Attributed Credit and Incremental Effect: What Multichannel Attribution Models Can Establish

Attribution is the mechanism by which most marketing budget is defended, and it is the revenue analysis most often asked to carry more weight than its construction supports.

The problem is not that the models are poor. Modern multichannel attribution is technically sophisticated. The problem is a category error committed at the point of use: an attribution model distributes credit for conversions that occurred, and it is read as an estimate of what each channel caused. Those are different quantities, and the difference is not a nuance.

1. What the research established, and how it was established

Li and Kannan [1] developed an empirical model of multichannel attribution and — this is the part that matters — validated it against a field experiment. The pairing is the point. A model estimated from observational touchpoint data was tested against an intervention, because observational credit assignment does not by itself establish effect.

The gap between them is selection. Channels that appear before conversions are frequently channels that people already intending to convert use. A branded search click before a purchase is at least partly an effect of the intention rather than a cause of it, and no credit rule applied to observational data can separate the two.

2. The four choices inside every attribution model

Any attribution output is the product of four decisions. They are rarely documented, which is why results appear to change on their own.

The third deserves emphasis. For a fixed dataset, the choice of credit rule can reorder channel rankings entirely. A comparison of channels under an undocumented rule is not a finding about channels.

3. Why the error persists

Two reinforcing reasons, one cognitive and one organisational.

The cognitive one is availability [3]: tracked channels are the ones present in the data, so they are the ones considered. Channels that cannot be tracked — a conversation, a community, a recommendation — are not merely under-measured, they are absent, and absence reads as zero rather than as unknown.

The organisational one is that attribution output is used to allocate budget, which makes it a performance measure. Ridgway's finding [4] applies directly: once credit becomes the measure, channel behaviour optimises for capturing credit — retargeting placed late in the journey, campaigns designed around trackable touches — which improves the measure without improving the outcome.

4. What attribution is legitimately good for

The critique is of one use, not of the technique. Attribution models support several things well.

• Describing the touchpoint structure of the journey — how many interactions, over what elapsed time, in what sequence. This is descriptive and does not require a causal claim.

• Detecting change over time under a fixed rule. If the rule is held constant and documented, movement in attributed credit is a real signal about something.

• Segment comparison. Enterprise and self-serve journeys differ structurally, and attribution describes that difference usefully.

• Generating hypotheses for experiments. A channel with rising attributed credit is a reasonable candidate to test — a use that treats the model as a pointer rather than as an answer.

5. Protocol: establishing effect where the decision requires it

When the decision is material — a substantial budget shift, a channel exit — attribution is insufficient and an incrementality test is the appropriate instrument.

1. State the decision and the threshold first. What magnitude of effect would change the decision? A test that cannot detect that magnitude is not worth running, and this is the step most often skipped.

2. Choose a unit of randomisation you can actually control — geography, account list, time period. This constraint usually determines the design.

3. Hold out. Suppress the channel for the control group for a period exceeding the typical sales cycle, or the test measures timing rather than effect.

4. Measure the outcome that matters, not the proximate one. Held-out regions producing fewer form fills but equal closed revenue is a finding, and it is invisible if the measured outcome is the form fill.

5. Compare against the attributed figure. The gap between attributed credit and measured incremental effect is the most useful number the exercise produces, and it calibrates how much to trust attribution for the other channels.

6. The valuation frame

Gupta, Lehmann and Stuart [2] demonstrated that customer-level value can be estimated with sufficient rigour to inform firm valuation. That work is relevant here as a contrast in standard: it is explicit about its model, its assumptions and its sensitivity, which is what allows the estimate to be used for a consequential decision.

The comparison is not flattering to routine attribution practice, where the model is a tool default, the assumptions are undocumented, and the sensitivity is never tested — and the output is nonetheless used to move budget.

7. Limits of this argument

Li and Kannan [1] establish a validated modelling approach in an online multichannel setting, not a general claim that all attribution is misleading. Gupta et al. [2] establish rigorous customer valuation, not a claim about attribution specifically. The heuristics literature [3] and the measurement literature [4] are general findings applied here by argument.

What is defended: attributed credit and incremental effect are distinct quantities, the distinction is present in the research that developed these methods, and the four embedded choices materially determine the output. The recommendation follows from the distinction and not from any claim about how wrong any particular model is.