There are very few genetic risk score models that outperform traditional observational disease markers (here[1] is a non-paywalled discussion of cardiovascular GRS performance as an example). The best GRS results tend to be in relatively genetically-homogeneous populations that are similar to the population in which the GRS model was developed. In some cases, knowing the ethnicity or simple family history of a patient can buy you a good portion of the AUC of the relevant GRS.
So if you have a classifier like ZIP, that (1) epidemiologists have done a bit of legwork correlating to classical markers like obesity (or their correlates, such as income and dietary/smoking/prescription patterns) and (2) tends to follow familial/ethnicity clusters in (3) a heterogeneous population, you can amass a fair bit of predictive power on the cheap for complex disorders where environmental variance plays a role, as well as beating the spread on behaviourally-determined mortality/morbidity factors.
It is likely that the predictive power of GRS-based approaches will improve for many conditions in the future (they are of course already powerful for Mendelian disorders).