Why ROAS cannot tell you what your ads caused

A marketing platform reports 5x ROAS. Your backend reports 2x. These figures are illustrative rather than an industry benchmark, but they expose a real measurement problem: different rules can assign different amounts of revenue to the same advertising. Choosing the lower figure because it comes from the system that records the transaction does not resolve what caused the sale.

The key question is the decision the measurement must support. Attribution-based ROAS assigns credit for observed revenue according to a set of rules. A budget decision can require a causal answer: how much revenue changed because the advertising ran? That requires an estimate of what would have happened without the advertising.

Platform and backend ROAS can answer the wrong question

Platform and backend reporting can use different rules to connect a sale with earlier marketing activity. Consider a customer who sees a Meta ad, later searches the brand on Google, clicks an ad, and purchases. A backend using last-click attribution can assign the order to Google. The earlier Meta exposure remains part of the observed customer journey.

This creates two measurement tasks. Attribution decides how observed interactions receive credit. Causal measurement estimates whether an intervention changed an outcome. If an executive wants to know whether an additional advertising dollar generated additional revenue, assigning credit among touches cannot establish the answer by itself.

Backend data records transactions

A transaction system can accurately record that €100 arrived, who purchased, and when. Its attribution rule is a separate layer. In the Meta-to-Google example, a final-click rule credits the later Google interaction while assigning no credit to the earlier Meta exposure. That accounting choice does not tell management whether the customer would have purchased without either interaction.

Platform dashboards create a different risk. Multiple systems can associate the same transaction with interactions they recorded, creating duplicate revenue when their reports are simply added together. Executives need to understand how each reported result was observed, modeled, or assigned before comparing systems.

The measurement method can influence which channels receive credit

A last-click rule assigns the conversion to the final qualifying interaction before purchase. In the earlier customer journey, the later Google click receives the credit. An earlier Meta impression can receive no backend credit under that rule even if it occurred before the search. The example shows how the rule shapes the recorded customer path.

A weak reported result can have more than one explanation: the campaign may have little effect, or the chosen attribution rule may assign little credit to the way that campaign operates. Attribution alone cannot determine which explanation is correct. An unclicked impression may have influenced a purchase, and the customer may also have purchased without it. Establishing incremental effect requires a counterfactual: an estimate of the outcome under a comparable condition in which the advertising treatment was absent or different.

Better attribution improves credit allocation

Multi-touch attribution assigns credit across multiple interactions rather than giving all credit to one qualifying touch. Richer customer-path data can provide more information for that allocation. Identity resolution, server-side events, deduplication, and additional touch data can change the information available to a model. The causal question still requires evidence about the counterfactual.

Return to the illustrative 5x platform ROAS and 2x backend ROAS. Reconciling those reports can reveal where transactions overlap and where attribution rules differ. A reconciled value still does not show how many customers would have purchased in the absence of the advertising. That missing quantity is why causal designs matter for incremental budget decisions.

The decision question should define the experiment. Management might test whether removing a particular channel in selected markets changes revenue, or whether a defined increase in spend produces incremental conversions. These are bounded causal claims, so their results apply within the tested conditions. The transaction system can supply the revenue outcome without having to assign causal credit to each customer’s individual journey.

Management should specify the decision first and choose a design capable of identifying the effect relevant to that decision. Experimental estimates carry uncertainty and depend on the validity of the design. The relevant question is whether the design can produce evidence strong enough for the budget decision being made.

Use each measurement tool for the decision it can support

Attribution supports decisions that require credit allocation across observed or modeled interactions. Teams can use it to investigate customer paths and compare activity under consistent rules. Its conclusions should stay within the rules and data used to assign credit.

Experiments address a different decision. A well-designed experiment can estimate whether a defined change in advertising produced an incremental change in an outcome under the tested conditions. Experimental estimates still carry uncertainty and depend on the validity of the design. For executives, their value is the direct connection between a defined intervention and a defined business outcome.

Automated allocation raises the stakes because measurement can feed directly into spending decisions. Executives should validate the objective, the measurement feeding it, and the causal relevance of that measurement before using it to support a material budget decision.

Key takeaways for decision-makers

• Match measurement to the decision: Platform and backend ROAS can assign revenue differently, but neither alone establishes what advertising caused. Use attribution for credit allocation and causal measurement when budget decisions depend on incremental impact.
• Treat backend attribution as a rule: Transaction systems can accurately record revenue while still assigning marketing credit according to a chosen attribution model. Executives should separate reliable transaction data from claims about what caused the transaction.
• Recognize how attribution rules shape channel performance: Last-click and other attribution methods can favor some interactions while giving others little or no credit. A weak attributed result does not by itself prove that a channel had little incremental effect.
• Use better attribution for better allocation: Richer customer-path data, identity resolution, and multi-touch models can improve how credit is distributed. They still cannot establish what would have happened without the advertising.
• Use each measurement tool for the decision it supports: Attribution helps analyze and allocate credit across interactions, while well-designed experiments can estimate the effect of a defined advertising change. Before automating material budget decisions, validate that the measurement provides evidence relevant to the decision.