieved a statistically significant click-through rate (CTR) lift (p < 0.0001), providing encouraging directional evidence, though this single observation cannot establish a general predictive relationship. SAER produces reproducible, explainable signals in approximately 30 minutes for $28, positioning it as a scalable pre-screening layer for recommendation development.