The Journey a Dashboard Cannot See
This chapter examines the customer journey behind an attributed conversion: how a demand-generating platform creates purchase intent, how trust determines the checkout surface, and how the reporting system recovers only some of the resulting signal. The useful judgment is not whether a marketplace sale is “really” associated with an earlier ad in every case. It is whether a metric can distinguish a change in where a purchase was completed from a change in how many purchases the advertising caused. The distinction becomes precise once demand, channel choice, and observability are represented separately.
Intent is created before a checkout is chosen
Consider a brand that sells one product and buys advertising on a platform that both reaches customers and operates a relatively new storefront. Call that platform the generator, G. The brand can also sell through a large, familiar marketplace, R. The two surfaces are connected by the customer, but they are not interchangeable from the customer's point of view. The generator may be where the product is discovered; the marketplace may be where the customer feels comfortable paying.
The model starts with the generator's advertising spend, sG. That spend produces expected incremental purchase intent according to a response function:
q(sG) = αsG^β
Here, q(sG) is the amount of intent created by the advertising, α represents the platform's effectiveness or scale, and β lies between zero and one, representing diminishing returns. The function says that more spend can create more intent, but each additional unit of spend produces less additional intent than the previous one. It is a compact way to describe a familiar commercial reality: reaching more people or buying more exposure does not make demand grow without limit at the same rate.
The important point is what q measures. It measures the purchase intent generated by the advertising, not the number of transactions that will later appear in G's storefront records. Once intent exists, the customer still has to choose where to complete the purchase. The generator's advertising and the generator's checkout are therefore separate stages.
The shape of the response curve matters because it gives the brand a reason to compare marginal effects rather than only totals. With β below one, the slope of q(sG) falls as spend rises: an additional unit of budget can still create intent, but it creates less of it at the margin. The parameter α shifts the overall effectiveness of the generator; the parameter β describes how quickly extra spend encounters diminishing returns. Neither parameter says where the eventual order will be booked. Two campaigns can create the same amount of intent and produce different dashboard results if their customers have different checkout preferences or if one journey is easier to match back to the original ad.
That separation is also a counterfactual discipline. To say that G generated incremental demand means comparing the observed outcome with what would have happened without the advertising opportunity. To say that R booked the order describes the observed transaction in the treated journey. The second fact does not supply the missing no-ad comparison, and the first fact does not imply that G should own the checkout record. A metric that mixes the two stages can look precise while answering neither question cleanly.
The model represents that choice with a distrust parameter, τ. A higher τ means more customers are reluctant to use the generator's on-platform checkout because of unfamiliarity, payment friction, weaker returns, less buyer protection, or general reputation. For a customer whose intent was generated by G, the probability of completing the purchase on R is τ, while the probability of completing it on G is 1 − τ.
That parameter changes the location of the transaction, not the amount of intent the advertisement created. If q(sG) units of incremental intent have been generated, the expected total number of incremental purchases remains q(sG), regardless of whether those purchases occur on G or R. The split is:
- On G: (1 − τ)q(sG) purchases are booked on the generator's storefront.
- On R: τq(sG) purchases are booked on the trusted marketplace.
This is the mechanism's first necessary separation. A customer can be persuaded on one surface and reassured on another. The second surface may remove the final obstacle to purchase, but that does not establish that it created the original intent. Conversely, the fact that the generator caused intent does not mean every resulting sale will be visible to the generator.
The distinction matters for a brand because the same campaign can generate two kinds of records. The generator can observe the people who click, browse, or buy on its own surface. The marketplace can observe the orders completed in its checkout. The brand may hold a broader transaction record, but its advertising dashboard can connect that record to the original exposure only when the required signal crosses the boundary between platforms.
Trust is not a decorative psychological variable in this model. It is the mechanism that changes which records are available to which observer. If trust in G's checkout is high, most intent becomes on-platform transactions and the generator's records provide a reasonably complete view of its results. If distrust is high, the generator can remain effective at creating demand while a larger share of the evidence is booked somewhere else.
The effect is easiest to miss when people treat the checkout location as a proxy for the entire journey. A customer might discover the product in G's feed, compare alternatives, remember the brand later, and then choose R because the final payment and returns experience are familiar. The marketplace is part of the path to purchase, but its presence does not reveal how much of the customer's intent was already created upstream. Nor does the generator's inability to observe the order prove that no influence occurred. The model deliberately keeps these questions separate so that a change in trust changes the reporting mix without silently changing the amount of demand generated.
This also clarifies what the brand should ask when two dashboards disagree. It should not begin by asking which platform has the more persuasive story. It should ask whether the dashboards cover the same population, checkout destinations, conversion window, and matching rules. If G counts only purchases it can connect to its own surface and R counts orders near its sponsored placement, the two numerators describe different subsets of the journey. A difference between them may be a difference in observability before it is a difference in effectiveness.
