In several courses, I have introduced and discussed the idea of (a) a logic model, and (b) a twist on the attribution model. I am still working through this idea, but feel it is formed enough to warrant a post. The logic model offers a simple and graphical approach to showing behavior because the marketing manager can move the customer through a purchasing process where the outcome requires some sort of measurable value. These outcomes could include:
- Completing a form
- Downloading an app
- Reading a white paper
- Watching a demonstration vidcast
- Participating in a webinar
These outcomes should provide some measurable value. As to the exact value - except purchase - I am leaving for a different entry and discussion. For now, I am focusing on purchase.
A logic model requires probabilities of outcome behaviors, or customer actions. The probabilities can be developed from prior efforts. For example, to develop probabilities for a an email campaign logic model, the marketing manager should examine ratios, or probabilities, from previous email campaign. From the prior effort, the probabilities for opening the email, clicking the link to the website, and making the purchase could be determined.
Next, I add a layer of complexity to the logic model by incorporating only direct revenue and costs of the campaign. I attribute the campaign cost only to those customers who purchase as a result of the campaign. I treat the remaining amount of campaign cost as residual, or waste in a pejorative sense of the word.
I calculate a net marketing contribution for the campaign by subtracting cost of goods sold, or cost of merchandise sold, and the attributable campaign cost from the attributable sales revenue. With this figure, I can determine the marketing ROS and the marketing ROI, or ROMI, for this campaign. These figures are preferable because I am only considering the customers who purchased and not the customers who exhibited all behaviors except purchase.
At this point, the effort remains incomplete. Instead of an exact amount for purchase, a better model would use a confidence interval. Similarly, a better model would report a power value to determine the probability of a type II error would occur. Hence, the model could be testable. That is, a research or analyst could compare expected to actual with a t-value.