How machine learning fuels conversion optimisation [White Paper]

By Otto Nyberg, on 27 September, 2017

When talking about machine learning in a commercial context, people often think that their conversion rate will increase just because advanced calculations relating to machine learning are being made in a server room. That is not the case. In order for machine learning to improve the conversion rate, there needs to be a complete use case where a machine learning algorithm figures out something interesting and the customer is able to use this information to their benefit.

In a standard example, you would have a treatment that has an effect of, let’s say, +10% on conversion if offered to every visitor. An example of this would be offering free delivery to every customer in an online store. If you instead targeted this treatment with machine learning, that 10% would set an upper limit to how much your conversion could improve. In the best case, we would find exactly the visitors that bought because they got that treatment and not offer this costly treatment to anyone else. This would land at the same +10% improvement in conversion. In such a case we would have optimised the benefit, i.e. minimised the expense while maximising the return.

In order to reach something better, you need a slightly different setup. If you had a treatment that has a positive effect on some of the visitors and a negative effect on others, you could actually reach a better conversion rate with machine learning than you could by offering that treatment to everyone. This would be a case where you reach some conversion rate precisely because you use it together with machine learning; a case where you could not reach the same conversion rate without machine learning.

- As it happens, this is precisely what we have done.


Case XXL

XXL, a sports equipment retailer in the Nordics, has used our targeting algorithm together with an automated chat message. We found out that, on their site, approaching every customer with an automated message produces a negative change in conversion rate. When we targeted 2% of the sessions, the average conversion rate went down by 7.6%. Note that the 2% were not entirely randomly selected, but were visitors that stayed longer on the site and were thus more likely to buy than the average visitor. When the automated chat message was targeted to 2% of the visitors with our machine learning algorithm, the resulting conversion increase was 6.6%!

Hence when targeting a treatment that on average decreases the conversion rate, we were able to turn the effect around and actually increase the conversion rate! In plain English, we targeted the visitors with the automated chat message that bought because they were targeted (visitors that needed help), and left the ones alone that would have been annoyed by the chat.

We have together with XXL showed that machine learning can produce real value in an online store, both for visitors and for the store. Using machine learning, we were able to find the visitors that needed help while leaving the ones alone that did not want help, thus increasing the revenues of the store.

For more details, read our recent whitepaper on uplift modeling:



If you want to hear more we are more than happy to continue the conversation, just come and chat with us or contact sales (at)!

Topics: Live Chat, Case study, Downloadable, AI, Ecommerce