Conjoint Analysis: Model Product Trade-Offs Without Mistaking Them for Demand

Suppose a pricing team receives a result saying that buyers are willing to pay 20 more for an advanced feature. The arithmetic may be correct and the decision may still be wrong. That number came from choices among particular packages, shown to particular respondents, across a particular price range. It did not include the sales conversation, procurement threshold, implementation cost, competitor response, or the possibility that the buyer never seriously considered purchasing.

That distinction is the useful core of conjoint analysis. It is a stated-preference method for learning how people trade one product characteristic against another when they cannot have everything at once. Instead of asking whether buyers value an integration, for example, a study might ask them to choose between a lower-priced package with basic integrations and a higher-priced package with broader coverage, a different support model, and a longer contract.

The result can sharpen a product or pricing hypothesis. It cannot turn a hypothetical choice into observed demand. Used well, conjoint analysis tells a B2B team which trade-offs deserve a real commercial test and where the answer is sensitive to the world the researchers designed.

What conjoint analysis actually estimates#

A conjoint study describes an offer through attributes and levels. An attribute is a dimension such as deployment model, support response, contract length, capacity, or price; its levels are the alternatives shown for that dimension. Respondents evaluate profiles built from different combinations of those levels. Depending on the study, they choose one profile, rate profiles, or rank them. A statistical model then estimates how changes in the presented levels are associated with the responses.

This setup solves a problem that direct questions often leave untouched. A buyer can sincerely say that security, service, flexibility, and a low price are all important. That does not tell a product team what the buyer would give up when stronger security raises the price or when flexible terms remove a volume discount. Conjoint makes the sacrifice visible because several characteristics move together. The ISPOR good-practice checklist treats the research question, attributes and levels, task construction, experimental design, instrument, collection plan, analysis, and reporting as connected parts of the study—not interchangeable setup details.

Utilities are relative signals, not units of customer value#

The fitted coefficients are often called utilities or part-worths. They express preference on the model’s scale, relative to the coding and reference levels used. A positive coefficient does not mean a feature creates positive margin, and a value twice as large does not automatically mean twice as much business value. Coefficients from separately fitted models are not inherently comparable either.

In a simple model with a linear price term, the utility of an alternative can be written as:

Utility = β₁x₁ + β₂x₂ + … + βₖxₖ + βprice × Price

When the coding and model make the ratio meaningful, marginal willingness to pay for an attribute change is approximately:

Marginal WTP = −(change in the attribute coefficient ÷ price coefficient)

Peer-reviewed conjoint research has used this coefficient-ratio approach to estimate willingness to pay, with uncertainty calculated around the estimates rather than inferred from the point values alone. For an explicitly illustrative calculation, suppose the coefficient changes by 0.60 and the price coefficient is −0.03 per price unit. The implied marginal WTP is 20 price units.

That means respondents traded the specified feature change against the specified price variation in a way consistent with 20 units under this model. It does not mean the list price can safely rise by 20. The estimate can move when the tested price range moves: in an Office for National Statistics split-sample study, two groups saw studies differing in their price ranges, and the group shown the higher range consistently produced higher median WTP estimates. The ONS cautioned that the results also depended on study design, sample composition, utility estimation, and the WTP calculation method.

Treat willingness to pay as a conditional trade-off estimate, never as a recommended price.

Preference simulations remain inside the designed market#

A market simulator takes estimated utilities and applies them to a scenario: package A at one price, package B at another, perhaps alongside a competitor or a “none” option. It can show how modeled choice changes when one level changes. This is useful for comparing scenarios under the same assumptions.

The percentage output is usually a share of preference, not a sales forecast. Even the vendor documentation for a commonly used choice-based conjoint simulator warns that simulated preference omits forces such as awareness, distribution, time in market, stock availability, advertising, and sales-force effectiveness. In B2B markets, the missing mechanics can also include procurement rules, security review, budget authority, discounting, switching costs, and implementation capacity. The model knows only the attributes represented in its profiles.

That does not make hypothetical choice evidence worthless. A field validation published in PNAS compared survey estimates with closely corresponding real-world naturalization votes in Switzerland; the paired conjoint design came closest to the behavioral benchmark, with estimates on average within two percentage points of the benchmark effects. The same validation study also found meaningful performance differences among survey designs. It supports a narrower conclusion than “conjoint predicts behavior”: close alignment is possible when the population, information, attributes, and decision setting correspond closely. External validity has to be demonstrated for the use at hand.

Design the study backward from the commercial decision#

The hardest part of conjoint analysis is usually not fitting a model. It is building a small artificial market that preserves the trade-off the business genuinely faces. If the study asks a cleaner or easier question than the real decision, an efficient design can produce a precise answer to the wrong problem.

