EXPEDIA GROUP TECHNOLOGY — DATA
Practical steps to develop a marketing attribution model
Imagine you see an exotic beach in an Expedia Group™️ TV advert. It prompts you to think about your next trip, and you search “Maldives hotels” using an online search engine, perhaps Google. From there, you click through to read reviews that help you narrow your choices, perhaps on TripAdvisor. The next day, after discussing hotel options with your partner, you directly visit the Expedia.com website to make your purchase. Of all the marketing touchpoints (TV, Google search, TripAdvisor, Expedia.com), which one was responsible for your purchase?
Travelers are exposed to a wide variety of marketing touchpoints, both online and offline. These touchpoints work together to shape shopping experiences and nudge customers towards conversions. To understand the return on business marketing investment and facilitate decision-making, it is crucial to be able to measure the contribution of each touchpoint.
To determine the value of marketing, we ask ourselves two questions:
1. What does the customer do after being exposed to marketing?
2. Would the customer have behaved differently if they had not been exposed to marketing?
There are three main marketing measurement approaches:
- Incrementality Experimentation
- Marketing Mix Modelling (MMM)
Attribution can shed light on question 1, while Incrementality Experiments can address question 2. MMM is a holistic measurement that provides top-level insights into both questions by measuring the impact of various marketing activities on sales/return.
This blog focuses on Attribution and the practical steps to develop an Attribution model. We will also touch on how Attribution fits into the Marketing Measurement Trifecta (Attribution, Incrementality, MMM).
What is Attribution and why is it important?
Attribution evaluates the impact of marketing touchpoints in a customer journey and aims to assign appropriate credit to each touchpoint.
It informs spending decisions through a timely and granular view of “return on advertising spend” (ROAS). It also enables the evaluation of marketing campaigns, highlights shopping path issues, and guides product development priorities.
The process of designing and choosing an Attribution model
Step 1: Define the scope
Choosing an Attribution model is a subjective decision and relies on business objectives. It can often be informed if we consider how customers interact with the business:
- Product sales cycle: How long is the purchase consideration window? There is likely a different consideration period for a replacement phone charger versus a family holiday.
- Customer type and visit behaviour: What is the primary purpose of the spend on a given marketing activity? Advertising can be good for acquiring new customers and retaining existing customers.
- Customer journey: Where does the business offering fit in the customer shopping journey? Customer intent moves from the upper funnel (undecided, search: “Maldives hotels”) to the lower funnel (decided, search: “Maldives hotel X on Y dates”).
It is advisable to determine what validation criteria will assess the model design:
- Accuracy: How good is the model at identifying which marketing activities led to the conversion and apportioning credit between them? A ground truth can be established through Incrementality Experiments.
- Fairness: Does the model fairly reflect the value of all marketing activities? Marketing partners are contracted under their own unique payment model that incentivizes specific shopping behaviour.
- Coverage: Does the model consider enough marketing activities? Attribution is only as good as the data you feed it, impression-based marketing would need third-party support to include those touchpoints.
- Relevancy: Does the model address defined business objectives?
It is also worth considering the following areas:
- Stability: Is the model consistent over time, and is it suitable across the markets and products in which the business operates?
- Customers: Are the outcomes aligned with expected customer behaviour?
- Interpretability: Does the model enable performance analysis investigations?
- Implementation: Can you afford both the time and resources for the implementation? Roll-out of a different Attribution model is not a quick task. Additionally, the more complex the design, the longer and more resources required to build and maintain.
Step 2: Design the model
If we look at the example customer journey above, three questions come to mind:
1. (Lookback window) How far should we look back to capture relevant user touchpoints? A short window risks missing signals, while a long window may introduce spurious touchpoints (eg relating to a previous purchase). Taking Product Sales Cycle (Step1) into account helps determine the lookback window. Stricter privacy restrictions have recently driven the trend towards a shorter period.
2. (Tiering logic) Out of all marketing touchpoints, what is the logic to determine how credit is distributed? Taking Customer Type (Step1) into account, we may decide to give greater weight to marketing that supports the customer-marketing strategy.
3. (Distribution of credit) We could credit multiple marketing activities (ie Multi-Touch Attribution) or have a winner-takes-all system (ie Single-Touch Attribution). Taking Customer Journey (Step1) into account can support choosing which type of Attribution model and position preference (eg last-touch v first-touch) is most appropriate.
Types of Attribution Model
Single-Touch Attribution: Rules-based models including:
- Last-Touch or First-Touch models attribute all credit to the last touch or the first touch
- Restricted Last Touch model attributes to the latest ‘prioritized’ touch.
Multi-Touch Attribution(MTA): Rules-based models including:
- MTA Uniform model attributes to all touches in the customer journey evenly
- MTA Custom-Shape model attributes more credit to ‘prioritized’ positions (e.g. U-shape favours first and last touches)
- MTA time-decay favours later touches than earlier touches.
