Conjoint Analysis: Definition, Methods & Worked Example
Three conjoint approaches, six study steps and a bottle worksheet showing how features and price become preference shares.
Conjoint analysis estimates how features and price affect preference by asking buyers to evaluate planned offers. It compares complete product configurations rather than rating features separately. A study needs a designed set of product evaluations and an estimated model; the worked example below uses assumed scores to explain the arithmetic. Our market research overview helps place this method alongside other ways to answer a business question.
Types of conjoint analysis
The names describe different parts of a study. Choice-based describes the response task; adaptive describes how later questions change; full-profile describes how much of each product is shown. A study can use more than one of these labels.
Approach | What respondents do and when it fits |
|---|---|
Choice-based conjoint (CBC) | Choose among product offers in repeated tasks. Use it for tradeoffs between complete configurations, with a no-choice option when rejection is realistic. |
Adaptive conjoint | Later tasks respond to earlier answers. Adaptive choice-based conjoint (ACBC) is one version; traditional adaptive conjoint (ACA) is another. Ask which task and estimation approach the proposal uses. |
Evaluate profiles showing every studied attribute. Traditional versions often ask for ratings or rankings; choice-based designs can also show full profiles. Keep the attribute set manageable. |
Sawtooth explains traditional adaptive conjoint and adaptive choice-based conjoint separately. Choose the design around the product tradeoff and what buyers can reasonably evaluate.
How to run a conjoint study in 6 steps
- Define the product decision, buying occasion and alternatives. State whether the study must choose a configuration, examine price tradeoffs or support a portfolio decision.
- Choose attributes and realistic levels. Keep descriptions specific, record impossible combinations and check that the planned variations can separate the effects you need.
- Design the repeated tasks and recruit the relevant buyers. Choose the response format, no-choice treatment, sample design and estimation method before collection.
- Pilot the full questionnaire, then collect answers. Check comprehension, task burden and data quality; retain the exact offers and response records.
- Estimate the model and check its predictions. Report uncertainty, coding and performance on tasks kept out of fitting; inspect differences relevant to the product decision.
- Compare products the business can make and sell. Test sensitivity to price and competitors, then carry the chosen configuration into cost, product and purchase checks.
The bottle example below explains the calculation with assumed coefficients. It illustrates the model output; it does not replace the design, data collection and estimation steps.
Example: choosing bottle features and price
This hypothetical bottle worksheet uses assumed coefficients, not collected buyer answers or a fitted model. It shows how the calculation works before considering a real study.
The bottle is intended for desk use and commuting. Keep the buying occasion fixed while comparing alternatives, so the result tells you which configuration to take into a product trial.
Attribute | Offer details |
|---|---|
Capacity | Levels in the example: 350 ml; 500 ml Design consideration: Both must be commercially feasible |
Lid | Levels in the example: Standard; travel lid Design consideration: Describe actual function, not an unverified performance guarantee |
Price | Levels in the example: $20; $25; $30 Design consideration: Same currency, unit and channel |
The bottle team can change capacity, lid and shelf price. Those are meaningful dimensions because buyers can recognize the levels and the business can deliver them. “Premium, sustainable and easy-clean” would bundle several mechanisms into one attribute and leave the recommendation unclear about which change matters.
Keep the buying situation the same. A commuting bottle may face different alternatives from a desk bottle. If channel, tax or delivery changes with price, record that as part of the offer. Otherwise, the model may attribute that unrecorded difference to price.
Use levels that span the credible decision space. Too narrow a range can miss the tradeoff the team needs; implausible extremes can produce choices unrelated to the intended launch. The tested range also bounds interpretation. A model learned from $20–$30 offers has no automatic evidence for a $60 version.
Exclude combinations for a product reason, such as a lid that cannot fit. But check whether restrictions prevent separate estimates. If the travel lid appears only on the larger bottle, how will the model distinguish the lid from capacity? Ask the researcher to show what the planned variations reveal.
Choice-based conjoint tasks
This example uses choice-based conjoint: respondents select among complete offers. It is a discrete-choice task. If a proposal contrasts conjoint with “discrete choice,” ask which task, planned variations and analysis it means; the labels alone do not specify different studies.
Offer feature | Offer details |
|---|---|
Capacity | Offer A: 350 ml Offer B: 500 ml Offer C: 500 ml |
Lid | Offer A: Standard Offer B: Travel Offer C: Standard |
Price | Offer A: $20 Offer B: $30 Offer C: $25 |
Ask which offer the person would choose on a relevant next purchase occasion. Include “none of these” when rejection is realistic. Forced choice records the least unsuitable offer even if the person would buy nothing, so the presence or absence of none is part of the inference.
The A/B/C display illustrates one unanswered choice task. It is not enough to fit a conjoint model; you need planned variation across tasks.
Show enough information to choose, with comparable images and explanations across offers. Describe the travel lid’s actual function without assuming leakproof performance. If one card has a photo and another only words, presentation may explain the choice unless that difference is deliberately tested.
