The severe limitations of manual audience selection in digital marketing can be overcome with predictive audiences estimated with machine learning methods and historical cookie data. Read on to see how RADLY has approached the challenge.
Digital marketing has taken huge leaps during recent years – both in the total volume, the proportion of programmatically bought media (direct buying and Real-Time Bidding – RTB) and the techniques of optimizing budget use. A critical factor in determining the efficiency of any marketing is the definition of audiences the advert is displayed for. A simple but inefficient way is to select such channels that the hypothesized users are relevant. However, typically only a small proportion of the reached audience is probable to be activated by the ad. In RTB digital advertising, one can define the audiences in multiple ways and with much more granularity.
The standard way to build digital campaigns relies on demographical or behavioural audiences bought from audience marketplaces and retargeting audiences defined by rules based on customer actions on a target website and browser type such as geographical location, device or language setting. This limits the efficiency of the built audiences to the capability of the planner. Often essential factors affecting the audience’s relevance are unknown or overlooked, and the resulting audiences contain only few customers that are activated by the ad.
Predictive audiences completely reverse the setting of RTB audiences. The principal idea is to learn the key factors affecting conversion probability from past actions and data of the customers. Rather than setting the rules by hand, machine learning is employed to find the recipe for the best performing audiences. The name, predictive audiences, implies another vital feature of the approach. Not only are the best performing audiences defined but also the probability of conversion is predicted to create a variety of different audience and action options. Should the marketer choose a limited audience with extremely high predicted conversions or a broader audience with lower conversion probability? What actions should be selected for each conversion probability group?
We at the RADLY are studying approaches to predictive audiences to overcome the problem of hand-built retargeting rules and prospecting audiences. The first criteria for success is to automate the audience selection process while reaching a similar level of performance as audiences created manually by human planners. The savings from reduced manual work are substantial. The second goal is to increase the performance of the predictive audiences further to significantly exceed the human capacity to create budget savings, increase revenue and ultimately, to enable more relevant marketing communications to the customers – who likes to be shown an advert with no relevance?
Methodologically the problem of predictive audiences is essentially a regression or a classification task. Historical data contains the touchpoints and marketing actions laid on the customer path as well as the possible conversion. Features are engineered from the customer path and used to explain the prevalence of the conversion. Regression methods such as Extreme gradient boosting trees (xgBoost), multilayer perceptrons (MLP) and recurrent neural networks (RNN) are prime candidates to solve the problem. RADLY pilots have shown promising performance with xgBoost due to its inherent regularization capabilities, which are necessary when dealing with the sparse and noisy customer path data in digital marketing. Pilots show that we can extract audiences based on their past behaviour with higher conversion probabilities than comparison groups. The next step is to automate the data processing and audience activation with DSP APIs to make the audiences to have the freshest cookies available for campaigns to use.
A fascinating insight was perceived from the predictive models. In many cases, features that were not related to the realized advert or website touchpoints were important in explaining the conversion propensity of a customer. This gives rise to an even more exciting application – autoprospecting. Prospecting for feasible customer candidates from previously unseen cookies is a difficult task, often requiring firing impressions across channels en masse to probe new potential customers. A further improvement to autoprospecting may include the inclusion of media provided general cookie activity data, e.g., browsed website categories and content.
With autoprospecting, the conversion probability can be estimated for any cookie – without the need to first gain ad impressions or client website activity. This way, prospecting can be limited to smaller and more relevant audiences to begin with, which will reduce budget waste and decrease the amount of unnecessary ad exposure.
Predictive audiences and autoprospecting are both very lucrative as well as realistic enhancements to programmatic digital marketing. RADLY aims to pilot and scale these applications to enable a higher level of work automation and to hugely increase the efficiency in digital marketing. As pilots have already been running, we expect to be able to present scaled use results and performance comparisons against human planners in the near future.
Contact the RADLY team for more information and pilot opportunities!