Assortment Rationalisation Framework for Dark Stores

The ask

As dark store assortments grew, a long tail of low performers started taking up limited shelf space, adding replenishment and picking work, and in some cases driving disposal losses. The business needed a scalable way to decide which SKUs stay, which need help, and which can go.

1. 01

   Assortment grows

   More SKUs added every month

2. 02

   Long tail expands

   Low-velocity SKUs pile up

3. 03

   Waste increases

   Disposal and inventory costs

Low sales is not the same as weak

Removing low sellers looks easy from the outside, but it's easy to remove the wrong SKU. A SKU can sell little for reasons that have nothing to do with demand:

• New launch

  Not enough data to judge yet.

• Low availability

  Supply held demand back.

• Low discoverability

  Customers couldn't find it.

• Poor pricing

  Priced above competitors.

• Niche subtype

  Needed to keep the category complete.

• Repeat-led demand

  Small but loyal audience.

The real question was whether low performance was structural or fixable. Short shelf life adds a third case: SKUs whose disposal risk outweighs what they sell.

Goal and non-goals

Principles

I designed it as an explainable funnel rather than one composite score, which made it easier to reason about, validate, and trust.

1. 01

   Protect clearly healthy SKUs early

   If a SKU contributes meaningfully to sales or units, it exits the funnel immediately.

2. 02

   Distinguish weak from suppressed demand

   Before concluding demand is low, check whether availability, discoverability or pricing is holding it back.

3. 03

   Explainability over black-box scoring

   Every recommendation traces back to a specific, auditable reason.

4. 04

   Recommend actions, not just labels

   Output 'increase availability' or 'reduce price', not just 'watchlist'. Actionable outputs drive adoption.

5. 05

   Preserve subtype and assortment integrity

   Never delist every SKU in a subtype. Keep at least one so the category has no gaps.

The signals

The key design decision was keeping suppression signals separate from demand signals. A SKU with low sales and poor availability isn't weak, it's constrained. For example, availability below 60% counted as suppression, not performance.

• Demand

  How much it sells

  GMV contribution, units contribution, dark store penetration, subtype penetration.

• Suppression

  What holds it back

  New SKU check, availability (% days in stock), discoverability (impressions, conversion, visibility).

• Commercial

  Price and loyalty

  Price competitiveness vs the lowest competitor price, and repeat purchase index (repeat buyers / unique buyers, 90 days).

The 9-step funnel

Every SKU enters at the top and is checked in order. At each step it either exits with a clear recommendation or moves on. Only SKUs that fail all nine checks are recommended for delisting.

Phase 1: remove clear winners

1. 01

   Sales contributionActive if top 80% GMV

2. 02

   Units soldActive if top 80% units

Phase 2: check for fixable constraints

1. 01

   Dark store penetrationActive if high store spread

2. 02

   New SKU checkWatchlist if recently launched

3. 03

   AvailabilityWatchlist if below 60%

4. 04

   DiscoverabilityWatchlist if low impressions, high conversion

Phase 3: evaluate structural strength

1. 01

   Subtype penetrationActive if high subtype share

2. 02

   Pricing (price competitiveness score)Watchlist if overpriced vs market

3. 03

   Repeat purchaseWatchlist if high repeat index

Disposal override

Perishable and short shelf-life SKUs needed separate treatment. Some weren't just weak, they actively lost money because expected sell-through didn't justify the disposal risk. A SKU can convert decently and still lose money if its shelf life doesn't match how fast it sells.

Click to enlarge

Guardrails

• New launch protection

  Recently launched SKUs are excluded until there's enough data to judge them.

• Availability distortion

  SKUs with under 50 to 80% availability go to the watchlist, since low sales may be supply-driven.

• Subtype wipeout prevention

  If every SKU in a subtype is flagged, the best seller stays as a fallback.

• Premium niche protection

  Low volume doesn't mean low value. High-margin niche SKUs go to manual review.

• Loyalty-driven SKUs

  SKUs with a high repeat purchase index stay even at low volume, because they serve loyal customers.

• Price correction path

  Overpriced SKUs go to price correction instead of delisting. The problem is commercial, not structural.

Three outcomes

• Active

  Structurally healthy

  Keep in the assortment.

• Watchlist

  Could recover with help

  Fix the underlying issue: new SKU, availability, discoverability, price, forecasting, repeat buyers, or subtype protection.

• Delist

  Still weak after every check

  Remove from the assortment.

In most watchlist cases, one fixable constraint was the difference between recovery and removal. Three example walkthroughs (SKU names are illustrative; signals and outcomes are from real framework outputs):

Click to enlarge

Click to enlarge

Click to enlarge

What it found

1. 01

   12,400+

   SKUs evaluated

2. 02

   62%

   recommended active

3. 03

   23%

   moved to watchlist

4. 04

   15%

   recommended delist

Freed shelf space went to new SKU trials, and some of them became high performers within the review period.

Click to enlarge

Click to enlarge

How it ran

The framework ran monthly on the 28th. Each category manager got an email with their category's watchlist and delist recommendations, and the reasoning behind each flag.

1. 01

   Historical backtesting

   Run against the past 90 days and compare with actual delist outcomes.

2. 02

   False positive analysis

   Audit delist recommendations with high cart inclusion or high margin.

3. 03

   Monthly category emailer

   Flagged SKUs and reasoning sent to category managers each cycle.

4. 04

   30/60/90-day monitoring

   Track disposal rate, availability and space utilisation after delisting.

5. 05

   Space reallocation simulation

   Model the impact of freed space on category revenue.

6. 06

   Auto-delist on the 5th

   SKUs with no category manager action or comment by the 5th were delisted automatically.

What I'd build next

V1 was kept explainable and rule-based on purpose, so category managers could trust it and override it. Later versions move from flagging weak SKUs to optimising what replaces them.

• V1 · Current

  Explainable funnel

  Rules-based, 9-step sequential evaluation, with disposal override logic.

• V2 · Near term

  Sharper economics

  Profitability per cubic volume, cart breaker score, customer returns analysis, better price sensitivity logic.

• V3 · Future

  Optimised assortment

  Optimisation-based planning, storage-type-aware decisions, and allocation constrained by demand, margin and volume.