Exact’s overarching goal was to develop a granular understanding of their unknown potential customers who were visiting Exact’s website. Their main KPI was to gain an in-depth understanding and transparent view of their users’ behaviour, feeding this data into DV360 (and other relevant channels) to maximise return on their marketing campaigns and apply exemplary marketing communications with prospects.

Incubeta was tasked with building a model that could identify Exacts’ main on-site products of interest, and the intent/interest levels of all users in the customer journey – and within the funnel.

We opted for a 2-phase approach called the ‘User Engagement Engine’ or UEE that permitted audience analysis and creation. 

• Phase 1: A manual scoring Algorithmic Approach 
• Phase 2: A Machine Learning Model Approach

Firstly, we analyzed audience behaviour on the Exact website, by mapping out all relevant events, page paths and metrics thresholds for converting users – we call these ‘features’. We then assigned scores to each feature and created three groups to which users would be assigned according to their scores. The groups were reflecting a Touch/Tell/Sell model.

Once a user receives a score, user information is pushed back into GA360. From here, we further segmented users by creating relevant audience lists that were then activated in DV360 for remarketing campaigns. 

Each event, metric and pagepath threshold was evaluated based on a specific lookback window and our process calculated the score for each user every day, according to the defined lookback window. The user is then assigned to an Engagement Level, to a language-country combination, to a product of interest and consequently to one of the groups: Touch, Tell, Sell. These factors are summarized in a string and pushed back to GA360 in a custom dimension via measurement protocol. 

In the second phase AutoML in GCP ran a model with a data set where the targets were the same as the first method – form submit and transaction. AutoML then provided a ranking of the features which we could use to reassign scores across all the features. We used a ML-model to look into all significant website events in comparison with ‘converted’ users. As a result the ML-model could predict the likelihood of conversion for each user – relevant to events up to that moment.