One attractive-looking task is not the design. Capacity, lid and price need planned variation across tasks and respondents so their effects can be estimated. The single A/B/C task here cannot recover every coefficient: several product differences occur together. Sawtooth’s choice-based conjoint guidance describes the attribute/level framework.
Try the full sequence before launch. People may understand the first task and lose attention later. Find where choices become difficult and check for answers driven by position. The design needs enough variation to estimate effects without asking more than the audience can reasonably complete.
Calculating utility and preference shares
Utility is the model’s preference score for an offer. Use the bottle choice worksheet below to reproduce the comparison in a spreadsheet. Every assumed coefficient is visible. The worksheet does not estimate a model or contain respondent data.
The worksheet uses the same capacity, lid and price attributes as the task above. The coefficients are assumed rather than fitted: offer constant 2; 500 ml versus 350 ml adds 0.5; travel versus standard lid adds 1; price contributes −0.1 per dollar. None has utility 0. This connection lets you trace each card into the calculation.
Task alternative | Offer details |
|---|---|
A: 350 ml, standard lid, $20 | Assumed utility calculation: 2+0+0−.1×20 Utility: 0 Illustrative share: 21.5% |
B: 500 ml, travel lid, $30 | Assumed utility calculation: 2+.5+1−.1×30 Utility: 0.5 Illustrative share: 35.5% |
C: 500 ml, standard lid, $25 | Assumed utility calculation: 2+.5+0−.1×25 Utility: 0 Illustrative share: 21.5% |
None | Assumed utility calculation: Fixed reference Utility: 0 Illustrative share: 21.5% |
The logit calculation is exp(utility)/sum(exp(all utilities)). Lower only B’s price from $30 to $25 and its utility rises to 1: the illustrative shares become 17.5% A, 47.5% B, 17.5% C and 17.5% none. The utility difference follows from the explicit price coefficient; it is not measured willingness to pay. Inspect the worksheet’s assumed cards and coefficients before changing prices.
A real study uses planned choice tasks and an estimation method to calculate scores, uncertainty and checks of model fit. One task with assumed coefficients cannot do that. Check costs and validation before choosing an offer to sell.
Those percentages are illustrative shares of preference, not market shares. They omit awareness, distribution, stock and other market conditions. Sawtooth's simulator guidance explains this important boundary.
Utility is a preference score within this model. A’s zero does not mean “no value.” It follows from the chosen coding and constant. B’s score of 0.5 gives it a higher exponential weight than A’s 0. Every alternative, including none, contributes to the denominator.
The none option affects the arithmetic. Its assumed zero utility gives it the same baseline share as A and C. Changing that assumption changes shares even with identical product cards. A real study needs evidence for how it represents rejection.
Price tradeoffs in the assumed bottle model
The assumed travel-lid contribution is 1 utility unit and the price coefficient is −0.1 per dollar. Holding the rest fixed, a $10 increase offsets that utility addition in the supplied arithmetic. This demonstrates a marginal tradeoff under a constant linear price term. It is not a measured $10 willingness-to-pay premium for the lid.
When an analyst converts a score to dollars, ask how price was modeled. Is the tradeoff inside the tested range? How much uncertainty or buyer-group variation is there? A convenient dollar figure can hide those assumptions.
Shares are also relative to the displayed set. The $25 B scenario competes against A at $20, C at $25 and none. If C is removed or a new competitor added, the denominator and question change. A larger share after removing a rival does not establish that buyers value B more.
Report preference scores, shares within a simulated choice set and actual market outcomes separately. Awareness, distribution, stock, promotion and purchase difficulty are outside this calculation. A market model may add that evidence; this worksheet does not.
Do not move a coefficient into a different product model or coding scheme. Constants, reference levels and price units work together. Reading one coefficient without them can reverse or exaggerate its apparent meaning.
Attribute importance and the tested level range
A level’s score is relative to its coding reference. Here, 500 ml adds 0.5 compared with 350 ml; the travel lid adds 1 compared with the standard lid. Those differences apply within this assumed model. Its chosen zero does not measure a product’s absolute value.
A simple range summary illustrates why attribute importance depends on the levels. Capacity spans 0.5 utility, lid spans 1 and price spans 1 across $20–$30. Dividing those ranges by their total 2.5 gives 20%, 40% and 40%. This is arithmetic from assumed coefficients, not estimated customer importance. It compares the ranges we chose to test.
Now narrow the price range to $20–$25 without changing the coefficient. Its utility range becomes 0.5; the same range summary becomes 25% capacity, 50% lid and 25% price. Buyers did not change. The design range changed. This is why a reported importance percentage should travel with the attribute levels, rather than become a universal statement that one feature matters twice as much.
Before comparing importance summaries, check the coding and treatment of price. A nonlinear price relationship or different buyer-group sensitivity changes the meaning. Ask which ranking supports the product choice and how uncertain it is.
Choice-set sensitivity
The ratio of B’s share to A’s is exp(.5)/exp(0), about 1.65. The common denominator cancels. Removing C changes their individual shares but leaves that ratio unchanged under this formula. That is a property of the assumed calculation, not evidence about how real shoppers respond when an alternative disappears.
