A marketing mix model that gives affiliate less credit than your network dashboard isn't automatically broken. The reports may be answering different questions. Before arguing for a budget change, ask what each number represents and whether both use the same sales, dates and costs. That makes for a better meeting than arriving with two charts and choosing the friendlier one.
Put the disagreement into numbers
Consider a fictional retailer whose affiliate dashboard reports $100,000 in approved sales and $10,000 in commissions for a campaign period. The simple sales-to-commission ratio is 10 to 1. Now suppose its measurement team estimates $25,000 in incremental revenue using a model, with substantial uncertainty. Dividing that estimate by commissions gives 2.5 to 1, but neither calculation includes all costs or product margin. These are illustrative figures, not research findings or a client result.
The first ratio describes sales credited under the program's rules. The second uses an estimate of additional revenue relative to a modeled alternative. The gap alone does not prove stolen credit, wasted commission or a faulty model. Ask the measurement team to explain the comparison and uncertainty before using either ratio to approve spending.
What the technical guidance supports
Google's Meridian documentation describes marketing mix modeling as a causal-inference approach whose interpretation depends on assumptions and model specification. Its data guidance calls for aligned time-series inputs and distinguishes paid media, organic media, non-media treatments and controls. That supports careful input review; it does not establish that every affiliate program is systematically undervalued.
Google also documents a particular paid-search problem: underlying demand can affect both search activity and sales. Query-volume controls can help address that confounding. This is a reason to inspect the model's treatment of demand, not evidence that affiliate activity caused a rise in branded searches.
Read the sponsored research in context
A May 5, 2026 release from Rakuten Rewards describes research it commissioned from Prohaska Consulting, drawing on more than two dozen interviews and case-study analysis. The release identifies concerns with grouping unlike affiliates together and modeling commissions that track completed sales. Those are the sponsor's reported findings, not proof that your own model undercounts affiliate. Use the report to ask better input questions rather than treating it as a channel-wide budget verdict.
Ask for a shared input sheet
For the fictional retailer, request a single worksheet showing the reporting period, sales definition, commissions, network fees, placement costs, refund treatment and any currency conversion. Then ask how affiliate activity enters the model. Is it represented by cost, clicks or another input? Are different partner activities combined? These are diagnostic questions for the model owner, not a requirement to force affiliate into a particular model.
For example, a missing $2,000 placement fee would change the dashboard's sales-to-listed-cost ratio from 10 to 1 to about 8.33 to 1: $100,000 divided by $12,000. That fixes the cost denominator. It still doesn't establish incremental return or profit. A small accounting correction and a causal measurement claim deserve separate answers.
Check assumptions before prescribing a test
Ask whether promotions, stock changes or demand shifts coincide with the activity you want to measure. If several channels always rise together, request the uncertainty range and an explanation of what the available data can distinguish. A model can be useful while leaving an individual channel decision uncertain. Raising a prior or changing an attribution weight merely to match the dashboard would avoid the question.
If an experiment is warranted, agree on the outcome, comparison groups, timing, likely spillover and required sample size with a qualified measurement owner. A geo holdout is one possible design, not an automatic answer for every affiliate program. Confirm the provider's actual capabilities and the feasibility of changing exposure before committing to a test.
Leave the meeting with a decision you can defend
For the fictional example, the immediate decision is to reconcile the missing costs, review the model's assumptions and defer a confident channel-wide budget claim until the discrepancy is explained. That is progress. The goal is a decision supported by the evidence you have, with a clear plan for the evidence you still need.
The earlier version presented the research as blanket confirmation and cited an unsupported 60% enterprise adoption figure. This update identifies the sponsor, distinguishes its reported findings from conclusions about an individual program, and removes the unsupported percentage. A stronger budget argument starts with the actual inputs and the limits of the evidence.