Operations
What Is Conjoint Analysis? A Plain-English Guide for Consumer Researchers
Conjoint analysis in consumer research turns choices between bundles into decision weights, then shows you when those weights are worth acting on.
What to take away
- Conjoint analysis measures the trade-offs people make when they choose, not the ratings they give.
- Choice-based conjoint is the common form in US consumer work because the task looks like a real purchase.
- Attribute importance shows what drove the choice. Holdout share error shows whether the model is worth trusting.
- Act on the shares when holdout share error is 5 percentage points or less. Above 8 points, use the study for ranking only.
- No design speaks for a market you never sampled.
Conjoint analysis is a survey method that shows people competing product bundles and records the one they pick. The output is not a preference score but a set of weights showing how much each feature, level or price point contributed to the choice. Choice-based conjoint analysis became the default in American consumer work because the task resembles a shelf, a configurator or a checkout page. The family of conjoint methods runs from simple rankings to full choice exercises, and the differences matter when you read a supplier deck.
The metric worth tracking
Holdout share error is the number to watch. It is the average absolute gap, in percentage points, between the shares a model predicts for a set of choice tasks and the shares those tasks produce when a fresh group answers them. Three points of error means the arithmetic matches behavior. Twelve points means it does not.
Attribute importance is the second number worth naming. It is each attribute's share of total decision weight, written as a percentage. The shares sum to 100, so an attribute at 40 percent carries twice the weight of one at 20 percent, inside that design only. The weights are only as good as the questionnaire under them, which is why market research surveys is worth a look before you approve a design.
How to read the numbers
| Metric | What it measures | How to read it |
|---|---|---|
| Holdout share error | Gap between predicted and observed choice shares | Model fit, lower is better |
| Attribute importance | Share of decision weight per attribute | Relative ranking, sums to 100 |
| Part-worth spread | Distance between the best and worst level of one attribute | Sensitivity inside one attribute |
A share that moves two points between two runs of the same design is noise, not a finding.
Read the error first, then the importance ranking, then the spreads. If the error is comfortable, the ranking is usable even when the absolute shares look generous, because a closed choice set inflates them. Turning utilities into a ranking is an analysis job of its own, and market research analysis covers the steps between an output file and a decision.
What conjoint shares cannot tell you
Conjoint shares are inflated by design. The choice set holds only your bundles, so real competitors, shelf position, stockouts and habit are absent, and every feature in the study looks more important than it will be in a store. Panel make-up matters too, since the structure of online panels decides who ends up answering. A low holdout error is not proof of a good study. It says the model reproduces the tasks inside it, and a study that sampled the wrong people can fit its own data neatly.
The importance figures are relative. Remove one attribute and every remaining share rises. That makes importance a poor number for comparing two studies with different attribute lists.
The model assumes respondents read the whole bundle and weigh it. Many do not.
Attribution and its limits
A conjoint result is an estimate inside a controlled exercise. It is not proof that changing a feature will move sales. Distribution, promotion and seasonality sit between the test and the till, and teams that read a share shift as a forecast tend to over-claim. Quantitative research from the field sets out how much weight a modeled result carries before it needs market evidence behind it.
Example
The figures below are illustrative rather than measured. A US beverage team runs a choice-based study with 900 respondents. The design carries four attributes: flavor, pack size, sweetener and a front-label claim. Holdout error comes back at 4 points, so the shares get used. Flavor carries 38 percent of importance, pack size 27, sweetener 21 and the claim 14.
The team drops the claim from the next wave and spends the room on flavor variants. Fieldwork is where these studies are won or lost, and primary market research covers what collecting your own data asks of a team.
When to stop measuring and decide
Stop when three conditions hold at once: holdout error at or below 5 points, sample coverage of the segments you sell into, and the top two attributes keeping their rank across two independent runs. If a second wave misses those conditions, a third rarely fixes the design. Sample size is the usual culprit, and determination of sample size depends on confidence level and margin of error.
Common questions
Is choice-based conjoint the same as conjoint analysis? No. Choice-based conjoint is one design within the family, asking people to pick from bundles rather than rank or rate them. It dominates consumer feature and pricing work.
How many attributes can a study handle? Practical consumer designs often stop at six to eight attributes, because attention drops as the task grows. More attributes mean more respondents.
Do I need a national US sample? Match the sample to the market you sell in. A national panel is no substitute for shoppers in the three metros where a launch actually happens.
Can a conjoint study replace a price test? No. Conjoint estimates how price levels trade against features. It does not measure what people do at a real checkout with a real budget.
