Digital Marketing ROI: From Clicks To Causality

Forbes

a day ago

Rahul Wankhede, Director Data Science and Marketing Analytics, Humana.
getty
Digital marketing is an indispensable engine for growth. Year over year, big brands allocate a significant portion of their annual marketing budget toward digital channels and—with an ever-increasing digitally savvy and available audience—effective strategies can promise precision, personalization and reach.
While it's easier to spend in digital channels than traditional ones, proving return on investment (ROI) at both the top and bottom of the funnel is complicated. That complexity is further amplified when you think of the impact on brand awareness or recall, which may not always yield immediate conversions.
The real question is incrementality: Did the ad cause the action, or would you expect it to occur organically?
To understand the trade-offs, think of measurement approaches as a funnel, ranging from simplest at the top to more complex—and possibly more accurate—methods at the bottom. With each layer, you can expect more rigor but also more stringent data requirements, sophisticated models, investment, deeper analytical expertise and patience.
Here are a few different approaches for measuring the ROI of digital marketing efforts:
At the top of that funnel is last-touch attribution, a model that assigns full credit for a conversion to the last interaction before the action.
Since it's easy to implement and interpret, last-touch remains a popular method. However, the methodology ignores all other touchpoints that may have influenced the customer along their path to conversion. The method doesn't include the causal impact of interactions like first touch, linear or time decay.
That's where multi-touch attribution (MTA) comes in. MTA attempts to use user-level data and models to assign fractional or partial credit to each touchpoint. The approach attempts to account for the full customer journey, offering a more granular view of performance. This method typically includes large volumes of granular data such as cookies, device IDs and timestamps across multiple channels. The model is also becoming more difficult to manage, with the world moving toward more stringent privacy laws and cookie deprecation.
Ultimately, even with the data stack available, MTA is prone to bias. Some consumers who view your ads are likely to inherently be more engaged or loyal, while some may just be online more, potentially leading to overstated impact. Without proper controls, MTA could mistake correlation for causation.
To supplement attribution, marketing mix models (MMMs) are a popular technique that leverages aggregated data—typically at the geographic and weekly level—to estimate the impact of marketing on sales. MMM can capture both online and offline media impact and is valuable when you add saturation curves to assist with long-term planning and budget optimization.
While MMM does not require user-level data, which makes it resilient to privacy changes, results can skew heavily based on model specifications, assumptions and data quality. It also tends to smooth over short-term fluctuations and isn't necessarily useful for more tactical-level optimizations.
Randomized control trials (RCTs) remain the gold standard in measuring the incremental impact of marketing. They work by comparing outcomes between a treatment group (those exposed to the marketing) and a control group (those not exposed), isolating the causal effect of the campaign.
There are several ways to do this:
• Randomization can occur at the market, geographic level or individual user level, where users are assigned to either a treatment or control group.
• You can use synthetic controls after a campaign, which constructs a comparison group based on modeled data when a true control group wasn't established in advance.
Running and implementing RCTs at scale, where reach remains such a crucial factor, is a challenge.
You also run into issues with exposure bias; for example, being in the treatment group doesn't necessarily mean they saw the ad. Other factors, like viewability and fraud, can skew results. Conversely, control users might inadvertently be exposed to the ad due to retargeting, low match rates or media leakage.
These issues can compromise the validity of the test. One solution is the implementation of placebo or ghost ads, which appear as real ads to the user but promote unrelated content, serving as a true control. With ghost ads, the control group is selected at the time of the bidding process, solving the question of ad exposure.
This method requires a robust internal data infrastructure—including impression logs, user IDs, clean conversion data and analytical expertise. Many organizations don't have the engineering resources to deploy these tests at scale.
Still, despite these challenges, well-executed experiments can validate and supplement predictive models, inform bidding strategies and answer key questions about channel effectiveness. When paired with MMM or MTA, RCTs act as a calibration layer, bringing rigor to existing methods.
While much of marketing measurement focuses on conversions, upper funnel metrics like brand awareness, consideration and recall are equally critical. Though harder to measure, these indicators play a significant role in driving long-term growth.
Surveys remain a primary tool to measure brand awareness, comparing consumer perceptions before and after ad exposure. Modern brand lift studies use randomized control designs to capture these shifts more accurately, while longitudinal brand equity trackers monitor changes over time to link campaigns with brand health.
These methods offer insights that performance metrics alone can't answer, essentially the emotional connection with an audience the brand is trying to win. They also come with challenges: response bias, sampling variability and difficulty with linking surveys to actual exposure.
Each approach—attribution, experimentation, modeling, surveys—has trade-offs. The key is integrating them into a unified strategy aligned with business goals, campaign needs and data maturity.
Success lies in using the right mix at the right time, validating results without unnecessary complexity and building on insights over time. While machine learning will help automate and surface patterns, human expertise remains essential. The complexity is real—but so is the opportunity.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?