This matters when alternatives are close substitutes or the choice set changes substantially. A simulator may produce a neat answer for conditions its model does not capture. Ask whether the new configuration and competing offers were checked with suitable data kept out of model fitting.
Save the exact choice set, change one feasible input and explain which assumption drives the result. Adding a speculative competitor does not validate the resulting demand estimate. Use the worked calculation to understand that dependency before making a real decision.
Product costs and configuration constraints
Begin with products the business can make and sell. Keep each candidate’s simulated preference, competing set, cost and channel conditions together. A larger bottle can appeal while needing more shelf space or tooling. A popular feature may be uneconomic at the required price.
State the objective. A penetration strategy, contribution objective and premium portfolio strategy can choose different offers from the same preference model. The model supplies the tradeoff; the business supplies what it is trying to optimize. A single “highest share” badge collapses those decisions.
Check buyer groups when the result could change which product you develop. Assess uncertainty and validation before creating a product tier for each group’s apparent winner. More tiers add operating and communication costs the survey does not measure.
Check whether reasonable changes in estimates or alternatives change the recommendation. Then test predictions on choice tasks not used to fit the model. Matching training choices helps diagnose a model; held-out tasks check how it predicts new choices.
The next action could be one prototype and a paid test at the intended shelf price. If every winning configuration requires an infeasible discount, the completed recommendation is to revise the product or economics. Both outcomes are more useful than a scenario comparison whose largest percentage has no owner.
Synthetic conjoint analysis: check response source and fit
Choice answers may come from people or a simulated population. The task can look the same, but the evidence differs. Check how buyers are represented, how answers were collected or generated and how they were used to estimate the model.
For a simulated study, ask for the audience data, actual choice records, tasks and answer counts used to fit the model. Check the estimation method and its predictions for tasks excluded from fitting. Plausible coefficients alone do not prove a model was estimated. The bottle coefficients teach the calculation; a commissioned study needs its own records.
Ask the supplier to show one complete chain: task card, eligible response, fitted output, configuration comparison and proposed action. Keep the no-choice treatment and feasible combinations visible. If the suggested travel-lid bottle wins only after a price reduction, take that configuration into the relevant cost comparison before recommending it.
Our task-level validation comparisons illustrate why evidence should stay tied to a particular question and measure. Those results concern the displayed tests; evaluating a new conjoint specification requires evidence relevant to its population, design and intended recommendation.
The report should show which choice data produced the estimate, whether reasonable uncertainty changes the recommendation and what happens when prices or alternatives change. Keep fitted preferences, shares within the chosen set and separately supported market predictions distinct.
What a conjoint report should include
Ask for the audience definition, all tasks, coding, restrictions, quality rules, estimation method and held-out checks. Save the exact scenario and assumptions so another reviewer can reproduce the chosen configuration.
Ask for a report example that reconciles product, price and feasibility. It should name what to develop, why it wins under the stated objective, what assumption matters most and which product or market evidence remains necessary. A model can produce precise decimals without resolving those questions.
Sapien can compare simulated complete offers and explain buyer-group differences. Before you commission a study, agree with us whether it needs a full conjoint design or a simpler comparison of complete offers, and how the result will be checked. If the task is only benefit priority, our MaxDiff and conjoint guide is the more direct starting point.
Copy the data: bottle choice worksheet
In a blank spreadsheet, put the following tab-separated cells in A1. Each line is a row. Keep the header and four alternatives together; the None row stays in the denominator.
Alternative Capacity_ml Lid Price
A 350 standard 20
B 500 travel 30
C 500 standard 25
None
Add the headers Utility in E1, Weight in F1 and Share in G1. Enter the formulas below. They use a decimal point and comma-separated function arguments; spreadsheet locale settings can use different separators.
Cell or range | Enter or do |
|---|---|
E2 | =2+IF(B2=500,0.5,0)+IF(C2="travel",1,0)-0.1*D2 |
E3:E4 | Fill E2 down through E4 |
E5 | =0 |
F2 | =EXP(E2) |
F3:F5 | Fill F2 down through F5 |
G2 | =F2/SUM($F$2:$F$5) |
G3:G5 | Fill G2 down through G5; format G2:G5 as percentages with one decimal place |
The baseline reproduces 21.5%, 35.5%, 21.5% and 21.5%. Change only D3 from 30 to 25: the worksheet recalculates 17.5%, 47.5%, 17.5% and 17.5%. Restore D3 to 30 before trying another change. Copy the sheet for each scenario and record which input changed, so the recommendation can be traced back to the exact alternatives.
These formulas supply the same assumed offer constant, capacity effect, lid effect, linear price effect and fixed None reference as the example. Changing the capacity or lid coding requires updating the formula deliberately. Keep price comparisons within the credible $20–$30 study range; a spreadsheet will calculate outside it without supplying evidence there.
Further reading
If you are choosing a price or bundle, start with our pricing research guide. It helps separate the commercial choice from the technique used to investigate it.
For the bottle, a Sapien study discussion can start with the configurations your team can make and the shelf prices that cover their costs. The useful return is a configuration to prototype, the assumption most likely to change that choice and the purchase evidence still needed.
