Proving the monetary value of A/B testing

Also see my previous article: The CRO Process.

I’ve created a method to calculate the impact of A/B testing in terms
of yearly profit. And not unimportant: this has been accepted by the
whole company. This method isn’t perfect, but it gives you as a CRO
manager quite some power to prove the effect of your team.

The prerequisites to do this:

• 
  showing what you normally would expect to get in terms of traffic,
  number of orders and average revenue (or even average profit) per order
  for the next 12 months. If you don’t have this, you could take data of
  the past 12 months as an alternative baseline.
  
• You ran some successful A/B test :).

* I take 12 months so I don’t have to worry about any seasonal effects.

Our
BI/Finance team can already provide me with the baseline prediction
(based on current numbers and including predictions on competitors,
market changes, investments etc. etc.) so I don’t have to worry about
that. Of course we also have some successful A/B tests at our disposal.

Now I can perform these steps:

1. Calculate the extra profit by just running this single A/B test.
2. Prediction/extrapolation: A) Increase the baseline 12 month prediction with the changes in conversion rate and/or average order value proven by the test.B) Make some corrections on A) to make sure you’re not overselling yourself.

1) Calculate the extra profit by just running this single A/B test

This
is just a simple calculation only based on your A/B test data. You
calculate the revenue (or better: profit) from your original variant and
all other variants. Then you calculate the difference if you would have
had the original running 100% versus running the test with all your
variants.

This
works for all tests (successful or not) and tells you if you made any
loss/gain during the test itself. Of course you only implement the
winner, but if you test multiple variants it could be that the net
result during the testing period is negative. This helps you to get a
grip on the monetary cost/risk of performing A/B test.

2A) Increase the baseline 12 month prediction

This
one is actually quite simple: you take your prediction (we have it on a
day-by-day basis) and apply the changes in conversion rate and average
order value that we saw in the test.

2B) Make some corrections

Now
this is where it gets tricky. 2A) will get you a very precise (and
usually quite high) number which I feel will need some corrections
(downwards) to make it realistic. I also prefer to put it in a range
instead of just one single number.

Some of the assumptions/corrections I do:

• I
  assume a diminishing effect per month. I’ve set this to 2% which means
  that after a full year, you will only have 78% of the original effect
  left.
  
• When
  I do a test in a single or just a couple of countries, I always assume
  the effect will be less when applied in other countries. In the same
  region I’ve set this to 25%, for countries outside the tested region
  this is set to 50% (this is not included in the example sheet below).
  
• A
  correction for the significance level (we test at 95%). This correction
  makes no sense when applied to one test, but when you do this for all
  experiments that you implement, you can assume a certain part (around
  23%,
  

After
all corrections I turn the single number that comes out of this
equation into the upper limit of a range and subtract a percentage to
get to a lower limit. For a country this is 20%, when applied to the
same region it’s 40% and for the rest of the group this is 50%. All
arbitrary numbers.

And the result of all this? Instead of saying “we have improved conversion by 4%” I can now claim “after implementation this test gives us €636K — €854K extra profit this year”.

Try it yourself!

For
obvious reasons I can’t share our own internal sheet (containing our
complete budget), but I made a simplified version in Google Sheets that
should make the basic calculations clear:

Let me know if you think some things are off or if you have any improvements.

Write down your assumptions

I
highly recommend writing down all your assumptions in your calculation
sheet. You will need to make quite some of them to make this work and
writing it down makes it tangible and easier to adjust afterwards if you
have more data to update your assumptions and the calculations based on
them.

Some more written down assumptions I have in my sheet:

• We
  work across 12 countries with different currencies while the budget
  predictions are all int he same currency. Changes in exchange rates will
  not be taken into account.
  
• During
  A/B tests we optimize for CR and AOV but the testing system cannot
  optimize for profit (technical limitation). Profit is a percentage of
  AOV but this is different for different countries and different
  products. This calculation assumes a stable relation between AOV and
  profit which is the best we can do right now.
  

This also helps when sharing this sheet internally because: buy-in!

Don’t do this on your own, get buy-in

Now
I didn’t build this sheet on my own: I also got our business analyst
and our financial controller involved to check the numbers, the
calculations and all assumptions. They helped improving the sheet to
match how finance calculate numbers and to get an official stamp of
approval.

Some final remarks

Of
course this is just a prediction and even in hindsight after 12 months
you will not be able to calculate if the changes in your numbers were a
result of your A/B test implementation.

This
fact means it still feels a bit weird to me to do such a prediction.
But because we made this together with BI/Finance this is currently the
best way we have to quantify the effect of our experiments. It also
provide us with a way to compare our experiments.

But
the best reason to do this is that it works really well. Putting a
dollar sign in your experiment reports makes a lot more people
interested in the results and helps you to increase your budget.