Conjoint Analysis

How to Design a Conjoint Study

8 min read

Step-by-step guide to designing a conjoint analysis study. Learn how to choose attributes, define levels, set choice tasks, and avoid common design errors.

How to Design a Conjoint Study

What Goes Into a Conjoint Study Design?

A conjoint study design is the set of decisions that determine what respondents see and how their choices translate into usable data. Get the design right and you'll produce utility scores that reliably predict market behavior. Get it wrong and you'll collect thousands of responses that tell you nothing.

The core decisions are: which attributes to include, how many levels each attribute gets, what type of conjoint to use, how many choice tasks to present, and how to construct the experimental design that ties it all together. Each decision affects data quality, respondent experience, and the questions you can answer with the results.

Choosing Your Attributes

Attributes are the product features you want to test. The goal isn't to include every possible feature; it's to include the features that actually drive purchase decisions.

Start With Decision Drivers

Talk to your sales team, review customer feedback, and look at competitive positioning. The attributes that belong in your conjoint are the ones customers weigh when they're deciding between options. If nobody has ever mentioned "logo color" as a reason they bought or didn't buy, it doesn't belong in the study.

Keep the Count Between 4 and 7

Fewer than 4 attributes and you're probably better off with a simpler method like MaxDiff. More than 7 and respondent cognitive load spikes. Studies with 8+ attributes produce noisier data because respondents can't hold that many dimensions in working memory while evaluating trade-offs.

If you genuinely can't cut below 8, consider adaptive conjoint (ACBC), which tailors the exercise to each respondent's most relevant trade-offs.

Always Consider Including Price

Price makes the exercise feel realistic and enables willingness-to-pay calculations. Without price as an attribute, you can measure relative preference but you can't translate it into dollar values. For commercial research, price should be in almost every conjoint study.

Keep Attributes Independent

Each attribute must be independent of the others. Don't test shelf price, units per pack, and price per unit as three separate attributes when price per unit is mathematically determined by the other two. If two features always ship together in reality, combine them into a single attribute.

Defining Levels

Levels are the specific options within each attribute. This is where many studies go wrong.

Be Specific, Not Vague

"Fast shipping" vs. "slow shipping" is too imprecise. "Next-day delivery" vs. "3-5 business days" vs. "7-10 business days" gives respondents something concrete to evaluate. Use the same specificity you'd use in a product description.

Stay Realistic

Every level needs to be plausible to a respondent. Testing a $5 price alongside $500 teaches you nothing about the price range that matters. A good rule of thumb for price attributes: span roughly 60% to 140% of the current market price. Wide enough to capture sensitivity, narrow enough that every option feels real.

Use 2-5 Levels Per Attribute

Two levels give you a simple binary trade-off. Five levels give you a detailed preference curve. Beyond 5, the incremental information rarely justifies the added complexity and sample size requirements.

Balance matters too. If one attribute has 2 levels and another has 5, the design needs more respondents to estimate the 5-level attribute with the same precision. Try to keep level counts roughly balanced across attributes when possible.

Write for Your Audience

Use language your respondents actually use. If you're surveying consumers, "10 GB cloud storage" works better than "10 GB S3-equivalent object storage." Test your attribute and level descriptions with a few people from your target audience before finalizing the design.

Setting Up Choice Tasks

How Many Tasks Per Respondent

The sweet spot is 12-15 choice tasks per respondent. Below 8, you won't collect enough data per person for stable individual-level estimates. Above 18, fatigue sets in and response quality drops. Most practitioners land at 12 for shorter surveys or 15 when the conjoint is the survey's primary focus.

How Many Profiles Per Task

Show 3-5 product profiles per choice task. Three is the minimum for meaningful trade-offs. Five works when you have few attributes and want more data per task. Four is the most common default.

Include a "None" Option

Always add "None -- I wouldn't choose any of these" to each task. Without it, you're forcing every respondent to pick something, which inflates demand estimates. The "none" rate also serves as a diagnostic: if it exceeds 40%, your profiles may feel unrealistic.

Add Hold-Out Tasks

Reserve 1-2 choice tasks that aren't used in model estimation. After you run the analysis, check whether the model correctly predicts respondents' actual choices on these hold-out tasks. Hit rates above 70% suggest a valid model; below that, revisit your design.

Generating the Experimental Design

The experimental design is the plan that determines which attribute-level combinations appear together in each choice task. You don't create this manually.

Modern conjoint platforms use D-efficient or balanced overlap algorithms to generate designs that maximize statistical information per respondent. The design ensures every level appears approximately equally often and that the effects of different attributes can be separated cleanly.

What you should check before launching:

  • Level balance: Does every level within an attribute appear roughly the same number of times across all tasks?
  • Orthogonality: Can main effects be estimated without confounding from other attributes?
  • Interaction coverage: If you need to measure interactions between specific attributes (e.g., brand x price), does the design support that?
  • Prohibited pairs: Have you flagged any attribute-level combinations that shouldn't appear together (e.g., a luxury brand at the lowest price point)?

Most platforms handle the first two automatically. Interaction coverage and prohibited pairs require you to specify constraints before the design is generated.

Avoiding Common Design Errors

Testing Too Many Things at Once

Resist the urge to test everything in one study. A conjoint with 10 attributes, 5 levels each, and 3 interaction effects needs an enormous sample to produce reliable estimates. It's cheaper and more reliable to run two focused studies than one bloated one.

Overlapping Attributes

If "ease of use" and "simple interface" both appear as attributes, respondents won't know how to distinguish them. Each attribute must represent a clearly distinct dimension.

Unbalanced Level Counts

An attribute with 2 levels next to one with 6 creates estimation problems. The 6-level attribute needs proportionally more data. If you can't balance the counts, at least account for it in your sample size planning.

Skipping the Soft Launch

Always run 20-30 respondents before full fielding. Look for straight-lining, abnormally fast completions, and confusion signals (respondents abandoning mid-survey). A soft launch catches design problems that no amount of internal review will find.

Putting It Together: Design Checklist

Before you launch, verify:

  1. 4-7 attributes, each representing a distinct purchase driver
  2. 2-5 levels per attribute, specific and realistic
  3. Price included (for commercial studies)
  4. Attributes are independent of each other
  5. 12-15 choice tasks per respondent
  6. 3-5 profiles per task plus a "none" option
  7. 1-2 hold-out tasks for validation
  8. Prohibited pairs flagged in the design generator
  9. Total survey length under 15 minutes
  10. Soft launch planned before full field

Design your first conjoint study -- try Quali-Fi free for 14 days.

Related Guides

Put it into practice

Ready to apply this in your research?

Quali-Fi makes it easy to run surveys, conjoint studies, and more, all in one platform.