According to what they wrote, there is no separate testing set, so they estimate accuracy based on the training set. They use cross-validation to reduce their overfitting bias. (I could be wrong since I haven't actually used their service, but this is the take-home message of the wording of their explanation. If they intended to convey otherwise, they used the wrong language.)