2% of households returning their refrigerators and buying a new one seems pretty high, although I don't have any data to back this up. How many people reading this have made multiple (separate) refrigerator purchases in a week?
With arbitrary assumptions about statistics, you can "statistically prove" pretty much anything. Until the 2% figure is substantiated, it's all baloney.
There's also the fact that advertisers spend many millions of dollars buying these ads, so they have a pretty large financial incentive to do it well.
You are assuming there's a way to do it well and not well, but what if nobody actually knows how it is - doing it well? What if marketing is just given a budget of X million dollars to spend on advertising, but nobody there actually knows whether doing X or Y works, but they know everybody does X, so if they do X, there's no chance they'd be fired for that?
This does not demonstrate that the machine learning algorithms are doing a bad job. Quite the opposite really. See:
https://twitter.com/patio11/status/875629380105416705