Describe the architecture for multi-touch attribution with time-decay

WHAT THIS TESTS: This question tests whether you can design a distributed data pipeline that solves the identity resolution problem and supports incremental attribution modeling. Interviewers want to see that you understand event tracking at scale, the difference between anonymous and authenticated user states, and how to transform raw clickstream data into business-level metrics without collapsing everything into a single last-touch report.

A GOOD ANSWER COVERS: First, identity stitching across channels. You need a unified identity graph that links anonymous device IDs, email hashes, and logged-in user IDs through deterministic matching and probabilistic fuzzy matching. Second, event collection and ingestion. Use a first-party event stream like Segment, Rudderstack, or a custom Kafka pipeline that captures every touchpoint with a canonical timestamp, channel source, campaign ID, and cost data. Third, data modeling in the warehouse. Load raw events into Snowflake or BigQuery, then use window functions to sessionize journeys and build a user-level funnel table. Fourth, the attribution model itself. For time-decay, apply a half-life function in SQL or a Python model so that touchpoints closer to conversion receive exponentially more credit; a common default is a seven-day half-life. Fifth, serving and validation. Expose the model through dbt marts or BI tools and validate it against holdout experiments to ensure the incremental credit assigned to each channel aligns with lift tests.

COMMON WRONG ANSWERS: A red flag is suggesting that GA4 or Adobe Analytics alone can solve multi-touch attribution at the user level, because these tools sample data and rely on device graphs you do not control. Another red flag is proposing nightly CSV dumps from ad platforms into Excel; this breaks identity resolution, lacks sessionization, and cannot scale past a few million rows. A third red flag is ignoring the distinction between correlation and causation by building a decay model but never running incrementality experiments to validate it.

LIKELY FOLLOW-UPS: The interviewer may ask how you handle the deprecation of third-party cookies and mobile device IDs. They may also ask how to attribute offline touchpoints like TV or direct mail, which requires probabilistic matching or geo-lift experiments. Another common follow-up is how to update the model in near real time rather than batch, which pushes you toward streaming warehouses like Materialize or Flink. They might also ask how you would allocate a fixed marketing budget given fractional attribution output, which tests your ability to connect data science to financial optimization.

ONE CONCRETE EXAMPLE: Imagine a user sees a Facebook ad on day one, clicks a Google branded search ad on day three, and converts after a promotional email on day seven. In a last-touch model, email receives one hundred percent of the credit. In a time-decay model with a seven-day half-life, the email might receive roughly fifty percent of the credit, the Google click roughly twenty-five percent, and the Facebook impression roughly twelve percent, with the remainder distributed to any minor touchpoints. The SQL implementation would join the conversion event to all prior touchpoints in a thirty-day lookback window, compute the age of each touchpoint in days, and apply the decay formula credit equals two raised to the power of negative age divided by half life.

Source: gitnexa.com