Every promotional dollar that doesn't move incremental units is a margin leak. Yet most retailers run promos blind to the real lift: they see sales spike during a discount week and assume causation. They don't see the customer who was going to buy anyway, now paying less. They don't see the adjacent category that cannibalized when the deal ran. They don't see the regional variation — a strong lift in one metro against a flat one in another, running at exactly the same depth.
By the time the promo closes and the finance team reconciles, the damage is done. The discount is history. And next quarter, the category manager runs another promo at the same depth, repeating the same invisible leak. The problem is not effort or intent — it is that the signal needed to price a promotion correctly arrives after the promotion is already over.
The Invisible Leak: Cannibalization and Over-Discounting
Picture a regional grocery chain running a promotion on a flagship brand. The marketing team picks a discount depth, sales in that week jump, and the headline reads like a win. But hidden inside that jump is the cannibal — and finding it requires separating three very different things that all look identical on a weekly sales report.
The first is genuine incremental volume: customers who bought because of the discount and would not have bought otherwise. That is the only lift worth paying for. The second is timing-shifted buying: customers who were going to purchase anyway, who simply moved their basket forward to catch the deal. They cost you margin and return nothing in net volume. The third is cannibalization: volume pulled away from an adjacent, full-price category that you also own, so the "lift" on the promoted item is partly a transfer from your own shelf.
On the weekly sales line, all three read as one number. A transaction-level analysis pulls them apart. It can show that a sizable share of promoted-item sales came from shoppers who bought the same item at full price the prior week — accelerated, not incremental — and that an adjacent category quietly softened in the same window. The headline lift and the real incremental lift turn out to be two very different figures, and only one of them justifies the discount.
The second leak is regional. The same discount depth does not buy the same lift everywhere. In a metro with high shopper density and low private-label penetration, a given depth can drive genuinely incremental volume. In a more price-sensitive metro with strong alternatives on the shelf, the same depth drives far less true lift — most of the uplift is deal-seekers accelerating purchases they would have made anyway.
Running one discount at one depth across all regions is the operational default for most retailers, because granular regional promo planning is labor-intensive and depends on granular demand forecasting that few teams have. The cost of that simplicity is margin. A model trained on transaction-level promo outcomes across quarters and regions sees that variation — and flags which discount depths actually pay for themselves in each region, instead of assuming the chain average holds everywhere.
Reading Incremental Lift Before the Promo Runs
The data that makes this visible already lives in your POS and your transaction history. What is missing is the model that separates signal from noise.
Start with the baseline: for this category in this region, what is normal weekly demand? A short-horizon demand model answers that from seasonality, day-of-week patterns, promo calendars and local events. This is the same class of signal that drives an 18% drop in stockouts — demand sensing is the foundation everything else sits on. Without a trustworthy baseline, you cannot say what a promotion added, because you never knew what would have happened without it.
Next: when a promotion runs, how much of the sales uplift is incremental versus accelerated or cannibalized? This requires training on a multi-period history of promotions, sales and customer movement. For each past promotion, you measure actual incremental lift by comparing the promoted week to the predicted baseline and controlling for the customer segments that moved. A promotion that looked like a large uplift but was mostly timing-shifted buying becomes legible — its real incremental contribution is finally separated from the noise that surrounded it.
Then: what depth of discount was actually required to drive that incremental lift? The relationship between depth and lift is not linear, and it is not the same across regions or segments. In some places a shallower discount drives the same true lift that a deeper one drives elsewhere. A model that learns that mapping makes the next depth decision precise instead of guessed, and it does so from your own history rather than a vendor's rule of thumb.
The outputs are concrete and merchant-facing:
- Regional lift curves: for each category by region, what incremental lift does each candidate discount depth actually drive — per region, not a chain average?
- Cannibalization scorecards: which adjacent categories leak volume when this promo runs, and by how much?
- Depth recommendation: to move the target units and hold the margin line, what discount depth is optimal for each region?
Over-Discounting and the Margin Cliff
The economics are sharp, even without putting invented figures on them. When a discount drives strong incremental lift, the promo pays for itself and adds volume you would not otherwise have had. When the incremental lift is weak, the same discount destroys margin — you give away price on units that would have sold anyway, and capture little genuine demand in return. The gap between those two outcomes is enormous, and on the weekly sales report they can look nearly identical.
