Marketing Attribution

Do you work in a place where every marketer reports their own success metrics, nothing is comparable, and when you add up everyone's contributions, you end up with 500% of actual sales? Why does this happen? What are the forms of marketing attribution? What can you do about it?

Marketing attribution can be confusing with every marketing platform, partner, and a crowded marketplace of vendors that want to sell you their solution. Every marketer has a natural incentive to grow their marketing budget (and therefore influence/career), so they look for the lens that speaks best to the contribution of their channel(s). Marketing partners also want to grow their share of media spend, so they stand up data science departments to prove the value of their channels and hire technical account reps to lobby for preferential treatment. These forces create an inherent bias towards overstatement.

Marketing is about meeting customer needs profitably. At the core of this, maximizing:

Customers Acquired * (Customer Lifetime Value - Cost of Acquiring Customers - Upfront Discount Investment)

In this equation, marketing has the most direct control over spend efficiency (spend / customers acquired) to acquire as many profitable customers as possible and budgets that vary based on how efficiently they perform. Sometimes marketing also controls pricing and discount, which leads to cross-channel cannibalization as discount seekers find the best deals online. Marketing has indirect influence on customer lifetime value in the quality of the customers acquired and how well those customers' needs align to product offered. With the most direct control over efficiency, for any given spend level, how many customers are credited (marketing attribution) to a channel makes most of the difference in how performance is measured.

Marketing attribution can be measured by digital platform (partner) reporting, offline media time lag models, last touch models, multi-touch models, voucher, surveys, media mix models, and lift testing programs. There is no perfect answer for marketing attribution, but some are less flawed than others.

Digital platform (partner) reporting typically works by taking credit for any customer transaction occurring within a window of when an impression was served and within another window of when an ad was clicked. The click window typically varies from 1-30 days and conversions are credited by a pixel placed on your site. Many platforms also have algorithmic credit models that they create, fully control, and will not explain the details of. Most larger platforms will not share data on which customers or orders they are taking credit for, so you have to take their word for it. In addition, platforms optimize towards their own attribution model's crediting system, which means they want to ensure they touch the people most likely to convert (rather than those that are likely to convert IF they receive that impression), which raises cannibalization rates. Because of the lack of transparency and misalignment of incentives, digital platform reporting is often the most inflated and least reliable form of attribution. It is also the most commonly used because large partners invest heavily in promoting these tools and making them highly accessible to marketers.

Offline media time lag models are offered when clicks don't exist in channels such as TV or radio. These models measure the time between when an ad is aired and when conversions occur, taking fractional credit depending how much time has elapsed. The most simple models have no baselining and simply take credit using an exponential decay curve. Better models establish a baseline and apply an exponential or gamma decay over the baseline. In either case, the decay model parameters make all the difference in how much credit is given. With proper client-owned lift studies used to internally set the exponential parameter or gamma parameters for partners to implement, offline media time lag models can be reasonable and timely measures of attribution.

Last touch models such as google analytics give credit to the source and medium based on the utm settings coming in from the last clickable media site that led to visit containing a conversion event. These models are somewhat limited to clickable media channels, as offline media might flow through direct traffic (and therefore not trackable back to the actual source of that traffic) or attributed to the most convenient channel that flowed traffic to your site. This creates a bias towards the most convenient clickable media channels. Because of tracking limitations, this may slightly underreport total attribution.

Multi-touch models are similar to last touch models, but give weight to impressions other than the last touch. These can range from simple allocation formulas that vary credit given by the part of journey to more accurate Markov chain models that evaluate the downstream impact of removing an impression from a user's journey. When implemented as Markov chains, these models can be highly accurate in the digital portion of a customer journey, but have blind spots for offline media. Because many large platforms refuse to share the individual level data and device manufacturers are moving towards reducing data availability to advertisers, these models are no longer viable without substantial work around probabilistic touch imputation.

Voucher uses discount codes associated to marketing channels to assign credit. In order to get a discount, a customer provides associated data around where they found that voucher. This method is more resilient to privacy and policy changes across platforms and devices. It shares similar downsides to last touch attribution in that it favors channels that are most convenient and clickable.

Surveys ask customers where they heard about your company, typically implemented in a checkout funnel. This has a similar advantage as voucher in resiliency to privacy and policy changes across platforms and devices. Surveys can be biased by the options offered (e.g. multiple options for similar channels and only one option for others) and for more complex marketing programs (with dozens of channels) cannot line up 1:1 with channel hierarchy and still be legible on mobile (where more than 10 options are hard to browse and honestly answer). Randomizing order helps remove some of the first box bias but will not address options stuffing.

Media mix models try to relate changes in spend with changes in performance. Most are regressions with some environmental controls. More advanced Bayesian models allow regression coefficients to be informed by "priors," where lift tests help set the probabilistic range of outcomes. Many companies use these because the most basic models are easy to implement and they are resilient to privacy and policy changes across platforms and devices. However, if marketing spend moves according to seasonal patterns, these models suffer from a problem of multicollinearity; because spends move together, attribution to individual channels is not measured accurately and the model strongly favors channels that only spend at high levels when seasonality is favorable. These favored channels tend to be inefficient by all other measures and are only used when there are no other channels that have room to scale. With years of regular lift tests informing Bayesian priors and appropriate environmental control variables, media mix models can be useful for steering the business.

Lift testing programs are the gold standard for attribution, but not all lift tests are created equal. Lift tests compare, during the same period of time, performance in a set of geographies or customers that were exposed to ads versus those that had no exposure. Sometimes groups are segregated by spend levels to measure media spend elasticity. The key advantage of lift testing is that you understand incrementality--how your marketing investment in one channel actually brings in customers across all other attribution methods. The disadvantages include time to run a test (typically 3-5 weeks) and potential performance loss while the test is running (control customers or geographies typically don't receive marketing from the tested channel).

In an ideal scenario, impressions are split only after a customer would have been exposed in the test group and withheld from the control group (after the impression bid is won) in ghost bidding and all impression (and won but withheld impression) level data is passed back to the client for validation. This is rarely (if ever) possible; platforms that offer ghost bidding typically insist on controlling all the data, meaning you have to trust their reporting in spite of the incentive problem. The least useful testing is in-platform testing using platform metrics, which suffers from fundamental reporting biases (outlined in the platform reporting section) as well as BAU-bias where campaigns already in flight are optimized, making it difficult for any other campaign or ad to beat an already running BAU campaign due to ramp up and competition from BAU. The best available option for lift testing is usually client-owned geo lift testing where geo group pairs are made to minimize the time series difference in historic variation, maximizing testing power at any given spend and testing timeframe. Channels that cannot be targeted by geography may require carefully timed pre-/post- difference in difference testing.

There are many forms of marketing attribution, each with its pros and cons. Understanding the range of methods and biases can provide healthy debate leading to an optimized marketing mix. The gold standard is regular lift testing to monitor channel performance periodically, mapping that to more timely attribution methods (like last touch or voucher) for weekly or daily steering. Other methods like surveys can provide additional color on offline channels. Finally, in conjunction with (and only in conjunction with) a robust testing program, media mix models can provide an additional spend-based lens between tests.