Every Model Has a Bias
Last-click attribution gives all the credit to whatever touchpoint happened right before a conversion, which tends to over-reward channels like branded search that catch people who were already decided. First-click does the opposite, over-crediting the very first exposure and ignoring everything that closed the deal. Multi-touch models split the difference, but the split is still a modeling choice, not a measured fact.
Incrementality Testing as a Reality Check
Incrementality testing — deliberately holding a channel back from a segment and comparing results — gets closer to a real answer for a specific question, but it's slower, more disruptive, and impractical to run on everything at once. It's better thought of as a periodic check on your model's assumptions than a replacement for day-to-day attribution reporting.
Consistency Matters More Than Precision
In practice, the most useful approach is picking one attribution model, understanding clearly what it tends to over- or under-credit, and staying consistent with it so that trends over time remain comparable — even if the absolute numbers from any single model wouldn't survive close scrutiny on their own.