How Attribution Turns a Measurement Gap into a Budget Error
This chapter diagnoses a specific failure in media measurement: a demand-generating platform receives less budget and attributed revenue after some of the demand it created is completed on a trusted marketplace. The question is how a lower observability factor becomes a lower spend decision, and under what additional conditions effectiveness can appear to backfire. The useful judgment is to separate the formal under-investment result from the more conditional claim that better advertising can reduce the generator's measured revenue.
The symptom is a smaller budget for the demand generator
Start with a brand that divides its advertising budget between the generator, G, and the trusted marketplace, R. The generator's advertising creates intent, but the previous chapter's observability factor, κ, means that only a fraction of the resulting conversions is visible as G's performance. The brand now applies a familiar management rule: keep investing in a channel until its measured marginal return falls to a hurdle, λ. The rule may be sensible when the measurement boundary is complete. It becomes distorted when the numerator systematically omits outcomes.
The distortion is not simply that the brand writes down a lower number. The lower number changes the point at which the brand stops buying exposure. With true information, the model's first-best generator spend is proportional to:
(mαβ / λ)¹⁄⁽¹⁻ᵝ⁾
Here, m is unit margin, α describes the generator's effectiveness, β describes diminishing returns, and λ is the brand's return hurdle. When the brand uses attributed rather than true incremental returns, the same expression contains κ:
(mκαβ / λ)¹⁄⁽¹⁻ᵝ⁾
The exact notation is less important than the relationship between the two choices. Attributed generator spend divided by true-information spend is:
κ¹⁄⁽¹⁻ᵝ⁾
If distrust diverts any purchases away from G and off-platform recovery is incomplete, κ is below one. Because the exponent is greater than one when β is between zero and one, the spend ratio can fall faster than the observability factor itself. Diminishing returns do not cause the measurement problem, but they shape how severely the problem is translated into allocation.
For intuition, imagine an observability factor of one-half and a response curve with β = 0.5. The spend ratio becomes (0.5)², or one-quarter. This is not a forecast for a particular advertiser; it is a direct implication of the model's threshold rule. It shows why “the dashboard is off by half” and “the budget is off by half” are not equivalent statements. The allocator is solving for a spend level on a curved response function, so a change in the measured return changes the point at which the marginal return reaches the hurdle. The curvature turns a measurement wedge into an allocation wedge.
The result also depends on which return is being compared. A channel can display a high average revenue per dollar because it captures customers who were already close to purchase, while its marginal contribution to new demand is smaller. The generator faces the reverse visibility problem: its marginal contribution can be real but recorded later or elsewhere. The threshold rule is not inherently wrong for comparing marginal returns; it is wrong when one channel's marginal outcome is systematically filtered before the comparison is made.
This is the first diagnosis: the generator is underfunded because the budget rule treats a filtered outcome as if it were the complete incremental outcome. The result follows from the stated response curve and threshold rule. It does not require the generator's advertising to be ineffective, and it does not require the marketplace to be more persuasive. It requires only that the generator's caused outcomes be less observable than the returns used to compare channels.
A simple way to see the feedback is to hold the true customer response fixed. Suppose the generator creates a given amount of incremental intent. If more customers choose R's checkout, the number of purchases caused by the campaign need not change. But the number credited to G falls unless the marketplace returns enough signal to restore the connection. The brand then observes a weaker generator numerator, lowers the generator's measured marginal ROAS, and reduces the spend that produces the original intent. The next period contains less generator activity not because a new experiment proved that the advertising stopped working, but because the previous period's measurement was incomplete.
That distinction matters for diagnosis. A low attributed return is a symptom, not a root cause. It could reflect weak creative, poor targeting, saturation, genuinely low incremental demand, or a high share of purchases that completes elsewhere. The attribution report alone does not distinguish those explanations when off-platform outcomes are missing. The formal model isolates one mechanism that can produce the symptom; it does not establish that every weak channel report has that cause.
The channel comparison can also become nonlinear. The brand is not choosing between “demand” and “no demand.” It is choosing how much to invest in a source of demand whose returns are partly hidden and a surface that may be good at recording or harvesting that demand. When the hidden share is large, the budget rule can favor the channel with the clearest records over the channel with the strongest causal contribution.
The loop rewards the harvester
The trusted marketplace can benefit from the generator's measurement problem even when its own advertising did not create the diverted intent. Suppose R sells sponsored placements and receives a share of the customers who arrive with intent created by G. If the marketplace claims some of those diverted purchases through its own last-touch records, its measured conversions increase by a share of τq(sG). The marketplace's apparent ROAS then rises, while the generator's apparent ROAS remains reduced by the missing or partially recovered conversions.
