In recent years, zero-shot anomaly detection algorithms have attracted attention as effective methods in practical settings where collecting anomalous samples is difficult. However, similar to the well-known order bias observed in large language models (LLMs), there is a possibility that these algorithms may also exhibit positional bias depending on the location of input anomalies in images. In this study, we investigate positional bias in the representative zero-shot anomaly segmentation model, Segment Any Anomaly+ (SAA+), through empirical experiments. The results reveal a clear tendency for the model to detect anomalies more frequently in certain regions, especially the upper center of the image. To mitigate this bias, we propose two novel calibration methods and evaluate their effectiveness in terms of anomaly detection performance and bias mitigation. Experimental results demonstrate that the proposed methods improve both precision and recall while mitigating positional bias.