While training, the logged position for each listing from past search result becomes one of the listing "features". As a result you end up with a model which has listing feauture as input.

When scoring the model you set the input corresponding to the listing position to 0.