Algorithm-Driven MTA models are complex algorithm-driven, machine learning MTA models such as Markov Chain MTA.
Some additional considerations during designing:
- Data acquisition: Should we buy data or build it ourselves? Third-party tools might provide more breadth of data and greater insights powered by advanced analytics capabilities. However, we would have less control over data, lose granular visibility and remove the chance of bespoke optimization.
- Data availability: How much can we see in the user’s cross-device journey? What we can use might be restricted by regulatory changes, browser updates, and privacy controls. One solution might be to rely more on user sign-in and less on cookies for tracking.
- Is Multi-Touch Attribution (MTA) the savior? Algorithm-driven MTA sounds exciting but, like many models, it is something of a black box. The business may struggle to support a model that cannot explain why performance declines. Rules-based MTA can be easier to interpret, however, it can dilute value by giving credit to non-incremental touchpoints.
- Cross-activity impact: How do marketing activities interact with each other? Paying attention to marketing collisions in the path may unlock further insights (eg Google Paid and Organic Search).
Step 3: Choose an attribution model:
The process of choosing an attribution model is certainly not straightforward, and there are some key considerations outside of the initial data insight that should be considered:
Using the right data in the model
It is imperative to sense check data-outputs against expected customer behavior; a purely data-driven approach may not identify these nuances:
- Non-marketing touchpoints: When a last-touch model design is preferred, credit may be given to activities that are not driving the purchase (eg “password reset” emails). Suppression rules may be required for this kind of touchpoint.
- Time lag for third-party data: There may be a delay in receiving third-party data that can distort the user journey (ie Is the last recorded touch really the last marketing activity used?).
- Cookie-stuffing marketing activities: A third party can drop multiple cookies on a user’s browser that do not truly represent marketing exposure. As such, these may need to be suppressed.
- User behavior: Differences in marketing performance may not be indicative of better advertising. It can also reflect customer demographic and self-selection eg Google v Bing users, generic search (“hotel”) v brand search (“Expedia hotel”).
Addressing competing interests of decision-makers
Attribution results are usually used to allocate marketing spend — it is a “zero-sum game,” and there may be perceived “winners” and “losers” of an Attribution change. Being ready to explain why marketing activities appear better or worse under different models is crucial. There are also differing customer journey focuses for each marketing activity: some might focus on the lower funnel (close to purchase), some on the upper funnel (exploratory), and some in the mid-funnel.
Building trust in model results
Attribution only works as a unified methodology when trust is built with stakeholders. Ensuring that everyone has confidence and is aligned around the chosen model reduces the risk of multiple incoherent methodologies being appended over time.
The challenges above make it clear how important it is to engage multiple teams throughout the process. We recommend the following best practices:
- Hold regular information and consultation sessions to introduce the project, align the identified business problems and share the potential benefits of moving to a new Attribution model. There is minimal risk of over-communication when making a change as big as this.
- Work with stakeholders to gather requirements and learn about any data nuances that are captured for their teams.
- Discuss operational considerations with stakeholders, including the development and deployment of the model. It’s worth noting that the choice of lookback window determines when results are available for analysis — a 30-day lookback window means waiting for a month.
- Provide interpretability of the data to inform decisions; build self-serve dashboards for stakeholders to visualize how potential models impact the areas that concern them.
What happens after choosing an Attribution model?
Step 4 & 5: Develop and roll out the new model
It is advisable to implement a new Attribution model against an existing one, running parallel operations before rolling out a full change. This allows for comparison and testing of the new model, hopefully ensuring smooth adoption and avoiding U-turn decisions.
Step 6: Review and calibrate
No single measurement can perfectly capture all media-exposure, conversion, and sales data, and give the business actionable insights on an ongoing basis¹. Attribution is part of the Marketing Measurement Trifecta, alongside with MMM and Incrementality Experimentation.
MMM takes a top-down approach and looks at the aggregated, macro-level view, while Attribution uses a bottom-up approach and focuses more on individual user interactions. Incrementality sits in between MMM and Attribution: it is more operational and tactical than MMM, but not as granular nor real-time as Attribution. Each measurement has its advantages and limitations, and they complement each other.
It is recommended to consider a unified approach, where Attribution, MMM, and Incrementality validate, calibrate, and enhance each other².
Shout out to Incrementality Allocation Analytics team, in particular, Jasmine Coll, Dario Nebuloni and Rory Bligh for their valuable input and support.
 Measurement Trifecta — A three-part approach to measuring media impact, by Google
Stay tuned and follow us to learn more about the other two measurements in the Marketing Measurement Trifecta: MMM and Incrementality!
Below is a high-level overview of the three measurements: