Conjoint Analysis & Simulation Tool

Analysis Tool Design Tool →

Analyze choice-based conjoint (CBC) data to understand how customers value product attributes. Upload experiment data, estimate individual-level preferences (part-worth utilities), simulate market shares, and optimize product configurations for maximum profit.

1 ✓

Load Data

2 ✓

Map Columns

3 ✓

Configure Attributes

4 ✓

Estimate Model

5 ✓

Analyze Results

6 ✓

Simulate Markets

OVERVIEW & WHAT THIS TOOL SOLVES

If you ask customers what they want in a survey, they say "every premium feature at the lowest price." In real markets, buyers have to make trade-offs among competing bundles. Choice-Based Conjoint (CBC) observes which complete product profiles customers actually pick across repeated choice tasks, decomposes those choices into dollar-equivalent feature values (part-worth utilities), and lets you simulate market share and profit before launching a single SKU.

🎯 Questions It Answers

Which product attributes (brand, features, warranty, price) truly drive customer choice—and which ones are nice-to-have fluff?
How much extra are customers willing to pay (WTP) for a specific feature upgrade?
If we launch Configuration A or raise price by $10 against real competitors, what market share and profit will we capture?

📂 What Data You Need

Long-format choice task data (one row per product alternative per task per respondent).
Respondent ID, Task ID, Alternative ID, and a binary Chosen (0/1) indicator.
Attribute columns (categorical features or numeric price) plus optional "None" (opt-out) rows.

💡 What You'll Discover

Individual Part-Worths & Importance: Respondent-level utilities and attribute importance rankings.
Willingness-to-Pay (WTP): Dollar value of every feature level relative to its baseline.
Market Share & Profit Simulator: Live competitive share forecasts, customer segments, and optimal product configs.

🔍 Key Nuances & Common Traps to Watch For Click to expand ▾

Why Individual-Level Importance Matters: Attribute importance is computed by taking each respondent's utility range within an attribute, dividing by the sum of their ranges, and averaging across respondents—so opposing customer tastes (e.g., half love Brand A, half love Brand B) don't cancel out to zero.
The "None" (No-Purchase) Option: The outside "None" alternative is modeled as a respondent-specific constant baseline utility with zero product features. A strongly negative "None" utility means customers are eager to buy something in the category; a positive "None" utility means customers walk away unless a bundle offers compelling value.
Reference Levels & Zero-Anchoring: Every categorical attribute designates one level as the baseline reference (fixed at 0 utility). A positive part-worth means customers prefer that feature level over the reference level, holding price and other features constant.

📐 Under the Hood: Model Equations & Variable Definitions Click to expand ▾

1. Respondent Random Utility Model (Part-Worth Decomposition) Individual-Level Utility

\[ U_{ijt} = \alpha_{j,i} + \sum_{a \in \text{cat}} \sum_{\ell} \beta_{a,\ell,i} \cdot I(\text{level}_{jt}(a) = \ell) + \sum_{b \in \text{num}} \beta_{b,i} \cdot x_{b,jt} + \beta_{\text{price},i} \cdot p_{jt} \]

💬 Read it in plain English: The total attractiveness (utility) that customer $i$ assigns to product option $j$ in choice task $t$ is the sum of the part-worth scores for each feature that product carries, minus the pain of paying its price $p_{jt}$.

Variable & Symbol Directory (Every Term Defined) Hover or focus any role badge for details

U_ijt

Total Product Utility OUTPUT ⓘ

Overall utility score respondent $i$ assigns to alternative $j$ in choice task $t$.

α_j,i

Alternative-Specific Constant (ASC) ESTIMATED ⓘ

Baseline utility for respondent $i$ choosing a fixed option $j$ (such as the "None" opt-out).

β_a,ℓ,i

Categorical Part-Worth Utility ESTIMATED ⓘ

Value respondent $i$ places on feature level $\ell$ of categorical attribute $a$ compared to its baseline level.

I(·)

Feature Presence Indicator (0/1) YOUR DATA ⓘ

Equals $1$ if product $j$ in task $t$ has level $\ell$ on attribute $a$, and $0$ otherwise.

β_b,i, x_b,jt

Numeric Attribute Slope & Value ESTIMATED ⓘ

$\beta_{b,i}$ is respondent $i$'s utility gain per unit of numeric feature $x_{b,jt}$.

β_price,i, p_jt

Price Sensitivity & Profile Price ESTIMATED ⓘ

$\beta_{\text{price},i}$ (typically negative) measures utility lost per unit of price $p_{jt}$, anchoring dollar WTP ($-\beta_{a,\ell,i}/\beta_{\text{price},i}$).

2. Multinomial Logit (MNL) Choice Probability & Share Rule Choice & Market Simulation

\[ P(y_{it} = j) = \frac{\exp(U_{ijt})}{\sum_{k \in C_{it}} \exp(U_{ikt})} \]

💬 Read it in plain English: A product's predicted choice probability is its exponentiated utility divided by the sum of exponentiated utilities across all competing options on the shelf ($C_{it}$)—guaranteeing probabilities between 0% and 100% that sum to 100%.

Variable & Symbol Directory (Every Term Defined) Hover or focus any role badge for details

P(y_it = j)

Predicted Choice Probability OUTPUT ⓘ

Probability that respondent $i$ chooses option $j$ in choice set $t$; averaging across all respondents yields simulated market share.

y_it

Observed Choice Outcome YOUR DATA ⓘ

The alternative actually selected by respondent $i$ in task $t$ during the conjoint study.

exp(U_ijt)

Exponentiated Attraction Weight OUTPUT ⓘ

$e^{U_{ijt}}$, which converts any real-valued utility into a positive weight so higher-utility bundles capture disproportionately larger share.

C_it

Competitive Choice Set YOUR DATA ⓘ

All competing alternatives $k$ shown together in task $t$ (or configured in the Market Simulator, including the "None" option).

STEP 1: LOAD DATA

Select a marketing scenario:

Select a scenario above to load a preset CBC study dataset with context.

Drag & drop CBC file

CSV with respondent_id, task_id, alternative_id, chosen, and attribute columns

No file uploaded.

STEP 2: MAP COLUMNS

Map your CSV columns to the required CBC data structure fields.

📖 What do these fields mean? (Click to expand)

Your CBC data should be in "long format" — one row per alternative that was shown to a respondent in each choice task.

Field	What It Is	Example Values
Respondent ID	Unique identifier for each person who took the survey. Typically auto-generated by your survey platform.	R001, R002, 12345, email@domain.com
Task ID	Which choice task (question) this row belongs to. Each respondent sees multiple tasks (typically 10–15), each showing a set of products to choose from.	1, 2, 3... or Task_1, Task_2...
Choice Option	Identifies which product/option this row represents within the task. Could be product names (iPhone, Samsung), generic labels (Option A, Option B), or "None".	iPhone, Samsung, BrandX, None, Option_1, Option_2
Chosen (0/1)	Did the respondent pick this option? 1 = Yes, they chose this. 0 = No, they didn't choose it. Exactly ONE option per task should have chosen=1.	0 or 1

💡 Example: If respondent R001 saw 12 tasks, and each task had 4 options to choose from, there would be 48 rows (12 × 4) for that respondent. In each task's 4 rows, exactly 1 row should have chosen=1.

Respondent ID Column

Task ID Column

Choice Option Column

Also called the Alternative ID.

Chosen (0/1) Column

INPUTS & SETTINGS

Step 3: Configure Attributes

Configure each attribute's type and identify special alternatives (None, Competitors). The tool will automatically dummy-code categorical attributes and create appropriate model terms.

📖 Understanding Attribute Types (Click to expand)

Each product attribute in your conjoint study needs to be classified so the model knows how to treat it:

Type	When to Use	Examples
Categorical	Named categories with no inherent order. Each level gets its own utility estimate.	Brand (Apple, Samsung, LG), Color (Red, Blue, Black), Style (Modern, Classic)
Numeric (Linear)	Continuous numbers where the effect is consistent. Every unit increase has the same impact.	Screen size (inches), Storage (GB), Year of manufacture
💲 Price (Special)	Used for dollar amounts like price. Why is this special? Because price coefficients are used to calculate: 💵 Willingness-to-Pay (WTP): How many $ a feature is worth 📊 Profit calculations: In market simulations Mark your price column as "Price" so these calculations work correctly.	Price ($), Monthly fee, Annual cost
Numeric (Quadratic)	When there's a "sweet spot" — too little OR too much is bad. The model fits a curve instead of a straight line.	Sweetness level (people like medium, not extreme), Delivery time (same-day good, but too fast = suspicious?)

🎯 Quick Rule: If it's text (brand names, features) → Categorical. If it's a number → Numeric Linear (or Price if it's money).

Special Alternatives

📖 When and Why to Use Special Alternatives (Click to expand)

In CBC studies, some choice options are "special" because they don't follow the normal attribute-based utility structure:

🚫 "None" / Opt-Out Alternative

What it is: A "none of these" or "I wouldn't buy any" option that lets respondents decline all products in a task.

Why it matters: Without this, you force respondents to pick something even if they hate all options. The "None" option:

Captures realistic "no purchase" behavior
Provides a benchmark for how good products need to be to win a sale
Enables more realistic market share predictions

For the Smartphone example: If your data has alternatives like "iPhone", "Samsung", "BrandX", and "None", select "None" in the dropdown.

🏢 Competitor Alternatives

What they are: Real branded products (not generic "Option A/B/C") that appeared consistently across choice tasks with their actual market attributes.

When to mark as competitor:

✅ A known brand (iPhone, Toyota Camry) that always appears with its real specs
✅ Products you're comparing against but not designing/optimizing
❌ NOT generic options with varied attribute combinations

Why it matters: Competitors get a "brand-specific constant" that captures overall brand value beyond their individual attributes. This is useful for simulation when you want to see how your new product competes against established brands.

For the Smartphone example: If "iPhone" always appears with real iPhone specs (not hypothetical combinations), mark it as a competitor. If all brands (iPhone, Samsung, BrandX, BrandY, BrandZ) have varied attribute combinations in the study, leave them unchecked—they're design options, not fixed competitors.

🤔 Not sure? If your conjoint study uses hypothetical products where attributes vary (the typical setup), you probably don't need to mark any competitors. Just identify "None" if it exists.

"None" / Opt-Out Alternative (optional)

Select if your study includes a "no purchase" or "none of these" option. Blank-attribute opt-out rows are auto-detected.

Competitor Alternatives (optional — usually not needed)

Alternatives with identical attributes in every task are auto-checked as competitors. You can override by checking or unchecking any box.

VISUAL OUTPUT

Average individual importance over the tested levels. Longer bars indicate larger relative fitted utility ranges. Mean individual importance summarizes each person's utility ranges before averaging; importance of average utilities measures the consensus profile where opposing tastes can cancel.

How to read this chart

Each bar represents the relative importance of one attribute in driving customer choice. Bars are sorted from most to least important and sum to 100%.

Bar length = an attribute's share of the total fitted utility range, averaged across respondents; it is not variance explained.
Tested levels matter: narrowing or widening an attribute's range can change its importance.
These averages can conceal differences between respondents. Check part-worths, respondent quality and segments before making a decision.

💡

Marketing Insight

Use importance to compare trade-offs within the tested ranges. No universal percentage makes an attribute negligible; consider feasible changes, costs and differences between respondents before choosing a strategy.

Show ±1 SD respondent heterogeneity whiskers

Average utility for each attribute level. Categorical levels are shown relative to their reference level (ref = 0.00); continuous attributes show their per-unit slope. Compare within attributes to see which levels win.

How to read this chart

Each bar shows the average part-worth utility for one attribute level or continuous slope, grouped by attribute and color-coded.

Categorical attributes (with ref = 0.00): Your chosen reference level for each categorical attribute sits at 0.00 (ref); positive bars add utility above that baseline and negative bars subtract utility. Switching to Zero-Centered (Mean = 0) shifts each categorical attribute's levels so their mean is 0.00 without changing utility spreads or simulated shares.
Continuous / Price attributes (slope per +1 unit): Instead of separate bars for each number, a linear attribute has a single slope (β_m) showing how much utility changes per 1-unit (or $1) increase. Multiply β_m × x_m to get its utility contribution at level x_m.
Outside Option: "None" (ASC_None — Walk-Away Hurdle): This is not a per-unit slope! It is the one-time lump-sum utility of walking away ("Would Not Buy"). A product configuration's total utility (U_product = Σβ_cat + Σβ_numx_m) must exceed ASC_None for a customer to prefer buying over walking away. Because continuous attributes (like Price) subtract β_price × Price from total product utility, ASC_None is scaled in raw-attribute utility units—always compare ASC_None against a complete product profile that includes a concrete price and numeric values (see the live Walk-Away Hurdle breakdown below!).
±1 SD Whiskers: Toggle on the heterogeneity whiskers to see where respondents strongly disagree (±1 standard deviation)—wide whiskers indicate polarized preferences and segmentation opportunities.

