Back to Blog

How To Properly Perform Statistically Relevant Split PPC Ad Testing

February 14, 2014

By Greg Finn
No Comments

The crux of a solid paid search campaign indubitably is the ad copy that drives the user to make the click. When auditing paid search campaigns from other companies, it is always a shock to see the (lack of) ad testing that is used. An A/B split test is an extremely simple way to test a baseline ad against a new iteration, and use simple statistics and confidence testing to prove a theory with an exact confidence. This is a crucial element that we leverage with client recommendations and tactics — as facts and science never lie. Want to see how easy it can be to calculate statistical relevancy? Follow along in the steps below.

1. Formulate a Test & Implement

In order to test, we’ll need two different versions of an ad to test. More importantly, we need to identify what the success metric is. Maybe it is number of purchases, maybe it’s number of clicks. For this example, let’s assume that our success goal is a “sign-up” and we are testing the ad copy effects on visitors.We will start with one “championship” ad that is currently active that we’ll call “Ad A”. The variation will be “Ad B” with all elements the same as “Ad A”, except for the ad headline.

2. Ensure You Are Collecting Accurate Data

As any statistician will tell you, the sample that you are using is key to cut down sample error. The goal is to pull our data from an unbiased sample that is representative of the population.

Ad Rotation In Google:
The way that AdWords is set up out of the gate won’t serve us in the way we need, as it will try to optimize for clicks. We would rather use a systematic sampling model that will provide us equal results in our test. We recommend using “Rotate indefinitely” for your test within the campaign settings tab in AdWords:

Google-AdWords-Rotation
Ad Rotation In Bing:
Like Google, Bing also automatically sets ads to show more often based on clicks. To turn this off you’ll need to change each Ad Group by hopping into the Ad Group settings. You’ll see the Ad Rotation option:

Bing-Rotate-Ads

3. Collect Data

Ahhh, the easy part. Sit back and watch the impressions, clicks and conversions pile up. The size you’ll need will depend on how accurate you’d like the test to be.  Obviously the more data the better and you can see some tables on error percentages based on differing sample sizes. If you want to project a timeframe required to hit your ideal test sample size, Visual Website Optimizer has a nice free tool to help.

4. Analyze Data

If you aren’t a savvy statistician, don’t fret. There are plenty of calculators that can help display the data you require. I personally use Thumbtack’s free A/B Test calculator for quick comparison. You’ll want to insert your number of trials (visitors in this case) to the successes (sign-ups). You can adjust the confidence level if you’d like, but a 95% confidence interval is standard in stats.

thumbtack-split-testing-calculator-fail
The p-value displays our confidence that the improvement is in fact statistically relevant. The lower the p-value, the higher the confidence. We’d like to have a p-value of between 0.05 and 0.01 for “statistical significance.” For the example above we can see a very low p-value that means the tests have reached statistical significance. Seeing that this is statistically relevant, we can then look at the results and the estimated improvement. This test shows that the variation would likely cause a decreased success rate that would actually be anywhere from a  24% – 143% decrease in conversions (with the average conversion decrease of 93%).

thumbtack-split-testing-calculator
This example however shows that our test is currently not statistically significant as our p-value is a whopping 0.21. There is nothing that can be statistically recommended from this test. So from the two examples, we would strongly recommend keeping the Baseline ad in Test 1 as visitors were 93% more likely to convert, while we would not have any recommendations for the second test. If the success ratios held firm on the second test continue we’d need a little more than 2.5x the sample size to be statistically significant.

And that’s it! Not only should this help you make more informed decisions for yourself and clients, but it also an easy method to back-up your work and tactics with science in case a client doesn’t agree with your strategy..

Like this? Want more great info?  Be sure to follow us on Google+:

Keep up with us on Twitter:

or of course 
Instagram and Facebook.

Greg Finn

Greg Finn

Greg is the director of marketing at Cypress North. He has been managing internet marketing campaigns for more than seven years and is a certified Google AdWords partner in search advertising. Greg has been both a speaker and moderator at SMX, the world's leading search engine marketing conference, and is also a contributing editor and regular writer at SearchEngineLand and MarketingLand.

See Greg's Most Recent Posts

Share this post

Leave a Reply

Search our blog

Start A Project

Categories

What's next?

Well...you might like one of these

Article

How to Create a High-Quality Email...

Optimizing Your Contact Database for Effective Email...

Read article

Article

Should Google Ads Recommendations...

Google loves broad match, and shockingly, I do too! A 2019...

Read article

Article

How to Set Up Google Ads Automated Rules

Learn How To Set Up Automated Rules In Google Ads &...

Read article