[mrtsepelev]

1) Yes, the problem with creating variations is still here and is worth mentioning. Today you need to know your product well and know your visitors well to create a valuable hypothesis. And this is what we're working on right now (disclaimer - I'm one of the co-founders of Landy).

2) The idea here is that with ML you should not analyze every dimension separately. ML is taking into the account all available characteristics and making decisions based on all of them together (like if the guy on OS X, who came from NY from the Facebook campaign in the evening - prefer to watch product video instead of watching screenshots - no problem, we'll show him video).

3) The real power of ML comes out when you could not obviously split your traffic based on the ad link (like utm_campaign=dogs). Direct and search traffic on your homepage are great examples in this case. Also, manual targeting requires a bunch of analytic folks, who will continuously analyze your traffic, setup and adjust optimization campaigns. Even in this case - it's still difficult to adapt to dynamic changes in traffic (like a new type of visitors, season changes, etc). So ML could not only improve results but also decrease the amount of human resources which is currently required for solving such complex problems.

So what I'm trying to say is that your assumptions definitely make sense in some cases. But we believe that there are still plenty of cases when ML could drastically increase your results and save your time.