# Google Experiments, Variance, and Why Confidence can really suck

There are many unique parts to optimizing on a lower traffic site, but by far the most annoying is an expected high level of variance. As part of my new foray into the world of lead generation I am conducting a variance study on one of our most popular landing pages.

For those that are not clear what a variance study is, it is when you do multiple variations of the same control and you measure all of the interactions against each other. In this case I have 5 versions of control which gives you a total of 20 data points (all 5 compared to the other 4). The point of these studies is to evaluate what the normal expected variance range is as well as the minimum and maximum outcomes from the range. It is also designed to measure this over time so that you can see when and where it normalizes down to as each site and page will have a normalization curve and a normal level of variance. For a large retail site with thousands of conversions a day you can expect around 2% variance after 7-10 days. For a lead generation site with a limited product catalog and much lower numbers, you can expect higher. You will always have more variance in a visit based metric system then a visitor based metric system as you are adding the complexity of multiple interactions being treated distinctly instead of in aggregate.

There are many important outcomes to these studies. It helps you design your rules of action including needed differentiation and needed amounts of data. It helps you understand what the best measure of confidence is for your site and how actionable it is. It also helps you understand normalization curves, especially in visitor based metric systems as you can start to understand if your performance is going to normalize in 3 days or 7. Assume you will need a minimum of 6-7 days past that period for the average test to end.

The most annoying thing is understanding all the complexities of confidence and how variance can really mess it up. There are many different ways to measure confidence, from frequentest to Bayesian and P-Score to Chi Square. The most common ways are Z-test or T-Test calculations. While there are many different calculations they all generally are supposed to tell you very similar things. The most important of which is what is the likelihood that the change you are making is causing the lift you see. Higher confidence means that you are more likely to get the desired result. This means that in a perfect world a variance study should have 0% confidence and you are hoping for very low marks. The real world is rarely so kind though and knowing just how far off from that ideal is extremely important to knowing how and when to act on data.

This is what I get from my 5 experience variance study:

To clarify, this is using a normal Z-Test P-Score approach and there are over the bare minimum conversions that most people recommend (100 per experience). This is being done through Google Experiments. The highest variance I have ever dealt with on a consistent basis is 5% and anything over 3% is pretty rare. Getting an average variance of 11.83% after 5 days is just insane:

This is just not acceptable. I should not be able to get 97% confidence from forced noise. It makes any normal form of confidence almost completely meaningless. To make it worse, if I did not do this type of study or if I did not understand variance and confidence then I can easily make a false positive claim from a change. These types of errors (both type 1 and type 2) are especially dangerous because it allows people to claim an impact when there is not one and allow people to justify their opinions through purely random noise.

If you do not know your variance or do have never done a variance study, I strongly recommend that you do so. They are vital to really making functional changes to your site and will allow you to avoid wasting so much resources and times on false leads.