Limits of A/B Tests
A/B tests don’t give you perfect decisions.
No matter what you do, you’re never 100% certain
If we’re not careful, winners aren’t really winners
Your conversions go up… and then they come back down
The Standard Solution
Run your test until you hit 95% statistical signiﬁcance.
Go to getdatadriven.com if you need a signiﬁcance calculator.
Martin Goodson’s PDF on poor testing methods: kiss.ly/bad-testing
This gives us the best data but not necessarily the best ROI.
So how far do we take this?
We modeled several A/B testing strategies. Using Monte Carlo simulations, we tested diﬀerent strategies over 1 million observations (people).
Will Kurt gets full credit for all this. @willkurt
1 Pick the minimal improvement The Scientist: 2 Determine your sample size 3 Determine degree of certainty (95%) 4 Start test but don’t check it early 5 If results aren’t signiﬁcant, keep control
Results for the Scientist:
1 Waits until 80% signiﬁcance The Reckless Marketer: 2 Calls a winner as soon as 80% gets hit
Results for the Reckless Marketer:
1 Waits for 95% signiﬁcance The Impatient Marketer: 2 Moves on to the next test after 500 people
Results for the Impatient Marketer:
The Realist 1 Waits for 99% signiﬁcance 2 Moves on to the next test after 2,000 people
Results for the Realist:
The Persistent Realist 1 Waits for 99% signiﬁcance 2 Moves on to the next test after 20,000 people
Results for the Persistent Realist:
The Blitz Realist 1 Waits for 99% signiﬁcance 2 Moves on to the next test after 200 people
Results for the Blitz Realist:
Let’s compare them using the area under the curve.
Don’t make decisions at less than 95% signiﬁcance.
You’ll waste all the time you spend testing
1 Be a scientist at 95% We have 3 viable strategies for making this work: 2 Only make changes at 99% 3 Sloppy 95% but make it up in volume
1 Pick the minimal improvement Be a scientist when you have lots of data and resources 2 Determine your sample size 3 Determine degree of certainty (95% 4 Start test but don’t check it early 5 If results aren’t signiﬁcant, keep control
If you don’t have the data or resources to be a scientist, go fast at 99%.
And if you still want to play at 95% without being a scientist, never stop testing.
How We A/B Test
First, get volume to 4000+ people/month.
Only make changes at 99% signiﬁcance.
Let the test run at least 1 week before checking results.
If not at 99% after two weeks, launch the next test.
If the next test isn’t ready, let it keep running while you build the next one.
The KISSmetrics A/B Testing Strategy 1 Get to 4,000 people/month for test 2 Only change the control if you reach 99% 3 Check results after 1 week 4 Launch the next test at 2 weeks 5 Let old tests run if you’re still building
This strategy isn’t perfect. It’s a balance between good data and speed.
Clipping is a handy way to collect important slides you want to go back to later.