📊

Interpretation Guide

Positive utilities indicate preference above the baseline; negative values indicate dispreference. Combine high-utility categorical levels with competitive pricing so total bundle utility clears both competitor offerings and the ASC_None walk-away hurdle.

Select Attribute to Inspect:

Spotlight Respondent:

Each violin shows the full distribution of individual respondent utilities for one level of the selected attribute, complete with an inner median/IQR box, dashed mean line, and jittered respondent dots.

Level / Parameter	Mean Utility	Median (IQR)	Std. Dev.	% > Baseline (Ref)	% Personal #1 Level

How to read this heterogeneity view

Why default to Zero-Centered (Mean = 0)? In baseline-referenced coding, the reference level is fixed at 0.00 for every respondent, which hides how much respondents disagree about the baseline level itself. Zero-centering each respondent's utilities within the selected attribute gives every level (including the reference level) its own full violin distribution!
Violin + Respondent Dots: Wide or bimodal violins indicate polarized tastes (some buyers love the level while others dislike it). Narrow violins indicate broad market consensus.
Respondent Trajectories: Connects each respondent's utilities across the levels of this attribute (or plots each respondent's utility curve across the tested numeric/price range). Steeply crossing lines—or segments colored after running K-Means Segmentation below—reveal distinct customer archetypes.
% Personal #1 Level: The share of respondents for whom this level has the highest utility within the attribute (ties split evenly). Even a level with a negative average utility can be the #1 favorite for a passionate niche segment!

Distribution of price coefficients across respondents. More negative = more price-sensitive.

How to read this chart

This histogram shows the distribution of individual-level price coefficients across all respondents.

X-axis = price coefficient value. More negative = more price-sensitive.
Blue dashed line = mean coefficient; green dotted line = median. A large gap between them suggests outliers or skew.
Black dashed line at 0 = divides price-sensitive (left) from price-insensitive or positive-coefficient respondents (right).
Spread matters: a wide distribution justifies segment-based pricing; a tight cluster suggests uniform sensitivity.
Some positive coefficients are normal in small samples — they may reflect prestige effects or noisy estimation.

⚠️

Common Mistake

Don't assume all customers have the same price sensitivity! A wide distribution justifies segmentation analysis and potentially tiered pricing strategies (economy, standard, premium).

MODEL RESULTS

Total respondents estimated: –

Mean tasks per respondent: –

Mean pseudo-R² (McFadden): – ?

Estimation time: –

Interpretation Aid

Pseudo-R²: Measure of model fit per respondent. Values 0.2–0.4 are typical for CBC studies; higher values indicate stronger predictive power.

Tasks per respondent: More tasks (12–20) produce more stable coefficient estimates. Fewer than 8 tasks may yield unreliable utilities.

APA-Style Report

Estimate the model to see the APA-style statistical report.

Managerial Interpretation

Business-focused interpretation will appear here after estimation.

Aggregated Part-Worth Utilities

How to read colors & symbols: ★ Best Level Most-preferred categorical level in attribute +0.304 Positive mean utility (above baseline 0.000 or positive slope) −0.436 Negative mean utility (below baseline 0.000 or negative slope) Note on statistical significance vs. heterogeneity: Green and red colors indicate preference direction (sign) relative to baseline, not a frequentist p-value threshold. Because this model estimates a separate ridge-penalized utility vector for every respondent, Std. Dev. measures real taste heterogeneity across people rather than sampling standard error (a small Std. Dev. relative to Mean Utility indicates strong consensus; a large Std. Dev. indicates polarized segments).

Attribute	Level	Mean Utility	Std. Dev.	Min	Max
Estimate the model to view aggregated part-worth utilities.

How to Interpret Part-Worth Utilities

Mean Utility: Average part-worth value across all respondents. Higher utilities indicate stronger preference.

