Billy's Blog: Multivariate Testing & Optimization

Note: This is the first post of a 3 part series, each focusing on one type of test result that is tough to deal with.

There are 3 types of optimization results that people never look forward to getting.

Unfortunately, anyone who runs enough tests will run into these situations. In the following 3 posts I will go over the 3 situations and outline how they happen and how to prevent and handle them.

This first post is about tests with highly mixed data.

How does this happen?
Typically mixed test data occurs when events out of your control, or you forget to control, impact your test and skew your test results.

Your average traffic finds your tested page either through search, browsing around your website or through your planned advertising campaigns. However they get to the page, tests are designed (or should be designed!) based on how visitors will get there.

The problem arises when, outside of the scope of those involved in testing or because of some oversight, a new type of traffic is driven to the page without any preparation being done for that traffic. While traffic is good for your sample size, it is bad because those new visitors are coming in with totally different motivations, assumptions and knowledge. This means they probably will react differently to your tested elements than the traffic you were driving to it originally.

Most often this happens to me because of a new marketing push, such as an e-mail blast, new display ads or promotions at a trade show. This can happen even more unexpectedly if some outside party drives a lot of traffic to your page. A news story or blog review that innocently links to your page, can suddenly becomes a big source of new traffic.

What can you do to prevent this?
The first step is to spread the knowledge around your company that this testing is going on and that anything that may impact the page and its visitors should be run by the optimization team first.

Next, always segment out significant traffic and track it separately. If you segment, the worst case scenario is that they perform the same as your current traffic and you did a little extra work. The alternative is having to trudge through your data, trying to separate the 2 types of traffic and possibly having to restart the test if you can’t separate them out.

Lastly, be aware. Watch your data and look for big changes. If you see something strange or a sudden shift, try to find a cause. It usually will be nothing, but if you do find something, quickly build a segment for it. Even if the “new” traffic has already started hitting your page, a segment should be setup as soon as possible.

What can you do if this happens?

I would still try to segment the data in any way possible. Even taking certain days/time out of your data, may be enough to salvage your results. Do your analysis with and without those days and see if the optimal page changes. You should take extreme care when doing this still though and make sure you have statistically relevant results.

My next solution is just to restart the test. Testing is about continual growth and you can’t always get what you want out of every test. Be happy that you got some extra traffic and try to take precautions to take it into account, or prevent it, the next time around.

Let me know if you’ve ever run into these problems before and how you handled them. Look out for part 2 of this series in the next week.

Photo Source (under CC)