After we began using Bayesian statistics in our algorithms, our R&D team created yet another innovation by designing a new, unique type of algorithm to maximize conversions.
Why show your website visitors variations of your site that don’t convert particularly well during the entire duration of a test, when there are other variations that perform better? AB Tasty helps you minimize the ‘waste’ of traffic inherent in every test, thanks to a dynamic traffic allocation algorithm.
The latter solves the tricky question of the ‘multi-armed bandit’ by finding the balance between data exploitation and exploration to continuously – and quickly! – optimize your conversion rates. We go above and beyond simply helping you decide what to do after an A/B test. Instead, we offer a ‘smart’, automated solution perfect for use when decisions need to be made continuously, when the time is short or when you’re working in a constantly changing environment.
What is dynamic traffic allocation?
Dynamic traffic allocation consists of using an algorithm to modify the quantity of traffic sent to each live test variation.
Why is it useful?
It’s useful for limiting loss of conversions during a test, (called ‘regret’), which can occur when a part of your website’s traffic is sent to a variation that ends up not being the winner.
Let’s look at the following A/B test: ConversionRateA = 1%, ConversionRateB = 1.5%, carried out with steady traffic of 10,000 visitors per variation.
The regret of the test is: r = 10.000 * (0.015-0.01) = 50 conversions lost. Over the duration of the test, we could have had 300 conversions (20.000*0.015), but the test made us lose 50. There would therefore only have been 250 conversions during this time period.
Of course, we can only make this type of calculation after a test is over when we have the exact figures for the conversion rates. However, this doesn’t mean nothing can be done to limit wasted traffic during a test…
How does dynamic traffic allocation work?
The solution to the above problem is to modify the test’s traffic allocation so as to send fewer visitors to ‘bad’ variations and more towards the ‘good’ variations.
Watch out, doing this manually is very risky because you might invalidate the results. However, there are algorithms that can channel traffic so as to minimize a test’s regret, while at the same time identifying the winning variation…
We’ve chosen the most reliable algorithm for the job, based on the following idea: we use the uncertainty of the conversion rate measurements to reach a compromise between ‘exploration/exploitation.’ We ‘explore’ a variation when we send it traffic, even if initial readings don’t say it’s the winner since we know these first conclusions aren’t reliable. We ‘exploit’ when we send traffic to a variation deemed the winner by the numbers already collected – this is how we avoid losing too many conversions (assuming it’s really the true winner).
Naturally, these two goals go against one another. Exploration means losing conversions, and exploitation means taking a risk if you don’t pick the winning variation! It’s therefore critical to accurately model the uncertainty of the measurements, and then to find the right compromise between ‘exploration/exploitation.’
We take into account the uncertainty of the conversion rate measurements for each variation, thanks to probability distributions. These graphs show where there’s the most chance of finding the true conversion rate value. The more the curve is high on the Y-axis, the higher the chance is that the corresponding X value is the real one.
Here’s one example:
The variation A has 7 successes from 600 visits (black curve), version B has 27 successes from 600 visits (red curve). The situation is clear: it shows us that the variation A’s conversion rate is likely between 0% and 0.2%, and that of version B is likely between 0.25% and 0.7%. Since these are distinct intervals, even if we can’t be sure of the measurements, we’re still able to say with near certainty that B is the winning version. There’s little room to doubt that version B is the winning version since the curves don’t overlap.
Here’s another example:
Variation A has 7 successes from 300 visits (black curve), version B has 14 successes from 400 visits (red curve). The simple conversion rate calculation gives us ConversionRateA = 2.39% , ConversionRateB= 3.63%. There seems to be a difference, so we’re tempted to say that version B is the winner, but this is inaccurate….by looking at the probability distributions, it’s easier to spot the uncertainty of these measurements. By noticing that the two curves overlap, we understand that there’s still room for doubt.
The ‘Exploration/Exploitation’ Compromise
Let’s continue looking at the last example. We notice that it’s just as likely that ConversionRateA equals 3% and that ConversionRateB also equals 3 % (where the two curves cross). With this type of approach, we can calculate the probability that A is the winning variation, even if for now, B seems to be the better one. We use these type of calculations to find the right balance between ‘exploration and exploitation’. We can estimate the usefulness of exploring and the risk associated with exploiting with the help of an algorithm, such as Thompson Sampling.
- is sure, over time, to find the winning variation
- Is guaranteed to lose fewer conversions than if we used steady traffic
- Will find the winning version more quickly (if there are more than 2 variations), than if we used steady traffic. The more variations there are, the more likelihood there is that there are a few (very) bad variations. These poor performers will quickly be identified and given less traffic than those that perform better. If steady traffic allocation is used, these (very) poor performers would continue to lose a non-negligible amount of traffic.
How can you use dynamic traffic allocation?
Using dynamic traffic allocation with AB Tasty is very simple, all you have to do is check the box ‘Dynamic Traffic Allocation’, and to choose the KPI to optimize. In the beginning, traffic will be uniformly allocated, and then will be adapted and automatically allocated in such a way as to identify the winning variation and minimize the loss of conversions.
Once the test is launched, everything stays the same as with a classic test (with uniform traffic allocation). Of course, all statistical measurements take the dynamic allocation into account. Interpreting test results is therefore exactly the same.