Multi-touch Attribution

Multi-touch attribution begins by building a path table: one row per recorded interaction, keyed to a person or account, ordered by timestamp and ending at the conversion row. Rules decide what qualifies as a touch, since a click always counts, an ad impression sometimes does, and an email open usually does not, and further rules decide whether consecutive touches from the same channel collapse into one. The chosen weighting then splits a single conversion into fractions across those rows, so channel reports contain decimals rather than whole conversions.

Two properties determine whether the output is usable: path completeness and path length. Where identity resolution is weak, one buyer becomes three anonymous visitors with three short paths, and the model over-credits closing channels because the earlier touches sit in unrelated rows. Loosening the definition of a touch lengthens paths and dilutes every channel, while tightening it produces cleaner but shorter journeys. Deduplication is the underrated lever: a retargeting campaign firing eleven impressions can dominate a path unless repeats are collapsed.

In practice teams run multi-touch on a warehouse table rather than inside a reporting interface, because they need to define touch rules themselves and join marketing events to CRM outcomes such as opportunities and closed deals. Fractional credit is then rolled up to channel level and compared against spend. Where a scorecard quiz sits mid-funnel, this is the only family of models that shows the ad which produced the visit and the quiz which qualified the person both sharing one closed deal.

Multi-touch cannot invent touches it never saw, and it cannot rank two channels that always appear together. If every converting path contains both paid search and email, no weighting rule separates their contributions; the split merely reflects the weights you chose. Fractional conversions are also awkward for teams used to whole numbers, and rounding across many channels makes small ones look busier than they are. At very low conversion volume the whole exercise becomes unstable from week to week.

Start with path diagnostics rather than credit: median touches per converting path, share of paths containing only one touch, and share of conversions that could not be joined to any prior touch. A rising share of single-touch paths means the stitching is degrading, and the model will drift toward last-touch behaviour without anyone changing a setting.

Then compare fractional credit against spend per channel to get a modelled cost per conversion, and watch how that figure moves when you change the weighting rule. If a channel keeps its ranking under linear, time-decay and position-based weighting, the finding is robust. If the ranking flips between models, treat the question as unresolved and settle it with a holdout test.

The first failure is counting ad impressions as touches without deduplicating them. A single display campaign then appears in nearly every path and quietly absorbs credit from channels that produced actual clicks. Decide explicitly which event types qualify, collapse repeated touches from the same channel inside a short window, and publish those rules beside the report. A model whose touch definition is undocumented cannot be audited or defended a quarter later.

The second is presenting fractional conversions to stakeholders who expect whole ones. Nine channels each credited with a decimal add up correctly but leave everyone unsure what a single deal actually looked like. Show two views: the fractional roll-up for budget allocation, and a small sample of real paths listed in order. Concrete journeys settle the disagreements that decimal tables reopen at every monthly review.