This is a reallocation externality. G pays to create the demand, but R's measurement surface can receive the evidence of the resulting sale. Under a budget rule that compares the two reported returns, G helps finance the appearance that R is the stronger channel. The marketplace may still provide genuine value through trust, payment, shipping, or fulfillment. The point is narrower: recording or facilitating a transaction does not by itself identify who created the incremental intent, and a sponsored placement near the sale does not settle that counterfactual.
The error is especially difficult to spot because the marketplace's record is not necessarily fabricated either. R really did receive the order, and its sponsored placement may have been present at the final touch. What the record cannot establish on its own is whether the customer would have arrived with the same purchase intent without G's earlier intervention. In the model, R's spend responds to arriving intent, so a regression or dashboard that treats the marketplace's observed sales as independent evidence of its own demand generation can attribute the upstream effect to the downstream surface.
That is the difference between harvesting and creating. Harvesting is valuable when a channel reduces friction or makes a willing customer more likely to complete the transaction. Creating demand means changing the number of purchases relative to the no-ad counterfactual. A single last-touch record can document the former interaction while remaining silent about the latter comparison. The brand's allocation rule needs the second question when it is deciding whether to fund more reach, not merely whether a checkout occurred after a sponsored placement.
The loop can be represented as a sequence of linked conditions:
The diagram includes a condition that must not be smuggled into the general under-investment result. More effective advertising does not automatically reduce the generator's attributed revenue. If α rises while trust and recovery stay fixed, true demand rises and the same κ scales a larger response. The attribution wedge remains, but the direction of measured revenue need not reverse.
The stronger “backfire” result requires a relationship between effectiveness and trust. The source considers a case in which sharper creative or better targeting reaches marginal buyers who are less willing to use G's checkout. In that case, τ rises with α. Better advertising has two effects at once: it creates more intent, which tends to raise the generator's returns, and it sends a larger fraction of that intent to R, which lowers the share that G can observe. If the second effect outweighs the first in the measured numerator, attributed ROAS, budget, and revenue can all decline as true effectiveness improves.
This is why the assisted-own-goal phrase must be used carefully. The ad is not scoring against itself in the causal market. The ad is scoring against itself in the measurement-and-allocation system. Its causal success produces transactions that the system records on a competing surface; the resulting record shifts future resources away from the source of demand. “Own goal” describes the feedback created by crediting and budget rules, not a claim that the campaign reduced the number of people who wanted the product.
The simulation makes that conditional mechanism visible. It sets the diminishing-returns parameter at β = 0.5, the recovery rate at φ = 0.1, and lets the diversion share rise with effectiveness according to τ(α) = 0.5 + 0.25(α − 1). Under those assumptions, first-best spend increases throughout the effectiveness range. Attributed generator revenue instead peaks at approximately α = 1.7 and then declines. By α = 2.6, the generator has lost nearly half of its peak revenue even though it is more than twice as effective as at baseline. The region to the right of the peak is the simulated own-goal region: a marginal improvement in effectiveness reduces the measured return that governs allocation.
Those numbers are evidence about a specified feedback mechanism, not a universal forecast for advertising markets. The result depends on the assumed relationship between α and τ, the recovery rate, the threshold allocation rule, and the model's single-generator, single-marketplace structure. The simulation does not estimate how trust changes for a commercial platform, nor does it prove that the marketplace's sponsored placements cause no genuine conversions outside the constructed setting. It shows what can happen when a causal response and an observability response move in opposite directions.
That qualification also suggests how the failure should be verified. A brand should not infer the root cause from the generator's attributed revenue alone. It would need evidence about where conversions are booked, how often off-platform outcomes can be linked, whether the diverted population differs in trust or other characteristics, and whether a measurement change alters the apparent return without changing total channel-complete outcomes. A randomized comparison with outcomes counted across checkout destinations can test the causal response. A change in crediting logic alone cannot, because it still operates on the same visible universe.
The distinction between symptom and cause prevents two opposite mistakes. The first is to accept low ROAS as proof that the generator does not create value. The second is to declare every low ROAS result an assisted own goal. The model supports a narrower conclusion: whenever trust diverts purchases and recovery is incomplete, attribution systematically lowers the generator's measured return; whenever that measured return controls spend, under-investment follows under the stated rule; and only when diversion rises enough with effectiveness can better performance visibly backfire.
The deeper issue is now larger than one platform's dashboard. Attribution is one way of observing a commercial journey, while the budget rule is a decision made from that observation. Different measurement regimes preserve different parts of the journey: some preserve visible touchpoints, some preserve aggregate sales, and some preserve a randomized comparison of total outcomes. The next task is to compare those regimes on the same question rather than asking a single metric to carry all three meanings.