How AI Forecasts Demand Cannibalization Across SKUs

Your new colorway "sold great," until you noticed the original colorway's sales fell by almost exactly the same amount. Total revenue barely moved; you just paid for a launch, split one demand pool across two SKUs, and doubled the inventory carrying the same sales. The launch didn't fail. The forecast did, because it never asked where the new sales would come from.

Demand cannibalization forecasting predicts how a new or promoted product pulls demand away from your existing SKUs, rather than adding net-new sales. It models substitution between similar items and the halo effect, where a launch lifts related products. The point is to forecast the net change across the assortment, so you don't double-buy for demand that simply shifts between SKUs.

Key takeaways

• A launch can succeed while the assortment stands still: if the new SKU's sales come out of its siblings, demand didn't grow, it moved house.
• Substitutes cannibalize, complements catch halo: the direction of the effect follows the relationship between products, which you can map before the numbers.
• The forecast that matters is the net: new SKU volume, minus what it pulls from substitutes, plus what it lifts in complements.
• The buy decision splits into several: the launch order for the new SKU, trimmed reorders for its substitutes, a nudge up for complements.

What is demand cannibalization?

Cannibalization is when one of your SKUs grows at the expense of another of your SKUs, instead of at the expense of a competitor or a non-buyer. The new colorway pulling buyers off the original, the bundle eating the solo product, the promoted variant hollowing out the full-price one. From the single SKU's view it looks like success; from the assortment's view, demand didn't grow, it moved house.

Cannibalization vs net-new growth

The question that separates them: would this customer have bought something from you anyway? If yes, and the launch just redirected the choice, that's substitution. If the launch converted someone who otherwise wouldn't have purchased (new audience, new use case, new price tier), that's net-new. Real launches are a blend, and the blend ratio is the single most important number in deciding how much total inventory the assortment needs after the launch. High substitution isn't automatically bad (a fresher product defending the same demand can be strategy), but it must be bought for as replacement, not as growth.

What's the halo effect and how does it differ?

The halo effect is cannibalization's positive mirror: a launch or promotion that lifts related SKUs instead of eating them. The new dress drives sales of the matching bag; the heavily promoted hero product pulls traffic that spills into the rest of the collection; the entry-price item recruits customers who trade up later. Same mechanism (one SKU's event changing its neighbors' demand), opposite sign.

Substitution vs halo: which neighbors move which way

The direction mostly follows the relationship between products. Substitutes (same job, overlapping audience: colorways, sizes-of-the-same-thing, near-identical price points) cannibalize each other. Complements (used together, or sequential in the customer journey: the accessory, the refill, the matching piece) catch halo. Unrelated SKUs mostly stay put, moving only with sitewide traffic. Mapping which of your products are substitutes and which are complements is the qualitative half of the forecast; the numbers hang off that map.

How do you model cannibalization for a launch?

Three steps, run in order:

• Estimate the substitution rate from similar past launches. When you last added a variant to this family, how much of its volume came out of the siblings? That share is your starting point.
• Apply it to the affected SKUs, weighted by how close each neighbor sits to the new product in price, look, and use case. The nearest substitute absorbs the most.
• Forecast the net change across the assortment, adding back any halo you expect on complements, so the total inventory plan reflects demand that moved as well as demand that grew.

Past launches are the evidence for every number in that sequence, which is why recording family-level results at each launch pays for itself at the next one.

A worked shape of the net math. The new colorway forecasts 400 units for launch quarter. History with this product family says roughly 50% of a new variant's volume substitutes from siblings: 200 units cannibalized, spread over the two nearest colorways. The matching accessory typically catches a small halo, call it 30 units. Net assortment change: 400 − 200 + 30 = 230 incremental units. Buy the new SKU at 400, but cut the siblings' reorders by their 200, or the assortment quietly carries 200 units of double-bought stock. (Forecasting the new product's own 400 is owned by the new-product forecast guide; promotions have the same cross-SKU logic on a shorter clock, covered in promotion demand and lift.)

Adjusting the buy across affected SKUs

The output of a cannibalization forecast is never one number, it's a set of adjustments: the launch buy for the new SKU, trimmed reorders for its substitutes, and a nudge up for complements catching halo. This is assortment-level thinking, and it's where per-SKU forecasting quietly hits its ceiling: each SKU's forecast can be individually defensible while the assortment total double-counts the same customers. Which is exactly the seam AI is built for. Conative AI's deep learning models forecast at the assortment level, learning substitution and halo relationships from how your products have actually moved together, and its product analytics show the family-level view (which variants split one demand pool, which designs genuinely expanded it) so the next launch decision starts from evidence. AI-powered demand forecasting that nets the assortment instead of adding it. Start your free trial.

Frequently asked questions

What's an example of product cannibalization?

A brand launches a second colorway of its best-selling bag. The new color sells 400 units in its first quarter, while the original drops by 350 against forecast. Total family sales barely grew: most buyers simply chose the new color instead of the old one. The launch split existing demand across two SKUs rather than creating new demand.

How do you measure cannibalization rate?

Compare the affected SKUs' actual sales after a launch against what their forecast said they'd have sold without it. The shortfall attributable to the launch, divided by the new product's sales, is the cannibalization rate. If the new SKU sold 400 and siblings fell 200 below their expected line, the rate is roughly 50%.

Is cannibalization always bad?

No. Deliberate cannibalization is often strategy: a refreshed product defending demand before a competitor takes it, or a better-margin variant absorbing volume from a weaker one. It becomes a problem when it's unplanned, because then the inventory math double-counts: you bought the new SKU for growth that was actually replacement, and overstock lands on both.

What's the difference between cannibalization and the halo effect?

Direction. Cannibalization is a launch or promotion pulling demand away from related SKUs, typically substitutes that do the same job for the same buyer. The halo effect is the opposite: the event lifting related SKUs, typically complements bought alongside or afterward. A single launch usually produces both at once, on different neighbors.

How does cannibalization affect your buy quantity?

It splits one decision into several. The new SKU gets its launch buy, but its close substitutes need trimmed reorders, because part of their demand is about to migrate. Skip the trim and the assortment double-buys the same customers. Complements catching halo may earn a small increase. The net assortment change is what your total inventory should track.

Can AI predict SKU substitution?

Yes. Machine learning models learn substitution and halo relationships from your own history: how product families moved when variants launched, which items trade demand and which lift together. That lets a launch forecast arrive as a net assortment picture, with the expected pull on each neighbor, rather than a standalone number that ignores where the sales come from.