Half the money I spend on advertising is wasted; the trouble is I don’t know which half.
Online Marketing: Data-driven Control of Advertising Banners
Through the Use of Predictive Customer Journey
An ever-increasing proportion of online advertising is played out via Real-Time Bidding (RTB). In RTB, advertising space on the Internet is automatically auctioned off within milliseconds while a user loads a page. An advertising strategy that treats all users equally is not very efficient and leads to wasted resources, since some users already have a very high affinity for buying while others have little interest in the services of the advertiser.
The goal is to display the right amount of advertising to the right users. Success is measured by the sales achieved and is reflected in the efficient and profit-maximizing use of the advertising budget.
The basis of each algorithm is the data: As with attribution, the data consist of customer journey information. This includes data about the customers’ movement on the website, from their first point of contact on the page to the completion of their order. Ideally, these data are supplemented by off-site contacts (contact points of the potential customer outside of the company’s own website, e.g. banner advertising that was not clicked on), shopping cart data, CRM data, and, if necessary, other external data.
The first step is to select the relevant characteristics. The amount of data is often huge and there is a lot of noise in the bits and bytes, meaning that the really relevant information is often scarce. Therefore, experience is required to extract the relevant information. Tracking artifacts also often falsify the data. These problems must be identified and resolved before the actual analysis begins.
The data are first divided into training and test datasets. The former is used to determine the relationship between the available metrics and the purchasing behavior (specifically, the probability that the customer will make a purchase within a certain time frame) of individual customers using statistical regression models.
The interdependencies are validated using the test data set, which aims to show how well the model performs with new data. This is done by comparing the purchase probabilities predicted on the test data set with the actual observed purchases of the test data set.
The performance of the model on new data (called “out-of-sample quality”) is then optimized.
In a final step, the purchase probabilities are divided into various segments: For example, into segments 1 (well below-average purchase probability) to 5 (well above-average purchase probability). The model is then ready for use.
INWT has developed an algorithm that determines the purchase probability for a specific internet user in real time, and then allows bids to be controlled. The PCJ-algorithm enables a differentiated advertising strategy according to the probability of purchase, which considerably increases the efficiency of advertising campaigns. The budget can be targeted where the advertising medium actually works and purchases occur, and losses due to too much or too little advertising are minimized.
Our model is individually adapted to our customers so that it is optimized for the behavior of their specific users, internet presence, and advertising strategy. This adaptability is an essential strength of the approach, which leads to a performance that cannot be achieved with standard models.
The following figure shows the performance of the algorithm developed by INWT compared to the current industry standard using a randomized A/B test. The algorithm achieved an increase of 22% in ad impressions, and 24% higher number of sales (conversions) at the same level of financial expenditure. A major strength of the algorithm is that it is also able to correctly identify customers with buying affinity that may have mistakenly classified competing providers as uninteresting, and for whom the payout of advertising is comparatively inexpensive.
Note: If sufficient historical data are available, not only the purchase affinity itself can be calculated, but also the change in the purchase probability through the insertion of advertising (i.e. the effect of advertising). For example, advertising can be avoided with customers who have such a high affinity to purchase that they will also buy even without the aid of advertising.