“Brand Aware” explores the data-driven digital ad ecosystem from the marketer's point of view.
Today's column is written by Jake Bradbury, director of data science and marketing analytics at Nordstrom.
In the May installment of Brand Aware, Amy Loesch did an excellent job of describing the reasons for and potential challenges of using multi-touch attribution (MTA).
Based on my experience, one of her considerations jumped out at me: Developing a robust customer identity graph is central to success when building attribution internally or working with a third-party vendor.
Connecting a customer’s marketing history and subsequent transactions is as important as it is difficult. Data scientists and MTA-solution providers tend to emphasize the math and technology driving their approaches. Certainly, the underlying math is important – measuring incrementality, not just correlation, is critical to support efficient marketing.
But before any modeling and analysis begins, models must leverage a customer identity graph. Even if a brand elects to use a vendor’s MTA solution, that solution still needs a custom approach to integrate first-party data. Without a robust graph, any MTA solution will cover a fraction of sales, likely of a brand’s best-known customers – hardly a representative sample.
Brands that service customers digitally and in stores face the added challenge of linking transactions that do not require customers to provide any information when making a purchase. Even with robust solutions from onboarding companies, this is very difficult to do.
Supplementing perfect data with good data
Despite these challenges, brands can build a reliable mapping of marketing touchpoints to customers and their transaction histories. Start with what you know, then supplement with additional information using reasonable assumptions.
For some customers and marketing touchpoints, there will be different levels of certainty in the data. That is not necessarily a problem, because for marketing measurement and optimization, we typically to not need 100% certainty. (Exceptions do exist depending on how companies use the measurement to target and customize future communications.)
The obvious goal is to link as many transactions as possible to marketing. On the customer-transaction side, website login is a direct way to tie transactions and site browsing behavior to customer identity. For brick and mortar sales, a robust loyalty program serves as a de facto store login. The rewards associated with the loyalty program incentivize people to identify themselves when making a transaction. At Nordstrom, this makes up more than half of all sales – 51% in 2017.
Email receipts and programs for purchases made in-store but shipped to homes provide yet another path between store transactions and customer identity. Finally, for transactions without a deterministic link to an individual’s customer identity, location information can tell marketers much about the customer who made the purchase. For example, using demographic and geographic data, it is possible to estimate a history of marketing exposures based on campaigns’ geographic reach.
For linking marketing history, clickstream data with proper tagging provides reliable information about the marketing campaigns and channels clicked by a customer. Marketers can use CRM systems to supplement with email, direct mail and other addressable marketing contacts.
Next, onboarding service providers linked to ad-serving platforms can provide log files with impressions at the customer level. These log files show which customers were served paid media impressions, not just those who clicked. Integrating this data is one of the big selling points for many MTA vendors.
There will still be certain digital platforms, such as walled gardens, for which customer-level data is not available. Often, those platforms will still provide customer-level view-through data. These transaction-level reports show which converting customers were served an ad prior to purchasing.
Losing visibility of exposures for people who did not convert will have implications for model training, but view-through reports are still useful for building a customer’s marketing history.
Finally, some digital impressions and broadcast media exposures will have no tracking to devices or individual customers. For these cases, marketers can estimate exposures to customers based on the campaigns’ geographic reach and customer location.
To make this all more concrete, let’s consider some hypothetical examples, beginning with an online transaction.
For this purchase, we know the customer from his login and can tie him to cookie-session data, including digital marketing clicks. If his historic session data persists, we can supplement his marketing history with other associated cookies through logins or email clicks. We can also match him to any impressions tracked with an onboarding customer ID in our ad server log files or from platform view-through reporting.
From here, we need to supplement known data with estimated ad exposures. Let’s assume that we ran a TV campaign in the DMA where he lives, and based on the GRPs, we estimate that the campaign reached 15% of the population in that DMA. We would assume the customer had 0.15 exposures to the TV campaign.
For an in-store transaction, we follow a similar approach. If the customer self-identified at the point of sale, through a loyalty program or by providing other personal information, we can find cookies and marketing clicks using a persistent record of logins associated with that customer.
Associating log-file impressions would follow the same process as with an online transaction, as would estimating exposure to broadcast media.
Lastly, for an in-store transaction with no way to identify the customer, we would still know where the transaction took place and could therefore estimate exposure to broadcast media based on geographic reach.
For marketers, understanding who saw what before purchasing will allow them to communicate with customers in more meaningful way. Limitations on how certain marketing platforms track that communication mean that this data won’t be perfect. However, we can still do much more with the data available today.