Worked example: where will a new product's sales come from?

Step 1: Build a simulator by hand

A conjoint study gives every product an appeal score, built from what shoppers said about its brand, flavor, size and price. The simulator turns appeal into share with one rule: each option's share is its appeal divided by the total appeal of everything on the shelf. Buying nothing counts as an option too.

Here is a small energy-drink shelf. Fill in the share of each option, in percent.

That is the whole simulator: change a product, recompute its appeal, divide again. For the curious: appeal is e raised to the product's utility, and this rule is the logit model from From choices to value.

Step 2: Launch the big can and trace its buyers

Volt adds a 500ml can of the same Berry flavor. The survey gives it the same appeal as the 250ml: 2. The total is now 12. Recompute every share, then work out how much each option lost.

Now read the simulator's answer the way a brand manager would. Of the new can's sales, what percentage comes from each source?

Volt's line jumps from 20% to 33.3% of the market, and only a fifth of the new can's sales come from Volt's own 250ml. Hold that thought.

Step 3: Does that match how shoppers behave?

Look at the last column of your first table. Every option lost exactly the same fraction of its share: one sixth. That is not a coincidence of these numbers. It is built into the rule from step 1: a new product draws from everything on the shelf in proportion to its size, whatever the product is.

Now think about who actually picks up a 500ml Volt Berry. Mostly people who already like Volt Berry. Someone loyal to Rival A barely notices it.

Step 4: The usual fixes, and where they stop

Practitioners know about this problem. The classic version is the transport example in many choice courses: add a blue bus next to a red bus, and the model pulls riders off the train as readily as off the red bus. Two fixes are common.

Similarity corrections (randomized first choice in commercial simulators). These work well when two products are copies of each other. A 500ml can differs from the 250ml in size and price, so it is only partly a copy, and that is the case these corrections are least sure about.

Groups drawn in advance (nests, or a consumer decision tree). The analyst decides before the study that, say, all Volt cans compete with each other first. Two catches. You have to guess the answer you were trying to measure. And shoppers group the shelf in more than one way at once: by brand, by flavor, by size. A single tree has to pick one.

Step 5: Let the survey tell you who switches

The natural fix is to stop guessing. Your conjoint survey already shows shoppers choosing among products that share a brand, a flavor or a size. A model can measure, for each of those attributes, how much shoppers treat products that share it as substitutes. No tree drawn up front: brand, flavor and size each get their own measured pull, and they can all matter at once. With that, a new product takes most from the products shoppers see as closest to it, by however much the survey says.

A 20-product shelf used to take hours to fit this way, so tools kept the simple rule and patched it. Chosŏn fits a 20-product shelf in about two minutes. Step 6 is that shelf.

Step 6: Do it on a 20-product shelf

Every free account includes a 20-product energy-drink shelf with 1,000 survey respondents, already fitted. The survey was simulated, so the true answer is known and shown next to both models. In the app, a panel walks you through it: load the launch, see where its buyers come from under each model and the truth, read the switching, and write your recommendation to Volt's brand manager. Then download your summary.

Start step 6 in the app (free account; your answers so far come with you)