The difference between them is not guesswork. It is a historical pattern hiding in your transaction data. Categories with high repeat-purchase rates and low price sensitivity — a staple a household buys on a fixed cadence — tend to show small incremental lifts even at steep discounts, because most of the volume is timing-shifted buying you were going to get regardless. Categories with high deal-sensitivity and low repeat — a discretionary or impulse item — tend to show steeper lift curves, where a modest discount moves volume that genuinely would not have moved otherwise.
A promo-optimization model reads those patterns. For each category it builds a discount-response curve — the actual relationship between depth and incremental lift — from historical promos. Then, when a category manager proposes a new promotion, the model surfaces the expected lift and margin impact at the proposed depth. If it flags that the depth will drive only weak incremental lift and a negative margin benefit at current penetration, the manager can pull the depth back, retarget the customer segment, or move that spend to a lever that actually works.
This is where explainability earns its keep. A forecast a merchant cannot interrogate is a forecast a merchant will override on instinct. Because every lift number traces back to its drivers — baseline demand, promo history, segment movement, cannibalization — a category manager can see why the model recommends a shallower depth, agree or disagree on the merits, and adjust one region without breaking the chain-wide logic. That is the difference between a forecast merchants act on and one they merely admire. The result is fewer promos that destroy margin in the name of a headline sales bump, and more precise allocation of promotional spend to the categories and regions where it genuinely pays back.
- 18%
- Less stockout from demand sensing
- 4–6 weeks
- Assessment to first ranked roadmap
- Per-region
- Lift curves, not chain averages
End-of-Life Markdowns: Timing Clears Overstock Without Torching Margin
A different problem hits retailers at the end of a season. Inventory from one assortment is still on the shelf when the next is arriving. That stock has to move, and the category manager's toolbox often has one blunt lever: markdown depth, taken on a hunch and held until it clears.
A forecast-driven approach is sharper than the hunch. A demand model that accounts for seasonality and remaining shelf-life can predict when demand for the outgoing assortment will fall away. That window — the point at which the sell-through rate collapses — is when the markdown should hit. Take it too early and you give away margin on units that would have sold at a shallower discount later. Take it too late and you burn shelf space and invite out-of-stocks on the fresh inventory that should be earning that space instead.
The model surfaces that timing. For a given overstock quantity, shelf-life remaining and full-price sell-through rate, it recommends the markdown depth and the week it should start. The result is markdowns that clear faster and at shallower depths than a flat seasonal rule would produce, protecting margin on the units that do move and freeing the shelf on schedule.
This lever is especially powerful combined with replenishment optimization. If demand sensing is catching a stockout early, the system can recommend a small, targeted pull-forward in that region — cheaper than a missed sale. If overstock is building in a store cluster, an end-of-life markdown at the right depth and the right week frees that space before it becomes dead inventory. The same models that keep the shelf full are the ones that clear it efficiently, because they are reading the same underlying demand signal from both directions.
The difference between real incremental lift and accelerated buying is the difference between returning margin and burning it — and most retailers never see the delta until the quarter closes.
Where to Start
The assessment phase is straightforward. Map your promotional spend over recent quarters, calculate the actual incremental lift for each promotion from transaction data, and surface the promotions that destroyed margin. That analysis is the anchor: it shows the magnitude of the leak and the categories where it is largest, in your own numbers rather than an industry benchmark.
From there you build a ranked list — which category-by-region pairings carry the highest promotional volume and the largest historical lift variance? Those are where the model pays back first. A 4–6 week assessment produces a prioritized roadmap naming the first categories to model, the data to gather, and where the recovery is concentrated, mapped against your POS, ERP and replenishment data quality.
The transformation phase is faster than it sounds. You build the incremental-lift model on high-velocity categories, wire it into the POS and markdown workflows buyers already use, and run promotions at model-recommended depths against holdout categories to validate the lift curves. Because the recommendation writes back into the existing order and markdown cycle rather than a separate dashboard, it lands where the decision is actually made. Then it hardens into a standing system: every promo passes through the model before it runs, and every promo measures actual lift afterward, so the model sharpens with each cycle.
The sustain phase watches for drift — new competitors, category migrations, shopper demographic shifts, new-product cold starts — and retrains the lift curves against new promo calendars and seasonal breaks. The same closed loop that keeps demand sensing honest keeps promo-lift prediction honest, catching the silent failure modes before they show up as either empty shelves or marked-down overstock.
“The difference between real incremental lift and accelerated buying is the difference between returning margin and burning it — and most retailers never see the delta until the quarter closes.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production — sovereign models, governed data, and agents you can audit. Start with a value-first assessment.