1. Name the decision that can change. State the action before designing the survey: choose which two package concepts enter a sales test, decide whether annual commitment belongs in the base offer, or identify a price corridor for further validation. “Understand customer preferences” is too loose because almost any output could appear relevant after the fact.

2. Represent the people who make the choice. In a B2B purchase, the user, technical evaluator, economic buyer, and procurement team may respond to different attributes. Decide whose preference the model is intended to represent and recruit people who plausibly hold that role in the target account type, geography, and use case. A large sample of convenient respondents cannot repair a mismatched buying role.

3. Derive attributes from the buying situation. Interviews, win-loss records, sales objections, product constraints, and competitor offers can reveal which differences buyers actually notice. Convert those differences into attributes only when respondents can understand them independently. “Enterprise readiness,” for example, may hide separate judgments about identity management, audit logs, deployment, and support. A vague umbrella attribute invites each respondent to answer a different question.

4. Set levels that are credible and actionable. Each level needs to describe a state the business could provide or a competitor could plausibly offer. The range must be wide enough to reveal a trade-off but not so extreme that one option becomes a joke. Restrictions may be needed to prevent impossible combinations, yet restrictions also reduce independent variation. The ISPOR experimental-design guidance explains that a design must vary relevant levels enough to identify the intended effects and that excluding implausible combinations can introduce correlation and imbalance.

5. Specify the model before optimizing the tasks. Decide whether the study must estimate only main effects or also interactions, whether alternatives are labeled, how price enters the model, how a status quo or opt-out is handled, and which segment comparisons are planned. An interaction between service level and contract type cannot be recovered merely because both attributes appeared somewhere in the questionnaire. The design must contain the combinations needed to distinguish it.

6. Pretest interpretation, not just survey mechanics. Put draft tasks in front of people from the target population and ask what they thought each level meant, why they chose, which combinations felt implausible, and whether they ignored an attribute. A published guide to pretesting discrete-choice experiments separates content, presentation, comprehension, and preference elicitation because a survey can function technically while respondents misunderstand the choice. In the guide’s applied example, pretesting led researchers to reword, reorder, add, and remove attributes and levels.

7. Plan sample size around the estimand. The number of attributes, levels, tasks, alternatives, planned interactions, expected heterogeneity, recruitment design, and required precision all matter. More tasks can add statistical information, but difficult or repetitive tasks can also create fatigue and simplifying behavior. The experimental-design guidance describes this as a balance between statistical efficiency and response efficiency. Run design-specific precision or power simulations where possible, and reserve enough observations for any segment result the decision truly requires.

8. Define the commercial validation before fielding. Decide which result would trigger a real test, what will be measured, and what would contradict the conjoint finding. That prevents the team from treating whichever attractive simulation appears later as the study’s mandate.

This sequence also exposes when conjoint is the wrong tool. If the decision turns on one reaction to one fixed concept, qualitative interviews or a concept test may answer it more directly. If the team needs actual conversion, renewal, discount, or margin behavior, it needs observed market evidence. Conjoint earns its cost when several implementable attributes must be traded together and the business can act on the distinctions.

Turn the model into a bounded commercial test#

The analysis should preserve the distance between a model result and a business result. Four common outputs sit at different points along that distance:

A defensible readout therefore carries its conditions with it. It identifies who responded and how they were recruited; shows the attributes, levels, task format, and price range; states the coding and model; reports uncertainty; and distinguishes planned analyses from exploratory cuts. It also tests sensitivity to plausible alternatives. If a WTP estimate changes sharply when the price range or functional form changes, that instability is part of the result.

Then the business can narrow the claim. Instead of “buyers will pay more for advanced controls,” a useful handoff might say that eligible respondents preferred a specified controls package over the baseline within the tested alternatives, with an estimated price trade-off that remained within a stated interval under the planned specifications. The next step is to offer the two feasible packages to appropriately qualified buyers and observe acceptance, objections, discounting, time to close, implementation burden, and contribution economics.

No single test supplies the whole answer. Sales outcomes may reflect rep behavior or account mix; product usage may reveal adoption without price acceptance; interviews may expose why a level was misunderstood without measuring the size of a preference. The point is to join different evidence types without pretending they are interchangeable. Conjoint contributes controlled variation across complex offers. The market contributes consequences.

The next move is smaller than a launch#

Before commissioning a study, write one sentence that names the commercial choice and the evidence that must follow the survey. If that sentence cannot be made specific, the study is premature.

When the sentence is clear, conjoint analysis can reduce a messy packaging or pricing debate to a testable hypothesis. The responsible next move is not a universal rollout. It is the smallest observed commercial test capable of proving the model useful—or proving it wrong.

Frequently asked questions#