You are choosing between a cheaper product with fewer features and a pricier one with more. Asking customers “which features matter?” tends to produce a list where everything matters, because features are free to rate and prices are not. Conjoint analysis fixes that by making people trade features against each other, the way a shopper does.
It belongs in the market research process once the question is sharp, so start from the first step of the research process before you build a survey.
Key takeaways
- Conjoint analysis measures how much people value each product feature by making them choose between realistic combinations of features, which reveals the trade-offs they accept.
- The model estimates a part-worth utility for every attribute level, and attribute importance is each attribute’s range divided by the sum of all ranges.
- A market simulator adds up part-worths for competing products and converts them to preference shares, so you can test what-if changes such as a price rise.
- Importance depends on the levels you test, and simulated shares are relative preference among the options you defined, not a sales forecast.
How does conjoint analysis work?
A conjoint study breaks a product into attributes (the dimensions a buyer compares, such as price or battery life) and levels (the specific values each attribute can take, such as $79, $99 and $129). Respondents then see combinations of levels and react to them. From those reactions, a statistical model estimates a part-worth utility for every level: a number showing how much that level adds to, or subtracts from, the appeal of a product.
People rarely judge features one at a time; they judge the whole package. With enough package judgments, you can work backward to the value of each part.
What are the main types of conjoint analysis?
- Choice-based conjoint (CBC): respondents see a few product profiles side by side and pick the one they would buy, or none. It is the most common type because choosing mimics a real purchase.
- Traditional (full-profile) conjoint: respondents rate or rank complete product profiles, each showing every attribute.
- Adaptive conjoint: the survey adjusts the profiles it shows based on earlier answers, which shortens interviews for products with many attributes.
MaxDiff is often listed alongside these, but it is a different method. We compare the two below.
What does a choice-based conjoint task look like?
Each task is one screen. Suppose you are testing wireless headphones with three attributes: price, battery life and noise cancelling. A respondent might see three profiles and a “none of these” option:
| Option 1 | Option 2 | Option 3 | |
|---|---|---|---|
| Price | $79 | $99 | $129 |
| Battery life | 20 hours | 30 hours | 30 hours |
| Noise cancelling | No | Yes | Yes |
They pick one, then see a new screen with different combinations, often around ten of them. A design algorithm varies the levels so price is not always paired with the same battery life, which lets the model separate one attribute’s effect from another’s.
How do you read part-worth utilities and attribute importance?
Here is a part-worth table from a hypothetical headphone study. The numbers are illustrative, not real data. Utilities are zero-centered within each attribute, so they sum to 0 for that attribute.
| Attribute | Level | Part-worth utility | Range (highest minus lowest) | Importance |
|---|---|---|---|---|
| Price | $79 | +1.20 | ||
| $99 | +0.10 | |||
| $129 | -1.30 | 1.20 - (-1.30) = 2.50 | 55.6% | |
| Battery life | 10 hours | -0.60 | ||
| 20 hours | +0.10 | |||
| 30 hours | +0.50 | 0.50 - (-0.60) = 1.10 | 24.4% | |
| Noise cancelling | No | -0.45 | ||
| Yes | +0.45 | 0.45 - (-0.45) = 0.90 | 20.0% |
Attribute importance is each attribute’s range divided by the sum of all ranges. It answers “how much does this attribute swing a decision, given the levels we tested?”
- Calculate each range: 2.50 for price, 1.10 for battery, 0.90 for noise cancelling.
- Add the ranges: 2.50 + 1.10 + 0.90 = 4.50.
- Divide each range by 4.50: price 2.50 / 4.50 = 55.6%, battery 1.10 / 4.50 = 24.4%, noise cancelling 0.90 / 4.50 = 20.0%.
The three shares add to 100.0%. In this example, price moves choices most, but note the wording: importance depends on the levels you chose. If you had tested prices of $95, $99 and $105, price would look far less important, because you gave respondents almost nothing to react to.
A negative utility such as -1.30 for $129 does not mean people hate the headphones; it means that level pulls the total below this attribute’s average.
