Architecting multi-touch attribution
Identity stitching and credit assignment.
Ingest touchpoints, resolve to one identity, order into paths, apply a model; last-touch is trivial, time-decay needs the full path.
Ignoring cross-device identity.
WHAT THIS TESTS Whether you understand that attribution is primarily an identity and sequencing problem before it is a modeling problem, and can contrast the engineering cost of two credit rules.
A GOOD ANSWER COVERS The pipeline has four stages. Ingestion collects touchpoint events from each channel: ad clicks, social engagements, email opens or clicks, with timestamps and channel metadata. Identity resolution is the hard core: the same person appears with different identifiers across devices and channels, so you stitch them into one entity using deterministic keys such as login, hashed email, and click ids like gclid, plus probabilistic matching as a fallback. Path construction orders each user's resolved touchpoints by timestamp up to the conversion, producing a sequence. Attribution then assigns credit across that sequence, and the results feed reporting on channel ROI.
LAST-TOUCH VERSUS TIME-DECAY Last-touch assigns all credit to the final touchpoint before purchase. Implementation is trivial: you only need the most recent qualifying touch, so even an incomplete path or weak stitching still yields an answer, but it systematically ignores assisting channels and overcredits bottom-of-funnel touches. Time-decay distributes credit across all touches, weighting recent ones more via an exponential decay with a chosen half-life. This demands the full, correctly ordered path per user, robust identity stitching since a missed early touch silently distorts weights, a defined lookback window, and a tunable decay parameter, making it far more sensitive to data quality and more expensive to compute.
COMMON WRONG ANSWERS Skipping identity resolution and treating each channel's conversions independently, which double counts. Ignoring cross-device stitching. Assuming time-decay is just last-touch with extra steps rather than a full-path computation. No lookback window.
ONE CONCRETE EXAMPLE A path is a Google ad on day 1, an email on day 5, Facebook on day 9, and purchase on day 10. Last-touch gives Facebook all credit. Time-decay with a 7-day half-life gives Facebook the most, email a meaningful share, and the day-1 ad a small slice, rewarding the assisting touches the last-touch model erased entirely.
Read the original → improvado.io
Get five bites like this every day.
Tezvyn delivers a daily feed of 60-second tech bites with quizzes to lock in what you learn.