Design a multi-touch attribution model

WHAT THIS TESTS The interviewer wants both a defensible attribution model and the data engineering to support it, recognizing that long, multi-channel journeys make single-touch models misleading.

A GOOD ANSWER COVERS Start with model choice. Heuristic options include linear (equal credit), time-decay (more credit near conversion), and position-based or U-shaped (heavier on first and last). Data-driven options such as Shapley-value or Markov-chain removal-effect attribution learn credit from observed paths and are harder to game but need volume and rigor. State the trade-off: heuristics are transparent and cheap; data-driven models are more accurate but opaque and data-hungry. For implementation, resolve identity so anonymous and known interactions map to one user, ingest timestamped touchpoints from ad platforms, web, and product events, sessionize them into ordered paths, and define a conversion event plus a lookback window. Then apply the chosen credit rule and roll results up by channel and campaign for ROI.

COMMON WRONG ANSWERS Falling back to last-touch, ignoring identity resolution so paths fragment, omitting the lookback window so ancient touches get credit, or proposing Shapley without acknowledging its data and compute demands.

LIKELY FOLLOW-UPS How do you validate an attribution model when ground truth is unknown? How do incrementality experiments or geo-tests complement attribution? How do you handle cross-device and privacy-driven identity loss?

ONE CONCRETE EXAMPLE A user clicks a paid search ad, returns via email a week later, reads a blog post, then converts after a retargeting ad. A U-shaped model gives 40 percent each to the first paid-search and last retargeting touch and splits 20 percent among the middle two. The pipeline stitches all four events to one user id, orders them within a 30-day window, and credits each channel so the marketing team sees that paid search, not just retargeting, deserves budget.