Baseline: expected without promotion
Treatment: price, exposure, audience
Uplift: actual less baseline
Spillover: pull-forward and cannibalization
Value: contribution and inventory
Activity-period growth is not incremental impact
Holiday, trend, store expansion, stock recovery, and channel traffic can raise sales without the promotion. Uplift is the difference between observed performance and an unobserved no-promotion counterfactual.
Evaluate net contribution rather than units alone. Discount, media, fulfilment, refunds, pull-forward, and related-item decline can reverse the economic conclusion.
Model promotion facts at reconciled grain
Store promotion ID, items, locations, audience, dates, rules, budget, placement, and version. Order lines retain list price, merchant and platform discount, quantity, refund, and activity key.
Connect configuration, coupon, media, and order facts with stable identifiers. Preserve unmatched spend and enforce conservation for order-level allocation.
Build a comparable baseline
Depending on data, use comparable weekdays, seasonal forecasting, matched stores or items, pre-trend, or an untreated control. Report a range rather than false precision.
Use only information available before launch. New and sparse items may require portfolio results or an explicit uncertain status.
Test control comparability and contamination
Controls need similar audience, scale, trend, stock, and exposure. Corporate media or spillover may contaminate untreated stores.
Freeze selection before viewing outcomes and diagnose pre-period fit. A stock-out control is not natural demand.
Measure purchase pull-forward
Discounts can move future purchases into the campaign window. Use a post-period aligned with normal repeat cycles and compare cumulative pre, during, and after behavior.
Report immediate and after-effect-adjusted uplift, with maturity. The window depends on product and cannot be universal.
Include cannibalization and basket spillover
A promoted SKU may replace another size or brand, or lift complementary products. Review item, category, price tier, and basket contribution.
Association is not causality. Predefine experiment metrics and guardrails instead of selecting favorable slices after the event.
Translate uplift into contribution and cash
Subtract product, merchant discount, media, platform, fulfilment, service, and cannibalization cost from incremental net revenue. Keep fixed-cost treatment explicit.
Inventory clearance may justify lower unit margin, while discounting scarce stock may destroy full-price sales. Match evaluation to the declared objective.
Use experiments or transparent quasi-experiments
Randomize store, audience, or time where feasible. Otherwise use matching, stratification, differences, or time-series baselines and disclose assumptions.
Track actual exposure and contamination. No observational model turns an estimate into certain causality.
Accept data, baseline, and maturity
Reconcile configuration, orders, and cost; validate pre-fit, inventory, stock-outs, pull-forward, cannibalization, net contribution, and uncertainty. Backtest both successful and failed campaigns.
Do not fill missing spend or refunds with zero. An incomplete ROI should remain incomplete.
BI0 operating boundary
Retain objective, hypothesis, baseline version, exposure, uplift range, contribution, inventory, owner, and next experiment.
BI0 may be evaluated for integrating promotion, order, inventory, and cost data. Causal methods, allocation, and experiment orchestration require project verification, and uplift is never guaranteed.
Public references
Take the next step with BI0.AI
Talk through a real business scenario and see how governed AI BI can fit your team.
Explore BI0.AI