We compare three models of representation of item order in a verbal STM task: item-item associations, item-position associations, and primacy gradient. A speeded probed recall task is used, in which a list of words is presented, immediately followed by a probe; participants must report as fast as possible the word that was in the probed position. In the number probe condition, a digit is presented and one must say the word in that position. In the word probe condition, the probe is an item of the list and participants must say the immediately following item. Response times (RTs) are analyzed according to probe type and position. The three models imply different predictions about RTs as a function of serial order in the two conditions. Our results suggest a serial, self-terminating search from the beginning of the list to the target position, except for the final position, which is directly accessible. The item-item and item-position association models are ruled out; the primacy gradient model accounts satisfactorily for our results, except for the finding of a larger recency effect with a number probe. Alternative interpretations are also discussed.