The Attribution Model You Chose Is a Theory About Causation You Never Tested

Attribution is almost universally treated as a measurement problem. Pick a model, configure your windows, connect your ad server to your CRM, and read the output. The debate in most media buying teams is limited to which model — last touch, first touch, linear, time decay, data-driven — best represents the customer journey. That debate is real and worth having. But it starts one step too late.

Before you select an attribution model, you are selecting a theory of causation. You are encoding a specific assumption about which touchpoints cause purchase decisions, how long that causal chain runs, and how credit should be distributed across it. The model does not discover that theory from your data. It enforces it. And because every model produces a number, and that number looks like evidence, the underlying assumption is almost never audited after deployment.

What Each Model Is Actually Claiming

Last-touch attribution claims that the final measured interaction before conversion is the decisive one. This is not a neutral reporting convenience — it is a falsifiable hypothesis about buyer psychology: that all prior exposure is irrelevant to the conversion outcome except as context for the last touchpoint. If that hypothesis is wrong for your category, your product, or your buying cycle, last-touch attribution will systematically misdirect budget toward closure channels while starving the awareness and consideration investments that actually moved buyers into the funnel.

Time-decay attribution claims that causal influence on a purchase decision is a monotonically decreasing function of time. Touchpoints closer to conversion matter more. This is intuitive for impulse categories with short consideration cycles. For enterprise B2B purchases with 90-day or 180-day sales cycles and multi-stakeholder decisions, it may have no empirical basis at all. A whitepaper download eight weeks before a demo request may be causally upstream of everything that follows — but time-decay weights it near zero.

Data-driven attribution sounds like an escape from these assumptions, but it is not. It replaces hardcoded weights with statistically estimated ones, but it is still optimizing for correlation between touchpoints and conversion events, not for causal contribution. If your measurement environment has selection effects — and it does — data-driven attribution will find patterns in those selection effects and report them as causal signal.

The Validation Step That Is Almost Never Taken

The standard for validating a causal claim is an experiment: expose some buyers to a touchpoint, withhold it from a matched group, and measure the difference in conversion rate. Most attribution models are deployed without anything approaching this. A model is chosen based on vendor defaults, industry convention, or internal political compromise between channel owners, and then run continuously against live campaigns. The model's outputs are used to make budget decisions. Those decisions reshape which channels scale and which shrink. The mix that results is optimized for what the model rewards, not for what actually drives revenue.

This creates a feedback loop that is invisible in standard reporting. If your attribution model overweights paid search, paid search scales. As it scales, it captures more last touches on conversions that other channels influenced upstream. This increases its measured attribution share further. Your budget allocation drifts toward the channel your model rewards, and your model's outputs look increasingly validated — because the channel it rewards is now touching more conversions. The model is not measuring reality; it is manufacturing the conditions that confirm itself.

Where the Causal Assumption Can Be Tested

The practical question is not whether attribution models encode assumptions — they all do — but whether those assumptions can be tested without a full experimental infrastructure rebuild.

Three checks are tractable for most media buying teams.

First, run channel-off tests for defined periods. If a channel is receiving substantial attribution credit, suppressing spend in that channel for a controlled window and measuring conversion volume against a comparable prior period gives you a rough signal of whether the attributed contribution is real. This is not a clean experiment, but it produces directional evidence that is better than assumption.

Second, compare your attribution model's output to your CRM's pipeline sourcing data. If attribution is crediting digital display with significant influence but your CRM records show that pipeline is sourced primarily through outbound and event, the models disagree about basic facts. That disagreement is worth investigating before you use either number to set next quarter's budget.

Third, track conversion rates across buyer cohorts segmented by touchpoint history, not by touchpoint count or recency. If buyers who encountered certain content types or channel sequences convert at materially higher rates regardless of the final touchpoint, your model's causal assumption about recency and credit distribution is probably wrong for your category.

The Specific Cost of Running an Unvalidated Model

The cost is not random noise in your measurement. It is systematic bias in a specific direction determined by the model's architecture. Last-touch models systematically undervalue top-of-funnel investment. Time-decay models systematically undervalue early-stage content in long-cycle categories. Linear models systematically overvalue touchpoints that are incidental to the conversion path. Each bias produces a different flavor of misallocation, but all of them compound over time as budget follows the model's signals.

For media buyers managing significant programmatic spend, the dollar magnitude of this misallocation is not marginal. If attribution is systematically overweighting a channel by even 15 percentage points of budget share, across a seven-figure annual plan that is a material misdirection — one that no optimization work inside the overweighted channel can correct, because the constraint is not channel performance, it is model architecture.

What Sound Practice Looks Like

The goal is not to find the attribution model that is true — none of them are — but to operate with explicit awareness of each model's causal assumptions and to run periodic evidence checks on whether those assumptions hold in your specific buying environment.

Document the assumption your current model encodes. State it as a falsifiable hypothesis. Design at least one observable check that would provide evidence against it. Review that evidence on the same cadence you review campaign performance. If the evidence contradicts the assumption, change the model or adjust how you weight its outputs in budget decisions.

Attribution will remain an approximation. But an approximation with a known error structure that is periodically tested is a fundamentally different decision input than an approximation whose error structure is invisible and assumed to be zero.