Buyer Journey Insights with Data-Driven Attribution

John Bencina, Erkut Aykutlug, Changshuai W., Stephanie Sorenson, Shao Tang22 Jan 2025

Problem

Marketers need to know which campaigns drive conversions in order to move budget. Last-click attribution gives all credit to the final touch, so it over-credits Search and Email and under-credits upper and mid-funnel channels (video, display, social). For paid media, privacy limits mean no user-level impressions exist, only proxy clicks and daily aggregates, which biases any path model. A path model and a marketing mix model (MMM) also disagree on channel totals unless forced to match.

Example from the article: LinkedIn's internal marketing team used last-click attribution, which showed flat credit for video, display and social campaigns even as spend on them grew.

Solution in brief

Train an attention model that predicts conversion from a member's touchpoint sequence plus member and company features, and read normalized attention weights as credit. Impute missing paid impressions from daily aggregates, then rescale credit per channel and quarter to match the MMM. Estimate channel lift by deleting a channel's touches from a path and comparing predicted conversion probabilities.

Questions this case answers
  • How to allocate credit for an outcome across a sequence of exposures when one class of exposures is observed only in aggregate
  • How to reconcile a model built on individual-level sequences with a model built on aggregated totals so that their estimates agree
  • How to split conversion credit across marketing touchpoints when last-click over-credits search and email?
  • How to attribute conversions to paid channels when impressions are not observed per user?

Tools and methods

  • Attention-based multi-touch attribution model — binary classifier of conversion; attention weights become credit
  • Paid media impression imputation — spreads aggregate impressions over members and paths probabilistically
  • Post-modeling calibration to MMM — rescales channel credit to MMM totals
  • Counterfactual path lift — compares conversion probability with and without a channel's touches
  • Apache Spark — daily ETL that builds marketing paths
Continue reading — it's free

Google or a one-time email code. No password.