In this research we used the machine learning technology of Google‘s Cloud Vision API for the first time ever to test ad effectiveness. We tested over 9.000 online ads from over 50 product categories. We ran these ads through the Vision API to find out what ad characteritics drive the most success in terms of CTR. As far as we know, this is the first ad effectiveness study using Machine Learning as the basis of the methodology in the Netherlands.
Getting all images from DCM
To collect all ads and their performance metrics from our ad servers we used the DCM/DFA Reporting and Trafficking API and Java client libraries. We loaded all creatives for which performance metrics were still available, and stored them specified per day. For animated ads, the URL’s where reconstructed and then each of the ads was automatically opened for 30 seconds using PhantomJs or SlimerJs. During these 30 seconds, screen captures where taken 10 times each second, giving us up to 300 frames for each ad.
Getting key frames from animations
We created a python script that analyzes animations based on the frame to frame changes to select important frames from the 300 we collected before. From this we also acquired properties of the animation, like animation length, the number of loops and the type of animation (continuous or presentation). The script also calculated the number of unique colors in each frame, whereas a low number of unique colors indicates a large amount of text or drawings, and a high number of unique colors indicates photography.
Running Cloud Vision API
Finally we uploaded all images into Google Cloud Vision API requesting Label, Text, Landmark, Logo Detection, Face Detection and Dominant Colors, and stored the results for analysis.
Most important findings
