You Don't Need a Single 'Uber' Model to Make Data-Driven Marketing Work

By Published on .

Consider both traditional marketing mix models and new ways of attributing ads' effects, argues John Nardone. But don't try to come up with one 'uber' model.
Consider both traditional marketing mix models and new ways of attributing ads' effects, argues John Nardone. But don't try to come up with one 'uber' model. Credit: istockphoto

Ever since digital advertising emerged as a meaningful allocation of marketing budgets, brands have been frustrated with the way it has been measured.

Digital consistently seems underrepresented in marketing mix models, which don't properly account for digital's comparative role or provide timely signals for how to optimize digital spend. In response to these shortcomings, the industry has turned to "fractional" or "multi-touch" attribution models, which attempt to divvy up credit for results among each point of contact with a consumer along the way. But most marketers have found that, in practice, these have fallen well short of their promises, leading many to disregard them altogether.

In order to shed light on marketer's discontents, Forrester published its Marketing Measurement and Optimization Solutions Wave in October of 2016. Their conclusion? Marketers need an "uber-model," one that combines the characteristics of marketing mix and fractional attribution, all at once.

I respectfully disagree.

Marketing mix and fractional attribution are different tools aimed to answer different questions. While combining them might sound great in principle, in reality you wind up compromising speed and accuracy, as well as cost.

In our experience, what most marketers want is for each of their models to do what they are supposed to do -- and not contradict each other. Marketers are quite capable of using these models symbiotically, but separately, without the need to integrate them into one "uber-model."

As it turns out, the problem with marketing mix and fractional attribution models isn't that they are separate. The problem is that both are stymied by the fact that digital data sucks. You simply cannot build decent models from bad data.

"Now, wait a second," you say. "We are modeling with incredibly granular data! My logs capture every click and impression! Isn't that far better than the weekly, aggregated data usually used for mix models?" Well … no.

In truth, the current structure of log files are a terrible representation of marketing stimuli. What we need for modeling is to represent the marketing stimuli experienced by each person over time. But log files suffer from a host of problems against this goal. Consider:

  • Cookies fail to capture a big percentage of conversion events (in mobile and certain desktop browsers, for example, not to mention among users who routinely flush their cookies);
  • Cookies greatly distort our view of reach and frequency against devices, let alone against people. Log files include fraudulent impressions and clicks;
  • Ads that never had a chance to be viewed are included; and
  • Top-of-funnel marketing and bottom-of-funnel efforts represent different behaviors, but are all mixed together in the log.

In order for log files to be more useful for insightful and actionable modeling -- be it for market-mix or attribution -- the following steps should be considered:

  • Integrate a cookieless tracking solution to capture the full path of impressions, clicks, visits and conversions for each user;
  • Incorporate device graph data to capture consumers' experience and actions across their various devices;
  • Integrate user level verification data to remove fraudulent impressions and impressions that were not viewed;
  • Remove impressions and clicks that are delivered after the conversion event, and those that represent unreasonable frequency against individuals;
  • Unify and join the data for each user to assemble the time-stamped history of engagement for each converter and non-converter; and
  • Split the log file for display into two sets of data: one for top-of-funnel events, and one for bottom-of-funnel events.

If you take these steps, your "processed log file" will be a dramatically better representation of the marketing that consumers are experiencing. That better representation will generate more accurate and actionable models, even if the models themselves are the same ones you used before. Best of all, your marketing-mix and fractional-attribution models will complement and reinforce each when each represents the same reality.

You will then have better models that each deliver on their own strengths and purpose, and you won't need to chase the mythical "uber-model" in order to improve your digital marketing performance.

Most Popular
In this article: