The premise of marketing attribution is if we understood the contribution of each marketing dollar spent, by channel, we could make decisions that would result in optimal outcomes.
In general, those outcomes are geared towards maximizing one beneficial result while minimizing one particular cost. There are two dominant approaches to attribution: econometric and direct.
The Econometric Approach: Quantifying Patterns
Econometric approaches are based on detecting and quantifying patterns in existing, historical data. These are sophisticated, observational, approaches. Data from CRM systems, usually in aggregate, are combined with spending information. If a spike or trend in a cost is correlated with a spike in a revenue time series, it is inferred that the two are related. Comparing the strength of these signals and explaining them yields a model. If a market is exposed to a massive frequency of messaging, and revenue increases, then attribute the increase in revenue to that frequency of messaging.
The Direct Approach: Attributing Individuals
The direct approaches are based on detecting and attributing individuals. These are sophisticated, experimental, approaches. If a consumer is exposed to a banner ad, place a cookie on that consumers’ browser. If the consumer converts within, say, 90 days, using that browser, and the cookie is still on that browser, then attribute that conversion to the banner exposure.
Recommendation Engines: Best of Both Worlds
Both approaches leave a lot to be desired. Econometric approaches aren’t very well suited to picking up on longer run branding effects. Direct approaches suffer from instrument and counter-factual challenges: cookies get deleted, consumers use multiple devices, and not all of the factors are controlled.
Recommendation engines offer two streams of data that should appeal to both econometric modellers and to direct experimentalists. A consumer auth’s-in. The recommendation engine generates stimulus. The consumer can make decisions in response to that stimulus. Two streams of data are created: an aggregate one that can be used for econometric modelling, and individual identities of people (not cookies).
The use of the auth-in, with consumer permission, certainly makes for better experimental outcomes at the marketing treatment level, and, an information stream can be used for econometric-CRM approaches. In other words, the linkage of marketing treatments in one channel can be reconciled at a higher fidelity with CRM systems. It’s cleaner than a cookie and better than a pure display ad. The use of recommendation engines has a really important contribution to make to the attribution effort.