You could run the test as a MVT rather than 2 different AB's to determine the best combination, then if needed you could run an AB with the winning variant to test it's true performance vs. current/default. Hard to tell if this is an acceptable solution without more detail, just a thought based on ...