Follow up: Another take on creative AI
Last October, we wrote a blog piece about how AI could assist creative work by automating A/B-testing and making simple adjustments to banner creatives.
To get everyone to speed, a quick summary of the previous post: We were interested in studying the possibilities of using machine learning to boost creativity in marketing. For this purpose, we built a Banner AI demo. The core idea of Banner AI is basically to automate A/B-testing and creative versioning in a smart and creative way. The demo itself was an algorithm coupled with a simple Tinder-style UI to collect opinions about various banner versions. The algorithm used the collected data to continuously modify the banners showed by mixing and matching different elements from different banners. We operated the demo in DMEXCO in Cologne 11.9-12.9.2019 and gathered quite a nice amount of data. The demo gave us great insight into both the algorithm and our banner design.
A demo is always a fun thing to play with, but in the end, you really want to see how and if the algorithm works in a real–life application. In late October, we got an opportunity to use Banner AI in an actual digital marketing campaign, when our client, Helen, got enthusiastic about the idea and jumped in for a pilot. And thus, the AI behind the demo was connected with an existing dynamic ads setup to create new banner versions based on the measured performance of the previous ones.
Our Pilot period lasted for five weeks, during which the Banner AI created 169 unique banners. Banners had a couple of copy text versions and a couple of different CTA texts, but mainly the variability in banners came from combining different colors in the design. The results of the pilot – well – exceeded our expectations.
We measured the success of the algorithm with the change in click-through-rate (CTR) compared to a group of banners that were kept static through the pilot. The total lift in CTR during the pilot was 10%. We observed that the first couple of weeks worked as a learning period, where the algorithm did more exploration than exploitation. After that, the lift in CTR started to be more prominent and during the last week of the pilot, the lift was as high as 30%.
With the pilot, we were able to lift marketing performance, combat Ad fatigue, and create new banner versions that would have taken a massive amount of time if done manually. The opportunities for using an algorithm like this to automatically test and modify banner creatives are intriguing. By harnessing AI power to creative modification, we can not only achieve a level of creative optimization that is impossible by hand but also take versioning to a whole new level. For example, animation length or styles are rarely tested and modified but could provide a substantial improvement in performance.
Interested to hear more? Contact the RADLY team for more information and pilot opportunities!