Categorical attributes (brand, camera): Each level gets a coefficient relative to a baseline (usually the first alphabetically). Positive values mean preference over baseline; negative means less preferred.
Linear attributes (price, screen_size): Single coefficient per attribute. Positive = "more is better" (e.g., bigger screen); negative = "less is better" (e.g., lower price).
None alternative (ASC_None): The constant utility for choosing "no purchase." This has no product attributes—it's a fixed outside option. Negative values mean respondents strongly prefer buying something; positive values indicate high barriers to purchase (people prefer doing nothing unless products offer substantial value).
Competitor alternatives (ASC_Competitor_*): Positive = brand loyalty to competitor; negative = preference for your brands.

Std. Dev.: Variation across respondents. High std dev (>0.5) suggests heterogeneous preferences—good candidates for segmentation.

Min/Max: Range of individual utilities. Wide ranges indicate some respondents strongly prefer/dislike this level.

MARKET SIMULATION

Price Configuration

First, specify how price should be treated in the market simulation.

Price Treatment

If price was varied as an attribute, utilities can be translated into willingness-to-pay (WTP) values.

Cost Structure

Define what it costs you to make one unit. The tool adds these up into a single unit cost for each product configuration, then uses it to compute profit as market share × (price − unit cost). This same cost structure also drives the Product Optimization profit search below, so you only define costs once.

Total unit cost = Base Cost + sum of the incremental costs for the specific attribute levels a product uses. Example: base $30, "256GB storage" +$40, "OLED screen" +$15 → unit cost $85.

Base Cost (per unit)

The cost every unit incurs no matter its configuration (e.g., core manufacturing, packaging, base components).

Incremental Costs by Attribute Level

The extra cost each non-baseline level adds on top of the base cost. The baseline level of each attribute is fixed at $0. Enter positive values for pricier options (e.g., more storage) or negative values for cheaper options. Leave at 0 if a level has no cost impact.

Scenario Builder

Define a competitive scenario with multiple product alternatives. The tool will compute predicted market shares using the estimated individual utilities.

Simulation Options

Force all to choose (exclude the "None" option from the simulation) Ignore competitors (exclude competitor products from the simulation)

By default, competitors and "None" (if present in study) are included in simulations. Use these options to run "must purchase" or "our brands only" scenarios.

Total market size (customers)

Used to convert share percentages into customer counts and total profit dollars.

One-attribute sensitivity for the first custom product

Compares changes to the first product card with the current baseline. Rival profiles and the outside option stay fixed; price, costs and profits use the same simulator. Numeric attributes use observed low, middle and high values. Preference shares describe this choice set; revenue and profit are conditional scenarios, not market forecasts.

Attribute	Level	Preference share (%)	Change (pp)	Unit cost ($)	Margin ($)	Profit ($)	Profit change ($)

Simulation Results

📊

Reading Your Results

Market share shows what % of customers prefer each option. If "None" captures >30% share, customers may not find current products compelling—consider improving core features or adjusting prices. Compare profit (not just share) to identify the optimal product configuration.

Market Share by Alternative

Predicted preference share based on individual-level utilities averaged across all respondents.

How to read this chart

Bars represent predicted preference share and sum to 100%.
Blue = your products, orange = competitors, grey = None (opt-out).
If "None" exceeds 30%, many customers don't find any product compelling — consider improving features or lowering prices.

Total Profit by Alternative

Profit = Share × Market Size × (Price - Cost). Identifies which products maximize revenue.

How to read this chart

Profit combines market share AND margin, so the share leader is not always the profit leader.
Products with low share but high margin can still be very profitable.
Negative profit means cost exceeds price for that product at its predicted share level.

Product	Preference Share (%)	Customer Count	Price	Cost	Margin	Total Profit

PRODUCT OPTIMIZATION

Brute-Force Configuration Search

Explore all possible combinations of attribute levels to find the product configuration that maximizes market share or profit. Useful for new product design decisions.

📖 How does brute-force handle numeric attributes? (Click to expand)

Brute-force optimization tests all combinations of attribute levels. But continuous attributes (like price or screen size) have infinite possible values! Here's how the tool handles this:

Attribute Type	What Values Are Tested
Categorical (Brand, Color)	All unique values that appeared in your CBC data.
Price (when optimizing profit)	A user-defined grid: you specify min, max, and step size. For example, $499–$899 with $50 steps tests 9 price points.
Other Numeric (Screen Size, Storage)	Only the discrete values that appeared in your CBC study. If your study used Screen Size = [5.5", 6.0", 6.5"], only those 3 values are tested.

