Name the question first
“Which channel gets credit?” is too vague. A lifecycle team deciding where to place nurture resources needs a different view from a finance team forecasting incremental revenue. Write the decision, time horizon, and acceptable uncertainty before opening an attribution report.
For short-cycle acquisition, a last-touch view can be operationally useful because it highlights the channel that completed demand. For longer consideration cycles, it can systematically hide discovery and education. No setting removes that tradeoff.
Know what each common model emphasizes
First-touch attribution emphasizes discovery. Last-touch emphasizes conversion completion. Linear models share credit evenly, while time-decay models favor recent interactions. Position-based models privilege the first and last known touches. Data-driven models use observed conversion patterns to assign weights, but their sophistication does not make missing data visible.
A useful choice is one whose bias matches the decision. If you are evaluating how demand enters the system, first touch may be informative. If you are improving a handoff close to conversion, last touch may be clearer. When leadership needs one reporting convention, state its bias next to the result.
A more complex model can be more precise about the data you captured and still be wrong about the world you did not observe.
Account for the invisible journey
Cookie limits, consent choices, cross-device behavior, offline conversations, walled gardens, and unattributed direct visits all create gaps. Those gaps are not random. Privacy-aware visitors, enterprise buyers, and long research journeys may be underrepresented in different ways.
Document the coverage boundary: which platforms, devices, regions, consent states, and offline outcomes are included. Track changes in the unattributed share. A sudden improvement can mean better instrumentation, but it can also mean a classification rule started absorbing traffic that used to be honest about being unknown.
Use comparison to reveal sensitivity
Choose a primary model for consistent reporting, then compare it with a deliberately different model. If paid social looks weak in last touch but strong in first touch, the disagreement is the finding. It suggests an assist or discovery role that deserves a more targeted test.
Look at channel conclusions, not tiny credit differences. Would the budget decision change under a reasonable alternative model? If yes, confidence should fall and the next step should be evidence gathering—not averaging two incompatible answers.
Move from attribution to incrementality
Attribution describes associations inside observed journeys. Incrementality asks what would have happened without the marketing activity. Geography tests, holdouts, matched markets, and carefully designed lift studies can answer that stronger question when the decision is large enough to justify the effort.
Use attribution for ongoing navigation and experiments for expensive turns. A practical reporting note can state the attributed result, the model and lookback window, known coverage gaps, and the strongest causal evidence available. That is more honest—and more useful—than a single number with false precision.