A step-by-step QA pass for your conversion path data

A step-by-step QA pass for your conversion path data

The question: every multi-touch model — Markov, Shapley, position-based — is only as good as the conversion paths feeding it. How do you verify those paths are real before you model on them?

The QA sequence:

— Pull a sample of 50 full paths and read them like a story. Do the sequences make sense, or do you see impossible jumps and duplicated touches?
— Check path length distribution. If 80%+ are single-touch, your tracking is collapsing journeys, not capturing them — and any multi-touch model will mostly reproduce last-click.
— Measure identity stitching. Cross-device and logged-out gaps fragment one journey into several; quantify your match rate before trusting path completeness.
— Inspect time gaps between touches. Suspiciously clustered timestamps often signal bot traffic or tag double-firing.
— Verify the first touch isn't systematically (direct). Stripped referrers amputate the top of the funnel, starving exactly the channels MTA is supposed to rescue.

The nuance: incomplete paths don't fail loudly — they bias quietly toward whatever is best-tracked, usually bottom-funnel. The model still returns clean-looking numbers. That false confidence is the real risk.

Bottom line for practitioners: read raw paths, profile their length, and measure your stitching rate before fitting any attribution algorithm. A Shapley value computed on broken paths is a precise answer to the wrong question.