When Phrasee generates multiple variants to test, it isn't just testing random pieces of language. It's setting up a scientific experiment—testing each variant against the other.
Phrasee will be trying to predict the performance of each variant based on your previous test results. However, it will also ensure that it's constantly generating fresh language and adapting to your changing audience, too. You wouldn't want to just test the same variants every week—your audience would get bored!
The aim is to find out what language works for your audience for this campaign and for subsequent campaigns. Phrasee is not designed to get maximum uplift for one campaign in isolation, and we would never base success on just a few campaigns. Phrasee is designed to take a long-term view.
By testing a wide variant space within the language Phrasee generates, it's learning how best to communicate with your audience as quickly as possible. If you reduce the number of variants to test, the deep learning engine will take significantly longer to understand which language best resonates with your audience and, therefore, slow down all learnings.