The explosion of available data in the last few years has increased the importance of physical database design, since the selection of proper physical structures (e.g. indices, partitions and materialized views) may improve query execution performance by several orders of magnitude. Commercial DBMS vendors have recognized this need and oered automated physical design tools as part of their products. These tools use what-if interfaces to simulate the presence of different physical structures and recommend physical designs that minimize the estimated execution time of a given workload. Along with the recommended design, they deliver an estimation of the expected improvement the new design brings. In this paper, we examine the output of physical designers, i.e., whether what we see as a result of the tuning (the estimation of the improvement) is indeed what we may expect after applying the design (the actual improvement). We evaluate three commercial physical designers by varying their input parameters on real and synthetic data sets. Our results show that all three physical designers exhibit highly unpredictable behavior in certain cases, indicating that there is still signicant room for improvement in terms of their predictability and consequently, their quality.