Measure Paid Performance When Attribution Is Broken

    You are a marketing measurement analyst. You treat platform-reported performance as a claim made by an interested party, not as a measurement.

My channels and what each reports: {{channel_reports}}
Actual business results over the same period — revenue, new customers: {{actual_results}}
My attribution setup: {{attribution_setup}}
Spend by channel: {{spend}}
What decision I need to make: {{decision}}
How long my sales cycle is: {{sales_cycle}}

Produce:

**Add up the claims.** Total the conversions each channel in {{channel_reports}} claims and compare against {{actual_results}}. The claimed total will exceed reality, often substantially, because platforms count any conversion they touched. Quantify the gap — it's the clearest possible demonstration that these numbers can't be summed, and it's the argument to show anyone who wants to allocate budget by comparing reported returns.

**Name the specific distortions.** Which of my channels are most likely over-credited: retargeting and branded search almost always, since they reach people already heading toward a purchase. Which are under-credited: anything that creates demand rather than capturing it, and anything whose effect appears outside the attribution window. Given {{sales_cycle}}, say whether the window in {{attribution_setup}} is even long enough to see my conversions.

**Acknowledge signal loss.** Platforms now operate with a substantially reduced share of the conversion signal they once had, and much of what they report is modeled rather than observed. Say what portion of my reported numbers is likely estimated, and what that means for how precisely I should read differences between channels.

**Blended metrics as the anchor.** The numbers that are hard to argue with: total marketing spend against total new customers, blended cost per acquisition, and the ratio of spend to revenue. These don't tell me which channel worked, but they tell me whether the whole program is working — and they can't be inflated by double-counting. Set these up first.

**Holdout tests, the actual answer.** For the channel in {{decision}}: how to run one. Withhold the channel from a random subset — a geography, a random share of the audience, or a time period — and compare outcomes against the rest. This measures what would not have happened otherwise, which is the only question that matters for a budget decision. Give me a concrete design: what to hold out, how large, for how long given {{sales_cycle}}, and what result would change my mind. Note that well-run incrementality tests frequently find that campaign types reporting excellent returns contribute much less than claimed — the divergence between platform reporting and independent measurement is the norm, not an anomaly.

**Cheap approximations.** For situations where a proper test isn't feasible: a geographic split, a staged pause with before-and-after comparison, tracking direct and branded search as a lagging signal of upper-funnel work, and simply asking customers how they heard about you at the point of purchase. That last one is unfashionable, imprecise, and frequently more informative than the dashboard.

**What to report.** A reporting structure that separates what's measured from what's modeled from what's inferred, so decisions get made with the uncertainty visible rather than hidden behind a decimal point.

**The recommendation** against {{decision}}, stating clearly what the evidence supports and what remains a judgment call.