Which 6 of 20?
Ranging when shoppers see near-substitutes.

The question

A category manager has 20 energy drinks on the shelf: five brands, four flavors and three pack sizes, at set prices and unit costs. Space allows 6, 8 or 10 products. Which ones should stay to make the most category profit?

The answer depends on where a dropped product's buyers go. In this category they switch along three attributes at once, to different degrees: shoppers who lose their product tend to move to another of the same flavor, the same brand or the same pack. Most of the in-the-moment variation in what they pick is shared that way. That is common in categories with many flavors and sizes, and it's the case where the choice of model matters.

How the test worked

• Survey. A standard choice-based conjoint: 1,000 shoppers, 12 choice tasks each from the 20-product shelf, plus 6 holdout tasks for tuning. Five independent surveys from the same population.
• Known truth. The shoppers' tastes and switching patterns were set in advance, so the true profit of any range can be computed.
• Methods. The standard approach in conjoint software, hierarchical Bayes logit, read out four ways: share of preference and randomized first choice, each on respondent estimates and over posterior draws. Randomized first choice over posterior draws is its strongest documented form. Against it, Chosŏn's model, which estimates the switching from the choices.
• Search. Every method tried every possible range (up to 184,756 of them for 10 of 20), so any difference comes from the model, not from the search.
• Scoring. Each pick is scored by its true profit, as a percentage of the best range's. Three decisions (keep 6, 8 or 10) on five surveys: 15 decisions per method.

Results

• The same wrong cut, every time. Keeping 6 of 20, every standard readout on every survey cut one product it should have kept and kept one it should have cut, giving up 2.8% to 4.8% of the best profit.
• The forecast. Every standard readout under-forecast the profit of its own range by 11% to 20%. Chosŏn's forecast was within 3%. A business case built on the standard forecast starts from a number that is too low.
• Why posterior draws don't fix it. Simulating over posterior draws accounts for uncertainty in what shoppers like. It adds nothing about how products substitute for each other. Randomized first choice adds one similarity setting shared by all attributes, so it can't learn that shoppers stay within flavor more than within pack.

What it means for a real range review

1.7% of category profit is the gap between a good range and the best one. The larger cost is the pattern: a method that misreads substitution makes the same mistake every time, keeps near-duplicates that compete for the same buyers, and under-states what the range will earn.

Limits

• This is a controlled test on simulated shoppers, chosen so the answer is known. One category design, five surveys.
• Substitution here is strong. When shoppers don't see some products as closer substitutes than others, the methods agree, and Chosŏn's Switching tab shows which case a study is in.
• Research models that estimate substitution can come close (error-component and cross-nested logits got within about 0.3% of the best profit in our tests); conjoint platforms don't ship them.
• We are looking for real ranging studies to validate against. If you have one with an outcome to compare, become a beta tester.