The success of demand response (DR) programs, and in particular, incentive-based services, is subject to accurately estimating the customer baseline load (CBL). While conventional CBL estimation methods employed for industries are primarily based on day-matching and control-groups, the challenge lies in refining these to robustly handle the higher demand variations exhibited by small and medium enterprises. For this, we propose an improved day-matching technique, where, for each DR event, we adaptively select a customer-group using similarity theory. Subsequently, day-matching is performed on the selected group to minimize any biases. An empirical expression is derived for utilities to approximate CBL estimation errors, and thereby assess the suitability of our method for new or returning customers. A comparative study against conventional CBL estimators is conducted using data from the Irish Commission for Energy Regulation. While all methods show similar performances for low demand variances, the proposed method is distinctly superior when the variance becomes higher.