What is UMM and how does it differ from traditional MMM or MTA?
Unified Marketing Measurement integrates econometric analysis (marketing mix modeling or MMM), behavioral customer analysis (multi-touch attribution or MTA), and in-market testing to drive a complete and consistent picture of marketing effectiveness. Typically, it also is more ambitious in scope, taking long-run brand effects, pricing, and distribution choices into account to paint a complete causal picture of revenue. UMM is much faster than typical MMM, because econometric results are driven into an MTA-like longitudinal human (or account, in the case of B2B) record. At the same time, it does not suffer from “offline blindness” as does MTA.What types of data does UMM integrate (e.g., media, sales, CRM, external factors)?
The best UMMs integrate a comprehensive set of stimulus data—anything that might drive sales (or whatever your dependent variables of interest is). These data typically include first- and third-party paid digital media touches; “above the line” advertising; earned media and PR; CRM touches like email and direct mail; pricing; distribution channel; macroeconomic factors; and competitive activity.
The key difference from traditional MMM is that these are gathered at the lowest level of granularity possible—ideally the individual, time-stamped level—and then only aggregated later when required. This ensures that a consistent set of training data for both behavioral and econometric models.
UMMs should also integrate a formal ledger of test results, in a machine-readable format like JSON. As new test results are acquired, they should be efficiently ingested into the ledger, and then used as priors for the unified model.
How does UMM handle cross-channel attribution?
Because UMMs look at all distribution channels—for example, retail, direct, and platform—they must take cannibalization into account (or, in rare cases, a beneficial effect.) In essence, when a channel like retail is added, this typical depresses sales through, say direct. Reach and price are the key drivers of this effect, but UMMs should give channel strategists what they need to make route-to-market decisions.
How does UMM handle incrementality?
Incrementality should be the standard output of the UMM. In other words, each reported metric—such as ROAS (return on ad spend)—should (1) not be duplicative with any other stimulus channel, and, (2) should take the base, or “what would have happened without marketing”, into account.
This can be a challenge with MTA solutions, which typically start with a sale and work backward to divvy up credit to the various channels that “touched” the sale, but UMM solutions solve for this by deriving causality from more sophisticated econometric (MMM) analyses.
What modeling techniques are used?
UMMs rely on Bayesian regression for the MMM step, causal inference for the testing step, and machine learning techniques like logistic regression or Shapley values for the MMM step. Each solution uses slightly different methods, but the most important thing to look for is parsimony. Overly complex methods can easily create nice looking but incorrect results.How frequently is the model updated and refreshed?
The MMM component of a UMM, which is the main causal engine, is typically refreshed at three levels. First, daily contribution re-estimations can be accomplished easily without rebuilding coefficients. In essence, this means that each channel’s impact will change based on its spend and the sales observed, but the causal relationship between spend and return will not be updated.
Second, coefficients can be reupdated frequently as new data come in. This “mid-level” refresh uses the same model structure, but allows the causal “power” to change, between, say, paid social video and sales. This is a lighter lift and might be redefined weekly.
Finally, the overall model structure—which includes the taxonomy of stimulus channels and how they interact—is typically rearchitected quarterly.
The beauty of a UMM solution, however, is that the MTA component is updated in near real-time. While the causal inputs might not change as frequently, a CMO should be able to understand marketing’s performance on any given day.
How does UMM integrate with our existing tech stack (BI tools, CDPs, cloud platforms)?
UMM models should sit on top of an owned marketing data lake that takes inputs via API from martech tools and CDPs, but should not exist inside these tools. Martech tools are inherently siloed—even if vendors claim to have a comprehensive view. Outputs can certainly be integrated into existing BI tools (like Looker or PowerBI).What is the onboarding process like, and how long does it take to be fully operational?
Universal Marketing Measurement is not software. It is a system and way of thinking for marketing analytics that includes data, people, models, processes, and reports, so “onboarding” is probably not the right way to think about. That being said, a UMM approach should start producing results within six months, and should be fully mature after about a year. A typical realistic timeline:- Months 0-3: Data lake built and starting to be hydrated; draft marketing mix model built; first tests designed
- Months 4-6: First “final” MMM model built; priors ingested from initial rounds of tests; draft MTA model built
- Months 7-9: First “final” MTA model built; priors ingested from MMM; initial real-time results flowing
- Months 10-12: Results tuning; testing “factory” up and running; real-time results embedded in marketing and financial dashboard for the company
Can UMM handle both online and offline data sources?
UMM is tailor-fit to handle both online and offline data sources. Because the MMM portion can estimate any stimulus impact due to its use of econometric time series analysis, all marketing touches’ impacts are estimated. Offline touches are then probabilistically imputed into the MTA portion to get a real-time view of impact. Tests are also perfect for measuring offline impact. For example, geographical tests with synthetic controls can be used to measure the incremental impact of almost anything—and these results can then be imported into the MMM and ultimately the MTA component to understand impact.How does UMM quantify marketing ROI across channels and campaigns?
Because UMM calculates ROI at the individual level (whether a new customer, won deal, or sale), it is relatively straightforward to calculate return on any dimension. A well designed and enforced taxonomy is critical, of course.
Concretely, say we wanted to calculate ROI for the Social-Video channel. We would sum the incremental contribution (in dollars of revenue or profit) of Social-Video across the timeframe of interest, and then crosswalk back to the promotional table for each of the attributed Social-Video part-worths to understand working dollar spend that drove these part-worths. At this point, ROI can be attained by dividing contribution by spend. These calculations can be scaled across the entire dimensionality of stimulus and response—including channel, campaign, product line, content, and basically any other relevant cross-section or categorical variable.
