Benefit Transfer (BT) is often applied when a primary valuation study is considered too costly or time consuming to conduct. It is commonly assumed that BT performance improves with increasing similarity between study and policy sites. However, no common criteria for defining similarity exist, making it difficult to operationalise the concept of similarity in a practical BT context. We propose a structured framework for distinguishing between different degrees of similarity. In particular, we differentiate between three dimensions: physical, population and attribute similarity. While the first two are often used in the literature, attribute similarity is not. To investigate the impact attribute descriptions have on BT, we define it as whether or not the same ecosystem service categories are emphasised in the valuation studies. Using value estimates for water quality improvements obtained from 17 Choice Experiments conducted in Europe, we empirically test unit value transfer performance along a similarity gradient. The results confirm that increasing physical similarities across commodities and sites generally lead to lower transfer errors. However, when using income adjusted value transfer, we surprisingly find the opposite. Finally, we demonstrate that increasing attribute similarity may offset dissimilarities in terms of the site characteristics.