What is Conjoint Analysis?
Written by Arnon Shimoni
✓ Expert
Last updated on:
Conjoint analysis is a survey-based statistical technique that measures how much buyers value individual product attributes by forcing them to choose between complete bundles rather than rating features one at a time. Because price is one of the attributes, the output tells you what each feature is worth in currency, which makes conjoint the only common pricing method that can design packaging and set a price level in the same study. Paul Green at Wharton developed the marketing application, building on conjoint measurement theory, with later contributions from V. Srinivasan at Stanford and Jordan Louviere on the choice-based variant.
Field | Detail |
|---|---|
Also known as | Conjoint, trade-off analysis, choice modelling, discrete choice experiment |
Developed by | Paul E. Green (Wharton); choice-based variant from Jordan Louviere |
Question format | Repeated forced choices between product profiles with varying attributes |
Output | Part-worth utilities per attribute level, price sensitivity, simulated market share |
Typical sample | 300-800 respondents (practitioner convention; more for many segments) |
Attributes it handles well | 4-8 in choice-based conjoint; up to 20-25 in adaptive formats |
Best for | Packaging, tiering, bundling, feature gating, willingness to pay per feature |
Weakest for | Brand positioning, emotionally loaded attributes, very small B2B populations |
Cost | The most expensive of the common methods, by a wide margin |
Why forced choice beats asking directly
Ask a buyer to rate the importance of SSO, audit logs, a dedicated CSM, and a 99.99% SLA and they'll rate all four important. Of course they will. Nothing is being taken away from them.
Conjoint takes things away. It presents two or three complete product configurations, each with a different mix of features and a different price, and asks which one they'd buy. Doing that fifteen or twenty times per respondent, with the configurations varied systematically, lets you back out how much each attribute level contributed to the choices. Those numbers are part-worth utilities.
Because price is one of the attributes, utilities convert into money. If moving from "email support" to "dedicated CSM" is worth 0.8 utils, and 1 util is worth €140 a month on your price attribute, the CSM is worth about €112 a month to that segment. That's a number you can put in a tier.
This is why conjoint is the method to reach for when the question is packaging rather than price level. Van Westendorp and Gabor-Granger price the whole product as one object. Conjoint prices its parts.
Types of conjoint
Type | How it works | When to use |
|---|---|---|
Choice-based (CBC) | Respondent picks one profile from a set of 2-4, repeatedly | The default. Closest to a real buying decision |
Adaptive (ACA/ACBC) | The survey adapts to earlier answers, narrowing on what matters to that respondent | Many attributes (10+), or a heterogeneous audience |
Ranking-based | Respondent ranks a set of profiles | Small attribute sets, older studies |
Rating-based | Respondent rates each profile on a scale | Rarely used now. Less behaviourally realistic than choice |
Menu-based | Respondent builds their own bundle from priced options | Configurable products, add-ons, à la carte pricing |
MaxDiff | Best and worst from a set, no price attribute | Feature prioritisation before you design packages. Technically adjacent rather than true conjoint |
For most SaaS and AI pricing work, choice-based conjoint is the answer, with adaptive worth considering when you have a long feature list and a diverse buyer base. Menu-based conjoint deserves more attention than it gets in software, since it directly mirrors how add-on and modular pricing actually gets bought.
Designing the study
The design phase decides the quality of the result. Everything after it is arithmetic.
Pick attributes that are decisions you can act on. An attribute is worth including if a plausible answer would change what you ship. "Support level" is actionable. "Innovativeness" is not.
Keep the count honest. CBC handles 4-8 attributes comfortably. Past that, respondents start simplifying (picking on price alone, or on one feature they care about) and your utilities describe their coping strategy rather than their preferences. Adaptive formats push the ceiling higher by narrowing per respondent.
Levels must be unambiguous, mutually exclusive, and plausible. "Fast support" is ambiguous. "4-hour response SLA" is not. And every level has to be believable: a €10 enterprise tier tells you nothing except that respondents noticed the free lunch.
Set the price range wider than you expect. If your price levels span €100-€300 and true willingness to pay sits at €400, the study can't see it. Anchor the range on a prior Van Westendorp or on benchmarking.
Include a "none of these" option. Without it, respondents must pick something, and you lose the ability to measure whether anyone would buy at all. In B2B, where the real competitor is usually "do nothing," this option carries a lot of information.
Screen for buyers. The classic B2B failure: surveying enthusiastic users who have no budget. Their utilities are real preferences and they don't predict purchases.
Reading the output
Three things come out of a conjoint study, in increasing order of usefulness.
Part-worth utilities. The raw scores per attribute level. Useful for internal analysis, hard to present to an exec team.
Attribute importance. The share of total utility range each attribute accounts for. "Price explains 34% of the decision, integrations 22%, support 9%." This is the slide everyone remembers, and it's the one most often over-read: importance depends entirely on the levels you chose to test. Widen the price range and price gets more important. That's an artifact of your design.
A market simulator. The actual deliverable. Estimation (typically hierarchical Bayes, giving individual-level utilities) lets you build a model where you specify a set of competing products and get predicted preference shares. Now you can ask real questions: what happens to share if we move the Pro tier from €200 to €240? What if we pull SSO down into Pro? What if a competitor launches at €150?
