What would have sold anyway?
Estimate the underlying demand path without the promotion rather than assuming the previous week or month is a valid counterfactual.
Promotion effectiveness · Causal Decision Intelligence
Did the promotion work, or did sales simply happen during the promotion?
Observed lift is not automatically incremental lift. The useful question is what would likely have happened without the promotion, what changed because of it, and whether that change was worth the margin, stock and commercial cost.
The central distinction
Promotion lift and promotion incrementality are not the same thing.
A before-and-after comparison asks whether sales were higher during the promotion. A causal promotion analysis asks how much of that difference is credibly attributable to the intervention itself.
Seasonality, distribution, media, holidays, competitor activity, price changes, customer mix and underlying demand can all create apparent lift. If those factors would have raised sales anyway, crediting the full increase to the promotion overstates its value.
Simple example
+18% observed lift does not mean +18% incremental demand.
If the counterfactual baseline was already expected to be +12%, the promotion's credible incremental contribution may be closer to the remaining difference, subject to uncertainty and model design.
Expected baseline without promotion
What needs to be separated
Five effects that can make a promotion look better than it really is.
Was demand already due to rise?
Day-of-week, holidays, seasonality and events can coincide with promotional timing and inflate naive lift estimates.
Did more stores or stock drive the result?
Changes in numeric distribution, inventory or availability can create volume changes that are wrongly credited to the promotion.
Did the promoted SKU steal from another?
An SKU can show strong lift while category or portfolio incrementality is weak because demand moved from another product.
Did we discount customers who would buy anyway?
A promotion can look successful overall while destroying margin in segments whose behaviour was unlikely to change.
Choose the measurement design
The right method depends on what the business actually changed.
Randomised test
Best when treatment can be deliberately assigned and a valid holdout is operationally possible.
Matched control / natural experiment
Useful when similar untreated stores, customers, regions or time periods can provide a credible comparison.
Treatment-effect modelling
Useful when promotions were not randomised and the data contains enough variation and relevant confounders to support estimation.
Promotion data: dates, mechanics, depth, duration, channel and eligible products/customers.
Outcome data: units, volume, revenue, margin and where possible category or portfolio outcomes.
Context: base price, distribution, inventory, media, holidays, weather, competitor or market variables where relevant.
Hierarchy: SKU, brand, customer, store, region or channel dimensions that match the decision.
Predictive and causal models do different jobs
Forecasting tells you what is likely to happen. Causal analysis asks what changes because of the promotion.
The original Graphite Note article on this URL demonstrated an important predictive idea: explicitly marking promotions and special dates can materially improve a demand forecast. That is still useful for planning stock and expected sales.
But forecasting a promotion-aware demand curve is not the same as estimating incrementality. A modern promotion workflow can use both: a forecast to establish expected demand and causal analysis to estimate the intervention effect where the data supports it.
Two questions, two outputs
Prediction: “What will sales be?”
Causal: “What will sales be because we run this promotion rather than not run it?”
Promotion-aware forecast
Estimate expected demand while accounting for seasonality, special events and planned promotion timing.
Incremental effect
Estimate the counterfactual and quantify the difference attributable to the intervention, with uncertainty.
Commercial action
Combine effect, margin, cost, inventory, customer response and constraints to decide what to repeat or change.
Legacy Graphite Note walkthrough
The useful 2023 forecasting example is preserved, but its role is clearer now.
The original article showed how marking a promotion as a special event changed a time-series forecast. The historical screenshots can be pulled into this Astro page with the included asset-fetch script; the page itself no longer depends on WordPress content.
See how the GN CPG & Beverages Decision Pack turns this measurement into a commercial decision →
Direct answers
Questions commercial teams ask about promotion effectiveness.
Is promotion lift the same as promotion incrementality?
No. Lift describes a difference in observed outcomes. Incrementality is the portion credibly caused by the promotion relative to a counterfactual of what would have happened without it.
Can we estimate incrementality without a randomised experiment?
Sometimes. Quasi-experimental and observational causal methods can be useful when the data contains enough comparable untreated observations, treatment variation and relevant confounders. The strength of the conclusion depends on the identification design and data.
Can a promotion increase sales and still be a bad decision?
Yes. It can cannibalise another product, discount demand that would have happened anyway, create stock problems, or produce too little incremental margin to justify the cost.
How much history is needed?
There is no universal number. The useful amount depends on frequency, seasonality, how often promotions occur, variation in promotional mechanics and whether comparable untreated observations exist.
What should come out of the analysis?
Not only a lift percentage. The useful output is an evidence-backed decision: which promotion, product, customer, timing or market condition should be repeated, changed, targeted differently or stopped.
Related commercial decisions
Promotion incrementality and price elasticity answer different questions.
Incrementality asks whether a promotion caused extra sales relative to what would have happened anyway. Price elasticity modelling asks how demand responds as price changes. Both can feed broader commercial decisions in the GN CPG & Beverages Decision Pack.
Do not conflate them
Observed lift ≠ causal incrementality ≠ price elasticity
They can use overlapping data, but the estimand and the business decision are different.
Decision IntelligenceFrom signal and prediction to a governed business action.AutoMLAutomate repetitive model-building work without losing business context.Causal Machine LearningSeparate drivers from correlations and estimate intervention effects.Predictive AnalyticsForecast demand, score customers and anticipate business outcomes.Decision OptimisationCompare feasible actions and rank them by expected business impact.Forecasting vs OptimisationSee where prediction ends and the business decision begins.Decision APIDeliver recommendations, expected impact and evidence into workflows.Embedded Decision IntelligenceAdd a decision layer to ERP, CRM, BI, workflow and vertical software.Build vs BuyA practical framework for deciding whether to build, buy or embed.
Bring one promotion question and the data behind it.
We will separate what can be forecast from what can be causally estimated, then structure the output around the commercial action.