Exactly, there will be no way to determine that why negative effects you measure were actually due to the emoji.

The only real way to deal with this is to run these types of A/B tests “forever”, which is not desired. Which is why I’m very much wondering, what is the appropriate way to deal with this?