Prior to SQL 2008 ISNULL was considerably faster than COALESCE. This makes sense as the code behind ISNULL has to deal with a fixed number of input variables (2) and COALESCE has to be designed to work with any number of variables, so this will invariably involve extra computations.

In SQL 2008, I've seen a thread where people say that the performance is now for all intent and purpose the same for both (I wonder if the optimiser just converts it to ISNULL if there are only 2 variables).

Personally, if I'm only working with two values, then I'll use ISNULL and if I need more I'll use COALESCE in the spirit in which the two functions were designed.