⚠️ Limitation: The true optimal might fall between tested values (e.g., optimal price = $572, but we only test $550 and $600). For more precision, reduce the price step size or use the Market Simulation to manually test specific configurations.

Select attributes to vary

Optimization goal

Price range to test

Profit is evaluated as share × (price − unit cost), where the unit cost comes from the Cost Structure you defined in Market Simulation (base cost + per-level incremental costs). Set that up first so profit reflects your real costs.

Min price

Lowest price the optimizer will consider.

Max price

Highest price the optimizer will consider.

Price increment

Step between prices tested (e.g., 10 tests $0, $10, $20…). Smaller = finer search but more combinations.

DIAGNOSTICS & ASSUMPTIONS

Model Assumptions & Data Quality Checks

Estimate the model to see diagnostic information.

CBC Study Design Requirements

Sample size: Minimum 100–200 respondents for stable aggregate estimates; 200–500+ recommended for segmentation.
Tasks per respondent: 8–20 choice tasks provide reliable individual-level estimates. Fewer than 8 may yield unstable coefficients.
Alternatives per task: 3–5 alternatives per task is typical. Too few (2) limits information; too many (6+) increases cognitive burden.
Attribute levels: Use 3–5 levels per categorical attribute. Binary attributes provide limited information.
Orthogonality: Experimental design should minimize correlation between attributes. Check design quality in your CBC software.

Model Assumptions

IIA (Independence of Irrelevant Alternatives): MNL assumes adding/removing an alternative affects all others proportionally. Violated when alternatives are perceived as close substitutes.
Linear utility: Utilities are linear combinations of part-worths. May not hold for complex interactions or threshold effects.
Homogeneous preferences (at individual level): Each respondent has stable preferences across tasks. Violated if respondents use different decision rules in different tasks.
No outside information: Respondents base choices only on information presented in the task. Violated if they apply external knowledge about brands or features.

Held-out prediction check

Pseudo-R2 Distribution Across Respondents

Distribution of individual-level McFadden pseudo-R2 values. Higher values = better model fit for that respondent. Very low R2 may indicate random or inattentive responses.

How to read this chart

X-axis: Pseudo-R2 value (0 = no better than random, 1 = perfect prediction).
Blue dashed line: Mean R2 across all respondents.
Left tail (near 0): Respondents whose choices appear random — consider flagging or removing.
Right tail (0.5+): Respondents with very consistent, predictable preferences.
A bimodal distribution may suggest two distinct response quality groups.

👨‍🏫 Professor Mode: Guided Learning Experience

OVERVIEW & WHAT THIS TOOL SOLVES

🎯 Questions It Answers

📂 What Data You Need

💡 What You'll Discover

STEP 1: LOAD DATA

Use a Case Study

Upload Your Own Data

STEP 2: MAP COLUMNS

INPUTS & SETTINGS

Step 3: Configure Attributes

Special Alternatives

🚫 "None" / Opt-Out Alternative

🏢 Competitor Alternatives

Analysis Settings

VISUAL OUTPUT

Attribute Importance (average of individuals)

Part-Worth Utilities

Respondent Preference Heterogeneity (By Attribute)

Price Sensitivity Distribution

MODEL RESULTS

APA-Style Report

Managerial Interpretation

Aggregated Part-Worth Utilities

Individual Respondent Utilities

SEGMENTATION ANALYSIS

K-Means Clustering on Utilities

Elbow Method: Finding Optimal K

Segment Sizes

Segment Profiles: Attribute Importance

Segment-Level Mean Utilities

MARKET SIMULATION

Price Configuration

Cost Structure

Incremental Costs by Attribute Level

Scenario Builder

One-attribute sensitivity for the first custom product

Simulation Results

Market Share by Alternative

Total Profit by Alternative

Willingness-to-Pay (WTP) Analysis

PRODUCT OPTIMIZATION

Brute-Force Configuration Search

Select attributes to vary

Price range to test

Top 10 Configurations

DIAGNOSTICS & ASSUMPTIONS

CBC Study Design Requirements

Model Assumptions

Held-out prediction check

Pseudo-R2 Distribution Across Respondents