This post is part of a series of case studies that discuss solutions implemented by Syntelli’s data science team. To read more of our success stories, visit here.
Client: Large Manufacturer of Home Appliances
Our customer, a large manufacturer of a variety of household appliances, needed to better understand the contribution of each of their marketing channels (direct, online display ads, radio and TV ads, billboards, and others) to their sales. In addition, they required an evaluation of the effect of each web page (such as product information, Where-to-Buy, FAQ, etc.) on the sales of different lines of product.
Client Challenge: Omni-Channel Marketing Attribution
The common solution to this kind of a problem was to use attribution modeling algorithms that exist in the Google Analytics engine. Marketing analysts defined ‘number of consumers landing on the Where-to-Buy landing page’ as the conversion metric, and relied on the built-in algorithms in Google Analytics to provide the effect of each marketing channel based on the defined conversion metric.
As helpful as this was, a straightforward solution was not enough. Not all consumers who bought an appliance viewed the Where-to-Buy landing page. Moreover, attribution modeling through Google Analytics evaluated only the channels used before conversion, and not web pages that consumers could have visited before their purchase. Furthermore, consumers may have also viewed other web pages after purchase. For example, they could have visited a page to download a manual or register an appliance.
Therefore, Syntelli’s data science team had to adopt a strategy, taking these challenges into consideration, to provide a realistic model displaying the relationship between sales and consumers’ website activity.
Our Solution
As the first step, we extracted website activity of consumers which was tracked by MixPanel.
Why use MixPanel instead of Google Analytics?
While Google Analytics only tracks web pages, MixPanel tracks every action a person takes on the website – like viewing a page, registering an appliance, or downloading a manual. It then assigns each of them to a unique id.
Our attribution model needed to navigate how the consumer was exposed to media and site content, and in what order the exposure occurred. We had to have our data match their online journey. Fortunately, we were able to use a graph database instead of a more traditional relational database system for this purpose.
Figure 1. Structure of the Graph Database. Several types of events are assigned to each user including viewing a page, buying or registering an appliance, or downloading a manual. Each of five nodes and edges between them has some properties such as time, city, etc. The Cypher language is used to run query and extract necessary information from the database.
Graph Databases are a relatively recent addition to the Data Science tool kit. It allows us to not only define relationships between our data, but also assign attributes to the relationships themselves. When we add ‘time’ as one of those attributes, it makes simple the once complex task of mapping in-site user activity. We implemented our graph database, in Neoj4. The information from MixPanel was stored in the designed graph database.
As our next step, unique IDs and purchase dates of people who bought at least one appliance were extracted, and channels and web pages that buyers had viewed before and after purchase date were listed.
Now that we had our data structured in a graph, we were able to apply a more comprehensive attribution model – page views and channels before and after purchase date.
Using a method known as Bayesian Attribution, we developed a probabilistic graph of all the different media and site content journeys leading to a sale. Using this graph, we determined the reduction in probability of sale by leaving out a step. The change in probability gave us the value of a given media channel or piece of content in the journey – i.e. its sales attribution.
While our enhanced attribution modeling method provided useful information for our customer about the effect of their marketing channels and web pages on sales, other interesting information about consumer behavior was also extracted. It included percentage of consumers who used the Where-to-Buy page to purchase, time between page visits and purchasing, as well as patterns in purchasing different appliances by a consumer. The manufacturing client used this information to improve their website structure, build their recommendation system and optimize their advertising campaigns across multiple (offline and online) channels.
