Conjoint analysis is a survey-based market research method that measures how people trade off product features against each other and against price — instead of asking what they want, you show them realistic bundles and watch what they pick. The payoff is a set of numbers (part-worth utilities) you can plug into a simulator to forecast share, willingness to pay, and the impact of any feature or pricing change. Below is how we run one end to end, and where most teams quietly break it.
Conjoint Analysis
Conjoint analysis is a quantitative research technique that infers how much consumers value individual product attributes and price levels by analyzing the choices they make across systematically varied product profiles.
Why direct questions lie and conjoint doesn’t
Ask someone “how important is price?” and almost everyone says “very.” Ask “do you want more features?” and everyone says yes. Stated-preference surveys are flattering and useless because nothing forces a trade-off. Real buying is a trade-off: you can’t have the cheapest, fastest, and most feature-rich option, so you compromise.
Conjoint analysis forces that compromise. By showing respondents competing product profiles and making them choose, you observe revealed preference inside a controlled survey — closer to a purchase than a Likert scale will ever get. The math then backs out the implied value of each attribute level from the pattern of choices.
This is the same trade-off discipline that separates real market opportunity analysis from wishful thinking, and it’s why conjoint outranks focus groups for product and pricing calls. It belongs to the same evidence-driven market intelligence toolkit as causal research and structured customer panel work.
Rule of thumb: if a research method lets respondents say “all of it, cheaply,” it’s measuring aspiration, not demand. Conjoint measures demand.
The vocabulary you actually need
Three terms carry the whole method:
- Attribute — a dimension of the product (price, battery life, brand, delivery speed).
- Level — a specific value of that attribute ($9 / $19 / $29; 6h / 12h / 24h).
- Part-worth utility — the score the model assigns to each level, expressing how much it pulls preference.
From part-worths you derive attribute importance (how much an attribute swings choice across its range) and feed everything into a market simulator to predict share and willingness to pay.
How to do conjoint analysis: the 7-step workflow
1. Define the decision first
Don’t start with “let’s run a conjoint.” Start with the call you need to make: What price for the mid tier? Which three features ship in the MVP? Will switching brand positioning cost us share? The decision dictates attributes, sample, and the simulator scenarios you’ll run. A study with no decision behind it produces a slide deck nobody opens.
2. Choose attributes and levels
This is where studies live or die. Keep attributes independent, realistic, and few — six to seven max for choice-based conjoint. Each level must be plausible and concrete (“24-hour delivery,” not “fast delivery”). Avoid dominated levels no rational buyer would pick, and never list a price range so wide it makes everything else irrelevant. Garbage attributes produce confident, garbage utilities.
3. Pick the conjoint method
| Method | How it works | Best for |
|---|---|---|
| Choice-Based Conjoint (CBC) | Respondents pick one profile from a set, including a “none” option | The default — pricing, share simulation, most B2C/B2B work |
| Adaptive CBC (ACBC) | Survey adapts to each respondent’s prior answers | Many attributes, or high-consideration purchases |
| MaxDiff | Forces best/worst choices across items | Ranking many features when price isn’t the question |
| Traditional full-profile | Rate or rank complete profiles | Legacy / small attribute sets; mostly superseded |
CBC is the workhorse. Reach for ACBC when you have a long attribute list, MaxDiff when you’re prioritizing features rather than modeling price.
4. Generate the experimental design
You can’t show every possible profile combination — that’s thousands of screens. A statistical experimental design selects a balanced, near-orthogonal subset so every level appears often enough and isn’t confounded with another. Software (Sawtooth, Conjointly, or an R package like support.CEs / idefix) builds this for you. Get this wrong and the model can’t separate which attribute is driving the choice.
5. Field to a representative sample
Recruit people who actually resemble your buyers — a beautiful study on the wrong audience is just expensive noise. For stable individual-level estimates with hierarchical Bayes, aim for roughly 300+ completes; tight subgroup reads need more. Police data quality hard: attention checks, speeder removal, straight-lining detection, and completion-time thresholds. The same customer panel hygiene you’d apply to any survey applies double here, because conjoint tasks are cognitively heavy and fatigue corrupts later screens.
