What is Multi-touch attribution?
Multi-touch attribution divides a conversion's credit across every source that contributed rather than giving it all to one. Common rules are linear (equal shares), time-decay (later touches weigh more) and position-based (usually 40% first, 40% last, 20% spread between). Every rule is a chosen assumption, not a discovered fact.
The common models
| Model | How credit splits | Suits |
|---|---|---|
| Linear | Equal share to every touch | Long journeys where you cannot justify weighting |
| Time-decay | Exponentially more to recent touches | Short cycles, promotional selling |
| Position-based | 40% first, 40% last, 20% divided among the middle | Journeys where discovery and closing both matter |
| Data-driven | Weights derived statistically from your own conversion paths | Large volumes, and only with real data-science support |
The honest framing
None of these models measures causation. A visitor’s journey is a list of things that happened before a purchase; the model is a rule for dividing credit among them, chosen in advance by a person. Run the same sales through linear and time-decay and channels will change rank — not because anything about the business changed, but because you changed the assumption.
That is not a reason to avoid multi-touch. It is a reason to be explicit: state the model on the report, keep it fixed long enough for trends to mean something, and treat a change of model as a change of instrument rather than a discovery.
Practical guidance
- Start with two single-touch models, not a complex one. First-click and last-click side by side reveal most of what multi-touch would tell you, and neither hides its assumption inside a weighting formula.
- Linear is the least arguable multi-touch rule. If you cannot defend why one position should weigh more, equal shares is the honest default.
- Data-driven models need volume. Below a few thousand conversions a month they are fitting noise, however sophisticated the description sounds.
- Every model needs cross-visit memory. The longer the journey you want to model, the more visitor history the tool must retain — which is a real privacy cost, and worth weighing against the precision you actually gain.