The canonical model of choice considers multiple alternatives evaluated over multiple attributes. Nevertheless, little is known about optimal search in this setting. This paper provides the first partial characterization of optimal sequential search in a problem with multiple searchable attributes and alternatives, full recall, and no order restrictions on search. When this partial optimal benchmark is applied to a well-known dataset it is discovered that subjects systematically deviate from optimal search by (1) searching too deeply within alternatives and (2) switching too adjacently between alternatives. Existing behavioral models cannot explain these patterns of behavior.