The simulator is where conjoint earns its cost. A study that ends at an importance chart wasted most of the money.
(Sample output of conjoint analysis. By Happybunny95 - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=52403562)
Conjoint for usage-based and hybrid pricing
Conjoint handles multi-dimensional pricing better than any other survey method, which makes it the natural fit for hybrid models. It still needs care.
Model the price dimensions as separate attributes. A hybrid product has a platform fee, an included allowance, an overage rate, and often a commit discount. Each becomes its own attribute with its own levels. Now the simulator can answer "would buyers rather have a lower base fee or more included credits," which is the actual packaging question in most AI deals.
Watch the interaction between allowance and overage. These two aren't independent in the buyer's head. A generous allowance with a punitive overage rate reads as a trap, and main-effects-only models miss that entirely. If you care about the answer, design for interaction effects and accept the larger sample it requires.
Use outcomes as levels, not technical units. Levels like "2 million tokens included" mean nothing to a buyer who doesn't know their token consumption. "Enough for roughly 5,000 support conversations" is a level they can evaluate. Convert to the metered unit after the study. Same principle as in AI agent pricing.
Include predictability as an attribute. Buyers of consumption products pay real money to avoid bill shock, and spend caps, alerts, and committed-spend structures have measurable utility. Most studies leave this out and then can't explain why customers chose the "worse" deal. It's why prepaid credits win deals that a pure pay-as-you-go rate card should have won on price.
That last point is my favourite finding in this space, and it's the one that most often surprises the people commissioning the study.
Limitations
It's expensive and slow. Agency-run CBC with a proper sample typically means 6-10 weeks and a five-figure budget. That's the real reason most companies don't run it.
Small B2B populations break it. If your total addressable buyer population is 800 people, you're not getting a 400-person sample. Enterprise pricing often has to run on win/loss analysis and structured interviews instead.
Respondents oversimplify. Give someone 12 attributes and they'll latch onto two. The utilities you estimate then describe a simplification strategy.
It's still stated preference. No money changes hands. Absolute willingness to pay from conjoint runs high, though the relative values (this feature is worth twice that one) hold up considerably better than the absolute ones.
It's poor at emotional and brand attributes. Concrete, specifiable attributes work. "Brand trust" and similar get systematically distorted.
It's a snapshot. The simulator assumes competitors hold still. They don't.
Related terms
Frequently asked questions
What does conjoint analysis measure?
How much buyers value each attribute of a product, in units you can convert to money, derived from the tradeoffs they make when choosing between complete bundles.
What is a part-worth utility?
The numerical value a respondent places on a specific attribute level, e.g. how much "24/7 phone support" contributes to their preference relative to "email only." Utilities are relative within a study and can be converted to currency using the price attribute.
What's the difference between conjoint analysis and Van Westendorp?
Van Westendorp finds a credible price range for a whole product using four open-ended questions. Conjoint prices the individual features inside the product and can simulate market share for different configurations. Conjoint costs an order of magnitude more.
What is choice-based conjoint (CBC)?
The most common form. Respondents repeatedly choose one product profile from a set of two to four, plus a "none" option. It's preferred because choosing mirrors a real purchase decision better than rating or ranking does.
How many attributes can conjoint handle?
Choice-based conjoint works well with 4-8. Adaptive formats can carry 20-25 by narrowing the question set per respondent, at the cost of a longer survey (30+ minutes rather than under 15).
How large a sample does conjoint need?
300 is a common floor, 500-800 is comfortable, and you need enough in each segment you plan to analyse separately. Hierarchical Bayes estimation produces individual-level utilities, which helps, though it doesn't rescue a sample that's too small to represent the market.
Can conjoint analysis set my prices directly?
It can set them within the range and configurations you tested. It can't extrapolate outside your price levels, and its absolute willingness-to-pay numbers run high. Use the market simulator to compare options rather than treating any single output as the answer.
Does conjoint work for usage-based pricing?
Better than the alternatives. Model platform fee, included allowance, overage rate, and commit terms as separate attributes, use outcome-based levels rather than token counts, and include predictability features like spend caps as attributes in their own right.
What is MaxDiff and how is it different from conjoint?
MaxDiff asks respondents to pick the best and worst item from a set, producing a clean ranking of what matters most. It has no price attribute, so it prioritises features without valuing them. It's a good, cheap precursor to a conjoint study.
What is a market simulator?
A model built from conjoint utilities that predicts preference share for any set of product configurations and prices you specify. It's the main deliverable of a serious conjoint study and the thing you'll use for years after the fieldwork ends.
Why is attribute importance misleading?
Importance is calculated from the range of utility each attribute spans, which depends on the levels you chose. Test a wider price range and price looks more important. The number describes your study design as much as your market.
When should I not run conjoint?
When your buyer population is too small to sample, when the decision is brand or positioning rather than configuration, when you need an answer in two weeks, or when you already have enough transaction history to measure elasticity directly.
Educational reference. A conjoint study that says "move SSO into Pro and raise it €40" is only useful if you can actually ship that. Solvimon holds packaging, entitlements, and rate cards as configuration, so the repackaging is a catalog change. See pricing methodology.
Ready for billing v2?
Solvimon is monetization infrastructure for companies that have outgrown billing v1. One system, entire lifecycle, built by the team that did this at Adyen.
