Understanding attribution in online advertising

Since 1994, when AT&T’s first clickable banner ad (ad) was placed on HotWired.com, online advertising has spread like wildfire. Every individual with a smartphone or a computer is exposed daily to a deluge of advertising messages. With about 5 billion internet users and 4.65 billion active social media users in April 2022, online advertising is big business. In 2021, digital marketing grew 35% to $189 billion.

While online advertising has grown rapidly, a major challenge remains: the problem of attribution. How can an advertiser attribute value to marketing activities across different channels and media? Attribution involves evaluating the contribution of individual advertiser actions (ad actions) – such as display ads, paid search, and direct emails – to the eventual purchase.

The question of attribution guides many marketing decisions. And yet, despite the colossal sums invested by advertisers in online advertisements, the methods to justify these “investments” have no theoretical justification. In our study on attribution in online advertisingmy co-authors* and I propose a new attribution metric with desirable properties, which overcomes the shortcomings of existing heuristics.

What is the question to your answer?

In the online advertising space, marketers and advertising agencies make complex decisions about advertising spend, such as allocating budgets between advertising channels and media, as well as optimizing tactical decisions for each channel. Contribution, or the value of each channel, is an important part of optimizing the media mix. It helps to better understand the customer journey and helps a business justify its marketing spend.

Commonly used rule-based methods assign generated value to different advertising actions on the user’s path to purchase based on predetermined rules. For example, the last contact rule assigns all generated value to the last ad action, while the uniform rule assigns the same value to each ad action. Assigning custom weights is sometimes applied to tailor weights to a series of advertising actions.

Although these heuristics are easy to implement and understand, they are neither accurate nor systematic. Awarding the last touch is akin to the common winner-takes-all narrative in football, where the player who scores the goal gets the glory. But what about the player who made the final pass, the team effort and the interplay of actions that made that goal possible? Can we give credit where it’s due, instead of relying on simplistic metrics that don’t fully capture the value of each link in the chain?

For example, it is possible that a user who is inclined to buy a product performs a Google search and clicks on a paid search, leading to a purchase. However, even if the paid search click did not happen, the user, who is in a state of “desire”, might have purchased the product by finding the product link through organic search, in what is called the counterfactual scenario. Currently, existing methods do not consider causality or counterfactual effects. Since a purchase may be driven by external factors or a customer’s precondition, attribution methods should consider the baseline or expected outcome if the user had not been exposed to the announcement.

Fundamentally, with current attribution methods, the key question to justify ad spend remains unanswered: how much does a specific ad influence the customer’s purchase decision?

A new attribution metric

Attribution is not unlike awarding credit to individual players in a cooperative game, since the value generated by online advertising is the result of the cooperative effect of actions taken across channels and media platforms.

Borrowing concepts from cooperative game theory, we propose a new attribution metric, which we call a counterfactual adjusted Shapley’s value (CASV). Our method inherits the desirable properties of the traditional Shapley value while overcoming its shortcomings in the context of online advertising.

In particular, our method captures causality and takes into account the difference in value when the customer saw the ad and a hypothetical benchmark where the customer was not exposed to the ad. Unlike the rules-based method, no value will be assigned to an ad that has no effect on the customer’s purchase decision. Our method is also fair: two advertising actions that have the same effect receive the same value.

With the availability of user data in the online advertising environment, our new attribution metric is computationally viable even for a large scale data set. The massive volume of economic activity taking place in the digital realm has enabled the collection of data on an unprecedented scale, allowing businesses to leverage user data to better understand customer behavior and improve service quality.

In our study, we showed the applicability of our method and validated the robustness of the model using real data with several million user journeys and a few hundred thousand purchases (conversions).

Start with the customer journey

When it comes to customer conversion, advertising actions can have disparate impacts on conversion, depending on where the customer is in the conversion funnel commonly known in the marketing literature. Through the lens of the customer journey and with the insights provided by user data, we can better understand customer behavior based on their “state” – from product ignorance to knowledge, interest and conversion.

Accurately defining the client’s condition is a crucial part of this method. Borrowing concepts from the conversion funnel, we estimate the value generated by each advertising action assuming that the customer’s browsing behavior is Markovian in nature, that is, we assume that the customer history can be summarized succinctly in the current state.

For the advertiser, the observed state serves as the basis for appropriate advertising actions and can improve the chances of reaching the desired customer demographics at the right time.

From the right question to the right metric

Having determined the “right” question advertisers should be asking, advertisers now need to better understand the metrics we choose to drive the business.

In our approach to attribution, we propose to start by defining the required properties of the attribution method. Based on this approach, our proposed metric is the only one that satisfies the four desirable properties: efficiency, linearity of assigned value, symmetry between channels with the same behavior, and assignment of zero value to non-contributing channels. Even if a different set of properties were more appropriate in a different context, the techniques from our study could be applied to derive an appropriate method.

Since metrics define the “rules” of the game in the complex landscape of online advertising, defining metrics is an important decision. However, it is not sufficiently questioned. What principles are they based on and what are the implications for others? Ultimately, attribution developments are a work in progress and asking the right questions is a good way to start.

* Raghav Singal, Tuck Business School; Omar Besbes, Columbia School of Business; Vineet Goyal and Garud Iyengar, Columbia University.