6. Estimate the model
Modern practice fits hierarchical Bayes (HB) to recover part-worth utilities for each respondent, not just the aggregate — which preserves preference heterogeneity you’ll need for segmentation. Multinomial logit (MNL) is the simpler aggregate baseline. Validate before you trust anything: check hit rates on holdout tasks, watch for reversals (a “$29” level scoring higher than “$9” means something’s broken), and confirm importances make business sense.
7. Simulate, then ship the decision
This is the part most teams skip and it’s the entire point. Load utilities into a market simulator, define real competitive scenarios — your product vs. the two configurations you’re choosing between, against current rivals — and read predicted share and willingness to pay. Translate that into a roadmap call, a price, or a positioning bet. Then validate in the wild with an A/B test or a pilot launch, because a simulator is a forecast, not a fact.
Where conjoint studies go wrong
- Too many attributes. Past ~7, respondents simplify by ignoring some, and your utilities measure attention, not value.
- Implausible levels. Prices nobody would pay or features nobody understands poison the design.
- No “none” option. Forcing a choice when the real answer is “I’d buy nothing” inflates demand for everything.
- Treating the simulator as truth. It predicts; the market decides. Pilot before you bet the roadmap.
- Skipping the decision. A conjoint with no question behind it is a research vanity project.
For the SEO and product crowd, conjoint pairs naturally with total addressable market sizing and commercial intent signals: conjoint tells you what configuration and price wins, demand data tells you how many there are to win. If you’re using it to prioritize a content or product roadmap, that’s exactly the kind of demand evidence we lean on inside our programmatic SEO program before committing engineering to a build.
How conjoint compares to other preference methods
| Method | What it measures | Forces trade-offs? | Good for pricing? |
|---|---|---|---|
| Conjoint (CBC) | Value of feature/price combinations | Yes | Yes |
| MaxDiff | Relative priority of many items | Yes | No |
| Direct rating (Likert) | Stated importance, item by item | No | Weakly |
| Van Westendorp | Acceptable price range only | Partly | Price only, no features |
If the question is “which features, at what price, win against which competitor,” conjoint is the only method on this list that answers all of it in one study.
Frequently Asked Questions
What is conjoint analysis used for?
Conjoint analysis is used to measure how customers trade off product features and price, then forecast outcomes. Teams use it to set pricing, prioritize which features to build, design bundles, estimate willingness to pay, segment buyers by preference, and simulate market share for new configurations against existing competitors before committing budget.
How many respondents do you need for conjoint analysis?
It depends on the model and how granular your reads must be. For a standard choice-based conjoint estimated with hierarchical Bayes, plan for roughly 300 or more clean completes. If you need reliable estimates within subgroups or segments, scale up proportionally — each subgroup effectively needs its own viable sample size.
What’s the difference between conjoint analysis and MaxDiff?
Conjoint shows respondents full product profiles with price and infers the value of each feature-and-price combination, making it ideal for pricing and share simulation. MaxDiff forces best/worst choices among a list of single items to rank their relative importance, but doesn’t model price or realistic product bundles. Use conjoint for “what to ship and price,” MaxDiff for “what matters most.”
Can I run conjoint analysis without specialist software?
Yes, but carefully. R packages like support.CEs and idefix generate experimental designs and estimate models for free, and bayesm handles hierarchical Bayes. The hard parts are building a balanced design and validating the estimates — areas where dedicated tools like Sawtooth or Conjointly remove footguns. DIY works when someone on the team genuinely understands experimental design.
Is conjoint analysis still relevant with AI Overviews and AI-driven research?
Yes. AI Overviews can summarize what conjoint is, but they can’t run your study or know your customers’ trade-offs. Primary, first-party preference data is exactly the kind of proprietary insight that AI can’t scrape — which makes conjoint more valuable, not less, in a privacy-era market where third-party signals keep degrading.