AI and attribution: better optimisation starts with better measurement

One purchase, several claims

Browser and mobile privacy changes have reduced observable customer journeys. Platforms estimate missing signals and apply their own attribution windows. Meta, Google and analytics can consequently report different revenue totals without any one report describing the entire business.

An AI assistant that accepts every reported number at face value can amplify the discrepancy rather than correct it.

The budget allocation risk

Demand creation often precedes the final measurable click. If investment moves towards channels that capture existing intent, prospecting may lose budget even when it contributes to later purchases. That can weaken the supply of new customers while reported efficiency initially looks healthy.

• Reconcile attributed revenue with actual store revenue.
• Track marketing efficiency ratio and new-customer CAC consistently.
• Use post-purchase surveys as supporting evidence, not definitive attribution.
• Design holdout or geographic tests when the business has sufficient scale.

Treat recommendations as hypotheses

Ask what the model can observe, which assumptions it makes and what evidence would change the recommendation. A responsible reporting process explains the gap between channel metrics and commercial results rather than presenting a single ROAS figure as settled truth.