I hate gambling. You insert a dollar in a slot machine, pull the lever, and wait to see if you win or lose. You do this knowing full well that you don’t really have control, and that it is all luck. And we all know the outcome is inevitably you throw money down the drain, the house always wins. This is not a game for me. As a CTO with a computer science background I prefer to reach my goals rationally and while being in full control.
Luckily, smart scientists developed an algorithm years ago to maximize your chances of winning when you stand before several slot machines. It’s actually a quite easy algorithm:
- For 10% of the times, you pull the lever of a completely random slot machine. You keep score of the percentage of wins vs. losses per machine.
- For 90% you pull the lever you know has the highest current win count.
That’s it! Of course, in the real world you have to play a LOT to reach statistical relevancy. The hardest part being of course to stay on the lookout for an old lady moving in to play on one of your machines. However, the theory is sound, it’s rational, and gives the player control.
This algorithm is called epsilon-greedy method for the Multi Armed Bandit problem.
So, you might wonder, what does all of this have to do with marketing? A lot more than you might expect. Marketing is more like gambling than you may think.
When a visitor hits your site, you can show him or her multiple messages and hope that they click on one of them. But which one is the best, which has the highest CTR? How do you know?
Traditionally, if you have multiple messages and you don’t know which is the best, you run an A/B test for a defined period of time, look at the resulting click through rates and then decide to pick either A or B as the winner. Next you run with the winner and you’re done. Right?
Wrong. While this could be a good strategy in some cases, it is actually a quite wasteful one in most. What if A winning is influenced by a group of specific prospects coming to your site during the test period, but over time B really is the better variant for your broader audience. Do you repeat the test every few weeks to figure this out? That’s a lot of work. And don’t forget, as you keep testing week over week, you are guaranteed to lose out on conversion by showing the weaker option to 50% of your visitors.
So that’s where the Multi Armed Bandit algorithm comes in. Using this method, you methodically switch messages for 10% of your visitors. For the remaining 90% you choose to serve out the message that has performed best so far. The brilliance is in the fact that loser and winner (10% and 90%) can switch along the way as yields change.
This algorithm is made for busy marketers, you set it up once and you don’t have to think about it anymore, the algorithm continuously tries to find the best message over time. It also starts automatically showing the right one when it reaches statistical relevancy (no need to worry about the math), removing the need to get back to the test and decide which one you want. Its ‘fire and forget it’.
And with the time saved and additional revenue generated you have more time to have another go at real slot machines.
BlueConic offers both options; A/B testing, and Always-on Optimization, which is a nicer word than ‘epsilon greedy method for the multi armed bandit algorithm’!
Want to learn more or see how marketers have been using this? Sign up for our upcoming webinar: “Start Using Always-on Optimization in Four Steps”.