Stop Asking Customers What They Want
Ask someone what matters in a laptop and they'll say battery life, performance, weight, screen, price. All of it. Ask them to rank those and they'll produce a ranking that feels right and predicts nothing.
This isn't people being unhelpful. Preferences are only well-defined against alternatives. "Do you care about battery life?" has no real answer. "Would you give up six hours of battery to save $300?" does — and it's the only kind of question whose answer you can build a product on.
The technique
Conjoint analysis presents people with realistic product profiles and makes them choose. No profile wins on everything, so every choice is a trade-off, and the pattern of what someone gives up reveals what they were protecting.
The method came out of mathematical psychology in the 1960s and was brought into marketing by Paul Green and Vithala Rao in 1971. It's now routine in the automotive, telecom, consumer electronics and pharmaceutical industries — anywhere a product is a bundle of features and somebody has to decide which ones are worth building.
The structure is three steps:
- Define attributes and levels. Battery life: 8 hours or 14. Screen: 13 inch or 15. Price: $999 or $1,299.
- Build profiles and ask people to choose. Combinations of those levels, presented in pairs, with no pair where one option dominates the other on everything — a dominated pair teaches you nothing.
- Fit a model. From the choices, estimate a part-worth for each level: how much that level contributed to options being picked. Because price is one of the attributes, the part-worths convert into money — you can read off what an extra six hours of battery is worth in dollars.
That third step is what makes conjoint different from a survey. You never asked how much battery life was worth. You derived it.
Try it on yourself
Eight choices. There is no right answer, and no profile is better than the other on every dimension.
Eight questions, and we work out what you actually value
No option wins on everything, so every choice is a trade-off. Make all eight and the pattern of what you gave up says more than any answer to “what matters to you?” would.
Choice 1 of 8 — which laptop would you buy?
Show the numbers
| Attribute | Times it was the trade-off | Times you took the better level | Share of your decision |
|---|---|---|---|
| Battery life | 0 | — | — |
| Screen | 0 | — | — |
| Weight | 0 | — | — |
| Price | 0 | — | — |
You never stated a preference. The ranking at the end came entirely out of what you were willing to trade away.
What you get out of it
The trade-offs, not the wish list. Everyone says they want everything. Conjoint tells you what gets sacrificed when something has to be, which is the only version of the question a product roadmap can act on.
A price for features. Because price is an attribute like any other, the model converts every other part-worth into money. "Customers will pay about $140 for the larger screen" is a sentence you can take to a pricing meeting. "Customers say the screen is important" is not.
Segments that differ. Averaging across everyone usually produces a product nobody wants. The interesting output is often that a third of the sample is intensely price-driven and a quarter would pay almost anything for weight — which is a tiering strategy, handed to you.
A market simulator. Once you have part-worths, you can predict how share moves when you change a feature or a price, including what happens when a competitor changes theirs. This is the part that earns the study its budget.
Where it goes wrong
Too many attributes. Respondents can hold about six in their head. Past that, they simplify — usually by fixating on price — and you get a study that concludes everyone is price-sensitive because you exhausted them.
Levels nobody believes. If one level is implausible, people don't trade against it, they discount the whole profile. Keep the range realistic.
Stated choices are still stated. Conjoint is much better than asking directly, but it's a survey, not a purchase. Hypothetical price sensitivity is famously milder than the real thing. Treat the output as a strong prior to validate in market, not as a measurement.
Attributes you forgot. The study can only weigh what you put in it. If the real driver is brand, or delivery time, or whether it integrates with something the customer already owns, and you didn't include it, the study will confidently apportion its effect among the attributes you did include.
The shorter version
You don't have to run a formal study to get most of the value. The habit is the thing: stop asking people what they want and start making them choose.
Put two real options in front of a customer. Make one better on the dimension you're unsure about and worse on another. Watch which they take, and how long they hesitate. Do that twenty times and you'll know more than a hundred survey responses would tell you — because you asked the only question preferences actually answer.