We at RADLY work every day to bring AI and machine learning solutions to the hands of marketers. So, how can machine learning boost creativity in marketing?
Marketers often face the question of which creative has the best performance and what is the optimal combination of creative elements in a banner. A common way of figuring this out is to conduct a series of A/B-tests, but that turns quite laborious if you have tens and tens of different possible element combinations to test. If this task sounds like a great challenge for AI, you are not wrong: how about having an AI detect the differences in efficiency and automatically create new versions of banners it believes will generate better performance in your target group?
To test the idea of a Banner AI, we built a little game that can be described as a “banner tinder”. This game demonstrates how the algorithm checks out performance metrics and modifies the banners based on what has worked and what has not. Unlike in the real-world case where banner performance metrics work as training data to the algorithm, our offline demo needed to collect the data in a more hands-on way. Luckily the DMEXCO hosted in Cologne 11.9-12.9.2019 provided us with a great platform to go around and request training samples for our AI.
With these 15 base banners, we started asking people for opinions, one banner at a time. As the data began to grow, we learned a couple of things: firstly, there is a lot of variation between subjective opinions between people, which brings much variety in our “small data” demo. Secondly, right and left are internationally tricky to distinguish, so visual cues on swipe directions are essential (in all fairness, KPIs like click-through-rate also contain misclicks, so a few accidental likes or dislikes should not tip the scales).
After two days of intensive requesting for opinions, our demo collected a beautiful set of data and automatically created more than 200 new versions of our RADLY banners. At the end of our data collection endeavour, the banner set had evolved quite a bit. The most notable changes were the survival of the lighter colours, a few favourite images and only using texts with black highlight.
The demo did give us valuable insight and feedback on dos and don’ts of successful banner evolution. What one can find marvellous, is that even if there might not be a huge potential to discover the golden egg of banner versions in every case, at least with the help of Banner AI, the versions remain fresh and possible miscalculations get automatically cleaned from the creatives.
In the end, it is debatable if one should describe our Banner AI with the adjective “creative”. Ultimately it only mixes and matches elements given to it in a rather uncreative way: by checking out the data and following the brand guideline rules given for it. Even if the Banner AI does not create entirely new things from scratch, at least it acts as a creative way to provide an automated helping hand to the creative designer in the field of creative versioning.
Interested to hear more? Contact the RADLY team for more information and pilot opportunities!