Industry: Retail Product: ModelRisk Application: Product placement
A category team had a planogram proposal: move a hero SKU to a premium eye-level end-cap across 220 stores. The gross case was easy to love — the hero SKU's weekly sales were forecast to climb by about $6,498 network-wide from the better location. The planning sheet netted out a modest 20% cannibalization of the neighbouring SKUs that would lose the space, subtracted the merchandising cost, and landed on a healthy +$3,128/week. A clear win.
It is a clear win only if cannibalization really is 20%. When the ModelRisk rebuild let the location lift, the cannibalization fraction, the attention drag on adjacent SKUs, and store traffic all vary together, the net category uplift collapsed to a mean of $724/week — under a quarter of the deterministic figure — and the probability that the move is even net-positive fell to 56.2%. The hero SKU genuinely sells more; the category barely does. The gap is the money the planogram quietly moves from one shelf to the next rather than creates.
The deterministic sheet treats placement as addition: hero gains $X, adjacents lose a fixed fraction of it, subtract a cost, done. But cannibalization is the least observable and most variable number in the chain — is the end-cap pulling in genuinely new shoppers, or just relocating purchases that would have happened two aisles over? A single optimistic guess (here, 20%) is doing all the work, and it is exactly the guess no one has data for. Worse, lift and cannibalization co-move: a high-traffic week lifts the hero SKU and intensifies the shopper-attention transfer away from its neighbours at the same time. Averaging point estimates of co-moving quantities does not give the average of their net — it gives a number that is almost always too kind, because the cannibalization line is the one that gets rounded down.
The whole case is the distance between two distributions — the gross hero lift, and the net category uplift after cannibalization, attention drag, and space cost:
The gross hero uplift (red) sits far to the right with a mean of $6,498 — this is the number that sells the proposal in a slide. The net category uplift (green), once cannibalization and the adjacent-SKU attention drag are subtracted, has a mean of just $724 and a P10 of -$2,338 — a meaningful slice of the distribution is below zero. The deterministic plan's +$3,128 sits between the two, closer to the gross figure than the net, because its 20% cannibalization assumption is too generous. The headline reads off the zero line: P(net uplift > 0) = 56.2%. The move is favourable on average, but it is barely better than a coin flip, and a point estimate hid that entirely behind a single confident number.
The mean decomposition makes the leakage explicit:
Starting from the +$6,498 hero gross gain, the category gives back -$2,932 to cannibalization (sales transferred from adjacent SKUs), -$1,042 to attention drag (the adjacents lose premium space and shopper attention beyond the direct transfer), and -$1,800 to incremental space and merchandising cost, leaving +$724 net. Cannibalization plus attention drag alone erase $3,974 — more than 60% of the gross lift. The deterministic sheet booked the $6,498 and a thin haircut; the simulation shows that nearly all of the apparent gain is recycled demand, not incremental demand.
Because cannibalization is the input no one can pin down, the team swept it and watched the probability of a net-positive outcome fall:
At 0% cannibalization the move clears zero 80% of the time; at the modeled mean of 45% it is at 56%; and the curve crosses the coin-flip 50% line at about 54% cannibalization. Above roughly 54%, the placement is more likely to destroy category value than create it. That single crossing point reframes the decision: the proposal is not "should we improve the hero SKU's location" — to which the answer is obviously yes — but "are we confident cannibalization stays below ~54%," which is a measurable, testable question. The slope of this line is the business case for a controlled store test before a full rollout.
A tornado on mean net uplift ranks the sources of uncertainty:
The location-lift multiplier is the largest mover at +/-$2,377 — how much the premium spot actually lifts the hero SKU is the single biggest swing, and the easiest thing to measure in a test. Cannibalization fraction is second at +/-$1,287, the villain of the net story. Attention drag on adjacents follows at +/-$655, then the lift spread (+/-$285) and the store-traffic factor (+/-$22), which barely matters because it scales gross and cannibalization together and largely cancels in the net. The reading is clean: a store test should be instrumented to measure the location lift and the cannibalization fraction specifically, because those two inputs own almost the entire uncertainty in whether the planogram pays off.
A planogram change that brags about its gross hero-SKU lift is measuring the wrong thing. Moving a product to a better shelf almost always sells more of that product; the only question that matters for the category P&L is how much of the gain is genuinely new versus borrowed from the neighbours. Here the gross lift was a confident $6,498 and the net category uplift was a marginal $724 with a 44% chance of being negative. Monte Carlo simulation in ModelRisk separated the recycled demand from the incremental demand and turned an apparent slam-dunk into a test-it-first decision.