I implemented a recommender using tensorflow, based on e-commerce data. This recommender is predicting the next item a user will buy. I will judge the performance of my fitted model, by getting the next estimate of the model and comparing, if the user did actually buy this item (compare precision_at_k with k == 1).

My precision is around 20% (meaning in 20% of the cases the user did actually buy the product the system recommended). Our shop has about 150 Products and my evaluation is based only on properties of the user, not his/her purchase history.

In this case, is 20% good? Are there other, similar systems I can compare to? What would be a good precision for, e.g. 100.000 products?