How does a market simulation turn utilities into shares?
Once you have utilities, you can build a simulator: pick a set of competing products, add up each one’s part-worths for the levels it carries, and estimate the share of respondents who would choose each. With choice-based models, the common approach converts a total utility U into a share using exp(U) divided by the sum of exp(U) across all products.
Take three products from the illustrative table above:
| Product | Price | Battery | Noise cancelling | Total utility | exp(U) | Share of preference |
|---|---|---|---|---|---|---|
| A | $99 | 30 hours | Yes | 0.10 + 0.50 + 0.45 = 1.05 | 2.858 | 48.4% |
| B | $79 | 20 hours | No | 1.20 + 0.10 - 0.45 = 0.85 | 2.340 | 39.6% |
| C | $129 | 30 hours | Yes | -1.30 + 0.50 + 0.45 = -0.35 | 0.705 | 11.9% |
The exp(U) values sum to 5.902, so A’s share is 2.858 / 5.902 = 48.4%. The three shares add to 99.9% because of rounding.
Now test a decision. If Product A moves to $129, its total utility becomes -1.30 + 0.50 + 0.45 = -0.35, the same as C. Its exp(U) falls to 0.705, the total becomes 0.705 + 2.340 + 0.705 = 3.750, and A’s share drops to 18.8%. A $30 price rise costs about 29.6 points of preference share (48.4 minus 18.8), and most of it goes to B. That is the kind of “what if” a simulator answers in seconds.
Treat the output as relative preference among the options you defined, not a sales forecast. It leaves out awareness, distribution and promotion.
What is the difference between conjoint and MaxDiff?
MaxDiff is sometimes described as a variation of conjoint, but it is a separate method. Both force trade-offs, yet they measure different things.
| Conjoint | MaxDiff | |
|---|---|---|
| What respondents judge | Whole products made of attribute levels | A list of single items (features, messages, claims) |
| Task | Choose the preferred profile | Pick the most and least important item from a small set |
| Output | Utility for each level, importance of each attribute, market simulation | A ranked score for each item |
| Best for | Pricing, product configuration, “what if” scenarios | Prioritizing a long list of features or messages |
If you need to know how much to charge for a feature, you need conjoint, since MaxDiff has no price-versus-feature structure. If you need to cut thirty candidate claims to five, MaxDiff analysis is faster and simpler.
How is conjoint analysis used in practice?
Teams use it to decide which features to build, to set price points and tiers, to test whether a brand name carries weight against features, and to compare their offer with competitors’ profiles. Every respondent has their own utilities, so you can also compare groups. That links naturally to segmentation: you may find a price-driven segment and a feature-driven one that a single average hides. Connecting findings across studies lets the next pricing question start from what you know.
What are the common conjoint design mistakes?
- Unrealistic combinations. A premium brand at the lowest price with every feature looks odd, and respondents notice. Prohibit impossible pairings, but sparingly, since every prohibition limits what the model can estimate.
- Too many attributes or levels. Past a handful of attributes, respondents start answering on one or two. Fewer, well-chosen attributes beat an exhaustive list.
- Levels with a narrow or arbitrary range. As the importance example shows, the levels you choose set the result. Cover the range the market actually spans.
- Levels that overlap or are vague. “Fast shipping” is not a level; “delivered in 2 days” is. Levels within an attribute must be mutually exclusive.
- Skipping the question first. A conjoint study answers a specific decision, such as “what price and feature set do we launch?” If you cannot state the decision, you cannot pick the attributes.
How do you analyze conjoint results?
Specialist conjoint software estimates the utilities and runs simulators. The rest is familiar survey analysis: checking data quality, weighting to the target population, cutting results by segment, and testing differences for significance. Halo Reports covers that stage, with crosstabs, banner tables, weighting, significance testing and charts built from your survey data.
Before you field anything, write down the one decision the study should inform and list the three to six attributes that decision depends on. If you cannot fit them on one page, narrow the question first, then build the levels.