Let's stop a bad meme before it gets going.
Multi-armed bandit is a general problem. There are many strategies for tackling it. An epsilon-greedy strategy is one strategy. It is not the only strategy, and it is not even a particularly good one. There are many others. In fact A/B testing is itself a solution to the multi-armed bandit approach.
In fact the standard way to classify strategies is their asymptotic average regret - how much time you spend pulling the bad lever. An epsilon-greedy strategy has linear regret - even after a version has won you keep pulling it a certain fraction of the time.
Let's consider an alternate approach. Run an A/B test, then forever after exploit it. This is also a linear regret algorithm. Given two levers and the way you're going to run the test, there is a probability that you will come to the wrong conclusion and forever after do the wrong thing. Most of the time your regret is fixed, but sometimes it is not.
So they are equal? Not exactly. A/B testing converges quickly, and has a very small amount of linear regret forever after. An epsilon-greedy approach converges more slowly. Furthermore to lower the long-term regret you have to drop epsilon, which slows convergence. By the time you lower it to a level that is comparable to an A/B testing approach, convergence is so slow as to be operationally infeasible.
Thus the conclusion of this article is that A/B testing is a better operational strategy than epsilon-regret. And it is. But there are known better strategies still. The algorithms that Myna uses, for instance, have logarithmic regret. In a multi-armed bandit situation, they will soundly beat A/B testing just as A/B testing soundly beats epsilon-regret.
At that point you're left with two questions. The first is whether multi-armed bandit is a good model for website optimization. The second is whether there are any operational advantages to one approach over the other.
The answer to the first, for traditional multi-armed bandit algorithms, is a sound no. While relative conversion rates remain comparable, absolute conversion rates vary for all sorts of reasons. (Time of day, day of week, longer cycles, external website improvements, etc.) That was my big criticism the last time this discussion popped up, and there was no good answer to it.
But behind the scenes Noel (the CEO of Myna) and I have been in a discussion. I believe that I can produce a multi-armed bandit algorithm that has logarithmic regret and will work in the face of ever varying conversion rates, as long as there is a consistent relative difference between the versions.
At that point you are left with the operational differences. If you want to test at a granularity of "user", a website often has to put most of its active population into the test, before collecting enough data to know which version was good. With A/B testing you will eventually put everyone into the better version. With a multi-armed bandit, half your current users are stuck in the wrong version.
In effect you have a limited population of levers available to pull, and have to pull a good chunk of them before you start getting feedback on the rewards.
You can avoid this problem by testing at the granularity of an impression. Whether or not this is OK to do is a business question, and the answer will not always be yes. If it is not, then a multi-armed bandit approach is inappropriate to use.