The Mechanism: Attribution Is Inference Wearing the Costume of Measurement
Attribution feels like measurement because it produces numbers with decimals, but it is inference about counterfactuals — what a buyer would have done had a given touch not occurred. That counterfactual is never observed. The customer who saw the retargeting ad and bought cannot also be the customer who did not see it and tells you whether they would have bought anyway. So every attribution model — last-touch, first-touch, time-decay, data-driven — is a rule for distributing credit across a sequence, not a reading of cause. The rules disagree violently on the same data, and that disagreement is the tell. When last-touch and a multi-touch model assign the same channel 40% and 18% of credit, no measurement error explains the gap. Two assumption sets do.
The deeper failure is that attribution systems see only the paths that converted. The denominator — the people exposed who never returned — is structurally invisible to the conversion log, yet it carries the information needed to estimate lift. A channel that reaches buyers who would have arrived regardless will always look efficient, because attribution rewards proximity to the purchase, not influence over it. Branded search is the canonical offender: it intercepts demand that already exists and bills you for the interception.
The Trade-off: Precision Versus Identifiability
Strategists reach for more granular models hoping precision will dissolve the noise. It does the opposite. The more parameters a model uses to slice credit across channels, creatives, and time windows, the more it overfits the idiosyncrasies of the conversion paths it happened to see, and the less it tells you about the world outside that sample. Granularity buys apparent confidence at the cost of identifiability — the ability to separate the effect of spend from the demand that would have existed anyway. You can have a model that explains last quarter beautifully or a model whose conclusions survive into next quarter. Rarely both.
This is why the honest unit of marketing knowledge is incremental lift, not attributed credit, and why lift can only be earned through deliberate variation: holdouts, geo experiments, staggered launches, spend that is deliberately withheld so the absence can be observed. That variation is expensive and politically uncomfortable — it means turning off channels that look like they work to find out whether they do.
The Failure Mode: Optimizing the Map Until the Territory Erodes
The characteristic disaster is not a single bad call; it is a slow, compounding drift. A team optimizes toward the channels their attribution model favors. Those channels are disproportionately the demand-harvesting ones, because harvesting sits closest to conversion and wins the credit fight. Budget migrates from demand creation — the brand, category, and awareness work whose effects are lagged, diffuse, and uncapturable by a click log — toward demand capture. The metrics improve. Cost-per-acquisition falls. And then, six to twelve months later, the top of the funnel quietly thins, because nothing was feeding it, and the harvesting machinery now has less to harvest. The dashboard reported success the entire way down.
The asymmetry that makes this trap so durable: the costs of over-investing in capture are deferred and hard to trace, while the rewards show up instantly in exactly the report the team is judged on. A rational operator, optimizing the signal they are given, will walk straight into it.
The Decision Implication: Govern by Portfolio and Test, Not by Model
The board's job is not to find the true attribution model. There isn't one. The job is to build a spending discipline that stays sane while the signal stays noisy. That means treating attribution as a directional prior to be checked against reality, never as the arbiter, and reserving the verdict on whether a channel works for experiments that can actually deliver it.
- Set the top-line split — creation versus capture — as a strategic decision held above the optimizers, so daily efficiency metrics cannot quietly cannibalize the brand budget that feeds them.
- Hold a standing experimentation budget for holdouts and geo tests; trust lift estimates over attributed credit wherever the two conflict, and demand a causal claim before reallocating at scale.
- Validate the whole portfolio against one aggregate truth — total media spend against total margin — because the parts can each look efficient while the sum quietly loses money.
Decide spend the way you would underwrite under uncertainty: size bets to what you can actually know, keep deliberate variation in the system so the signal can improve, and never confuse a confident number with a true one.