Our clients invest millions of dollars in digital sites, apps, media and social ventures. But with the distance often so great in time and space between a digital interaction and the eventual purchase of a car, or a piece of electronics, or a bag of dog food, it was impossible to understand which specific content pieces in digital had the largest influence on an eventual decision to purchase.

Our challenge was to come up with a stand-alone model that could break digital interactions into core components and derive the relative value they contributed to the eventual purchase of something, even if that purchase happened offline and weeks or months after the interaction. This data would be used to optimize creative in every channel so that our clients were always showing the most persuasive content to its viewers. We would know that a retweet on Twitter would drive $3 of value to an eventual purchase while a like on Facebook might only drive $0.25. We would know that content about features was less valuable as an influence to purchase than content about brand. We would know that a coupon might drive $3 to a sale, but actually detracted from the lifetime value of a customer. With this information, we could prioritize spending and optimize content at a fidelity previously unknown.

We constructed our model using a team of in-house data scientists and statisticians, and have production-tested and validated models for clients across auto, consumer electronics, and CPG verticals.

The models we have built are an evolution of more prototypical attribution models. We marry behavioral data across site, social and mobile properties, with offline sales data, qualitative influence and perception data, and industry constants to derive the actual dollar value earned from any digital activity (e.g. a click, a download, a share, etc.). The model looks at every single thing a consumer can do. Be it on site, social, mobile or media, the model ranks and optimizes those channels, content themes, and experiences that drive directly to sales- even if those sales are happening offline. We can also run data for specific marketing segments through a model to understand if a piece of content that is underperforming generically, is actually very persuasive to a subset of customers (i.e. according to loyalty, geography, purchase path, etc.).

The models themselves are relatively lightweight to construct and are updated and adjusted quarterly to account for changes in competitive landscapes and customer needs. The more data we can add (for example, CRM data or lifetime value trends), the more accurate the model becomes at predicting trends in longer sales cycles.

The Relative Value Model has become the roadmap for the creative teams, who now have an explicit understanding of the pieces of content that made the most difference and impacted a purchase. This has also proven instrumental in determining what channels should get priority investment based on likely impact to future sales. For one auto client, we found that a particular set of video views were on the chopping block due to the high cost of production. When we ran the videos through the model, however, we were able to see that watching those videos was 55X more likely to get a customer to a dealership than any other single digital action. Rather than halting production, the auto client used this data to justify not only continuing the program, but they used it to create new media campaigns that specifically drove to the program.