The sequence is easier to see when the commercial event and the measurement event are shown together:
The left side of the sequence—spend creating intent—is a causal story. The right side—where the purchase is recorded and whether the signal is returned—is an observability story. They touch one another, but they are not the same variable.
The observability factor
To connect those stories, let φ be the recovery rate for purchases completed on R: the probability that such a purchase is nevertheless matched back to the original G impression. The model assumes that purchases completed on G are matched perfectly. It does not assume that marketplace purchases are always lost; some may be recoverable through identifiers or other signals. The key is that the recovery rate can be less than one.
The number of conversions attributed to G is therefore the sum of two parts:
q̂G(sG) = (1 − τ)q(sG) + φτq(sG)
The first term is the on-platform share, which is observed. The second is the off-platform share that survives signal recovery. Factoring out total intent gives an observability factor:
κ(τ, φ) = (1 − τ) + φτ = 1 − τ(1 − φ)
Attributed conversions are κq(sG), while true incremental purchases remain q(sG). When there is no distrust, τ = 0, every purchase occurs on G and κ equals one. When every marketplace purchase can be sent back and matched, φ = 1, κ also equals one. In the intermediate case—some distrust and incomplete recovery—κ is less than one. The comparative statics follow directly. Holding recovery fixed, an increase in τ moves more of the same generated intent into the off-platform share, so κ falls. Holding distrust fixed, an increase in φ returns more of that share to the measurable numerator, so κ rises. The model therefore predicts a larger attribution wedge when checkout trust is lower or cross-platform recovery is weaker; it does not need to assume that the ad became less persuasive.
The formula is useful because it shows exactly what the apparent performance loss means. Suppose the journey creates one hundred incremental purchase intents. If distrust sends a share to R and only part of that share is recovered, the generator's dashboard will record fewer than one hundred conversions. That shortfall is not necessarily a shortfall in demand. It is the consequence of multiplying demand by an observability factor below one.
The same logic applies even when the brand can see a few off-platform matches. Partial recovery can create false reassurance: the dashboard contains genuine linked purchases, but the recovered portion may not be representative of all diverted purchases. In the model, φ is a probability applied to the off-platform share, which makes the relationship transparent. In a real system, recovery can depend on identity continuity, reporting windows, platform policies, and the kinds of customers who cross surfaces. Those complications do not invalidate the mechanism; they make a single observed ROAS even less self-explanatory.
The same relationship appears in ROAS. If m is the product's unit margin, true incremental ROAS for G is:
ROAStrue = mq(sG) / sG
Attributed ROAS is based on the matched conversions instead:
ROASattr = m q̂G(sG) / sG = κ ROAStrue
In this model, the measured value is exactly the true incremental value scaled by κ. Greater distrust widens the gap because more of the purchase journey moves to the marketplace. Greater recovery narrows it because more of the off-platform outcome becomes linkable to G.
That scaling relationship should not be mistaken for a causal estimate. κ tells us how much of the generator's caused outcome remains visible under the stated reporting rules. It does not tell us that the unobserved customers would have failed to buy without the ad. Nor does a low κ imply that the marketplace caused all of the transactions it recorded. It says that the measurement boundary has retained only part of the causal path.
This is why the distinction between missing evidence and evidence of no effect is so important. A generator with a low attributed ROAS may have weak creative, poor targeting, or genuinely little incremental impact. But the same observed number can also arise when the advertising creates demand that completes through a trusted marketplace. Attribution alone cannot tell those explanations apart if the off-platform transactions are absent from the data it uses.
Changing the crediting rule does not solve that identification problem. Last-touch attribution can assign a visible conversion to the final observed interaction. Multi-touch attribution can divide a visible conversion among several observed interactions. A more elaborate model can use timing, clicks, views, or other patterns to redistribute the visible total. None of these choices can assign credit to a purchase that never entered the observable set.
The limitation is structural rather than stylistic. If the marketplace does not pass a usable conversion signal to the brand, the attribution system has no record on which a crediting heuristic can operate. A new rule may change the shares of the conversions already seen, but it cannot infer the missing transactions without adding evidence or assumptions from outside that rule.
The full journey can now be stated compactly. Advertising spend creates intent according to the response curve. Trust determines whether the customer checks out on the generator or the marketplace. The checkout location determines where the sale is booked. Signal recovery determines whether the brand can connect that sale to the original ad. Attributed ROAS reports the last of these conditions as though it were a complete account of the first.
Once the observability factor is visible, the next problem is no longer mysterious. A brand that uses attributed ROAS to set its hurdle is not merely reading a lower number; it is making an allocation decision from a systematically filtered view of demand. The question is how that decision feeds back into spend—and whether the platform that created the intent can be made to look weaker precisely because the journey worked.