This project is a success based on key takeaway::

a. Emotions data can be linked in real-time to existing audiences with a match rate of 40% to 50%

b. There’s a correlation between audience emotive response and engagement - greater engagement corresponds with more polarised emotional response

c. Emotional dimensionality can be reduced to enable a condensed view of emotional response and engagement for a range of audiences

d. There is a correlation between emotive response and social media action

e. It’s possible to more effectively target and plan media based on predicted social actions

On slide 6 the figures on the heatmap are indexes relative to the average levels of emotion for a video. Indexes above 100 represent above-average emotion, whilst numbers below 100 represent below-average emotion. Not all Proprietary Audiences from our DMP were analysed – we focused on a set of broader audiences to ensure results were not skewed from small samples. Quantifying the emotional impact on smaller audiences is also possible but requires increasing the sample size of our emotion measurement studies.

Observing the impact on social media highlighted the variation in social media action rates compared across audiences. Notably, a number of audiences generated social media action rates above the average action rate. The implication is that selective targeting of specific audiences generates greater social response than untargeted videos. By selecting a more optimal mix of audiences with positive emotional response to a video, we observed double digit % lifts compared to a scenario where no emotional data targeting was applied. (Slide 7 of material)

This initiative can be easily scaled and repeated and already has been for the 12 videos in this study. Audience Emotion maps can now be easily created for any video and used in real-time to inform creative team or media-planning.

To achieve the desired outcomes, we brought together emotional response data (from webcams) and Audiences (created in our DMP) that we target programmatically in real-time. We ran 12 surveys for 12 video advertisements spanning a range of industry sectors and markets. Each survey had approximately 1,000 respondents. We focused on 6 key emotions (Happy, Sad, Disgusted, Confused, Surprised, and Scared) and also measured Engagement, which reflects any emotive reaction to the stimulus, all second-by-second. In this raw measurement form emotive video response is particularly informative for creative and content development, but needs refining for media planning and targeting.

To better use emotion measurement for media planning, we joined the webcam emotive responses with our audiences from our Data Management Platform using cookie synching, resulting in a 40-50% match. This allowed us to not only quantify ebbs and surges of emotions throughout the creative duration for each audience, but also quantify overall levels of emotion for each audience.

One interesting observation was how the variance in levels of emotion across the six emotions increased with increasing levels of Engagement. This suggested an emotional 'space' across Engagement and levels of (positive/negative) emotion. We used multidimensional scaling to collapse the 6 emotions into a single dimension to make the audience emotion dataset more accessible. Each audience could then be positioned within this emotional space, mapping out their relative response. Slide 6 in the supporting materials highlights an example.

We also wanted to observe the impact upon social media actions. We developed models and trained them on approximately 2,083 videos watched by over 371,245 audiences. These models predict different social media actions on a 'per 1,000 YouTube views'. Positive correlation between predicted action and actual action was observed. The best performing model was then deployed to predict social media actions for each audience studied.