How it works
A conjoint study in three steps.
Each one shown in the real app.
From the first attribute to the market simulator. Every step below is a short recording from Choicery, with the details one click away.
Step 1 of 3 · Build & publish
Build the study and go into the field
Questions, attributes and product images in one survey, published with one click.
- Your own questions, attributes, levels and product images in one survey
- One click generates a balanced design, checked before anyone answers
- Publish once and share one link
Details and defaults
You decide
- Which method fits the question: CBC, CBC Shop, DR-CBC, BW-CBC, MaxDiff or Anchored MaxDiff
- The attributes and their levels, as text or as images
- How many tasks each respondent answers and how many products each task shows
- Which level combinations must never appear together
- Whether a “None of these” option or a buy/no-buy follow-up belongs in the task
- How the task looks: Inline, Stacked or Compact Grid, or a shop or shelf layout for CBC Shop
Choicery takes care of
- The design itself. Balanced Overlap by default: levels appear about equally often and level pairs stay balanced. D-optimal and random designs are one click away.
- Many versions. Respondents see different, balanced sets of tasks, 200 versions by default.
- Diagnostics. Design efficiency, level balance, independence of the attributes, repeated levels and a standard-error estimate for every level.
- Prohibitions that keep impossible products out of every version.
- Fixed tasks that serve as holdouts, so the model can later be tested on choices it never saw.
- Product images that follow the levels, so a different feature also shows a different picture.
- A locked experiment. Once the first responses arrive, the design can no longer be regenerated, so every respondent answers the same study.
Defaults and limits
- Design method
- Balanced Overlap by default, D‑Optimal or Random on request
- Versions
- 200 by default, up to 1,000
- Study size (Pro)
- Up to 20 attributes with 20 levels each, 30 random tasks plus up to 20 fixed tasks, up to 9 products per task
- MaxDiff (Pro)
- Up to 100 items with balanced item and pair frequencies; Anchored MaxDiff adds a calibration question
- Downloads
- Design as CSV and a diagnostics report
- A generated design
- Diagnostics you can read
- A preview of every task
- Design CSV and report
You decide
- The questions around the choice tasks: screening, usage, attitudes, demographics
- The order of pages and what gets shuffled: answer options, grid rows, ranking items, whole pages or questions across pages
- Who sees what, with skip and display logic built from AND/OR conditions
- Experimental groups for a between-subjects design, with their weights
- The look: style template, colors, fonts, logo, button texts and progress bar
- How people arrive: the survey link, a short custom link, a QR code or personalized panel links
Choicery takes care of
- Layouts that work on phones, checked in Preview & Test at desktop, tablet and phone sizes. Previews save nothing.
- Drafts and published versions. Edit freely, publish the changes in one step, and keep a publish history.
- Logic checked on the server, so a hidden question never collects data, even from a manipulated browser.
- A checksum on every response (HMAC-SHA256), so stored answers can be checked for later changes.
- Panel IDs read from the link and stored with the response, with an option to accept only one completed response per ID.
- Protection against junk starts. Link-preview crawlers do not count as participants, and the server limits how fast surveys can be started and pages submitted.
- Lean data. No IP address, browser or device details are stored with a response. Responses live in the database with a backup file per project, on servers in Germany.
Fieldwork settings
- Respondents
- No account and no install, on desktop, tablet or phone
- Availability
- Start and end date, pause at any time, close after a maximum number of completes
- Shared devices
- A mode for labs, tablets and kiosks where several people answer on one device
- Language
- Buttons and participant messages in 11 languages
- Privacy notice
- Hosted in Choicery or your own link; YouTube and Vimeo load only after a click
On the fieldwork dashboard
What you can check while responses come in.
Progress
Completion rate, completed and incomplete responses, average and median completion time.
Drop-off
The questions where incomplete responses stopped, ranked by how many people left there.
Timing
The distribution of completion times and the time spent on each page.
Fieldwork pulse
Responses today and in the last 24 hours, by day and by hour of the day.
Coverage
Respondents per conjoint block and per experimental group, so imbalances show early.
Data health
Checksum and backup status, and relative speeders flagged for review.
- A published survey
- Responses with checksums
- Exports at any time during fieldwork
- A bin to restore deleted responses
Step 2 of 3 · Analyze
Estimate what every respondent values
Turn every respondent's choices into preferences, with tested defaults.
- Hierarchical Bayes in one click, with covariates if you want them
- Segments, AI-driven profiles and interaction search when an average is not enough
- Every estimation downloads as a replication package
Estimations are shown sped up. Everything else is recorded in real time in the real app.
Details and defaults
You decide
- Which responses go in: filter by completion, participant ID, survey answers or date
- Whether numeric attributes such as price are coded linear or categorical, and whether price is held negative
- Burn-in and draws, or simply the defaults
- Whether a holdout task or a validation split checks the model
Choicery takes care of
- The estimation. A Julia HB sampler on Choicery's servers, with progress per chain and iteration. If the server is busy, you can join a queue that shows your position.
- Convergence checks: R-hat, effective sample size and the acceptance rate of the sampler.
- Fit and validation: root likelihood (RLH), holdout hit rates against chance and a respondent holdout.
- Results per respondent and for the sample: utilities, importance and charts, ready for the simulator.
- The rest of the questionnaire: descriptives, frequencies, group comparisons, crosstabs with chi-square, correlations and regressions, plus Kano, Van Westendorp and Gabor-Granger results.
- Data quality screens for speeders, straightlining, long strings and low variance. They flag responses for review and never delete anything automatically.
Defaults and limits
- Default model
- Hierarchical Bayes, multi-chain MCMC in Julia
- Where it runs
- On Choicery's servers in Germany, or on your own machine via the replication package (Pro)
- Diagnostics
- R-hat, ESS, acceptance rate, RLH, holdout hit rate, Cohen's kappa, MAE, RMSE
- Saved runs
- Up to 10 per method and choice task, so runs with different settings can be kept
When the average is not enough
Further models on the same data, in the same project.
Latent ClassSegments
Finds groups with distinct preferences, picks the number of groups by BIC and describes who is in each one, with or without covariates.
AI-Driven Profile AnalysisCharacteristics
A neural network learns how respondent characteristics shape preferences, ranks those characteristics with SHAP and predicts preferences for new profiles.
Interaction SearchCombinations
Tests every pair of attributes for an interaction, with a false-discovery-rate correction so many tests do not produce chance findings.
MaxDiff scoresRankings
Preference scores for every item, purchase probabilities for Anchored MaxDiff, and a holdout check of the scores.
- Utilities for every respondent
- Attribute importance
- Convergence and fit diagnostics
- Segments, if you need them
Step 3 of 3 · Decide
Test decisions before you make them
Put your product into a simulated market and see how choices shift.
- Put your product next to its competitors
- Choice share with confidence intervals, willingness to pay, reservation price and elasticity
- Personas show what a specific customer would choose
Details and defaults
You decide
- Which products compete: your own, the competitors and the “None” option
- Which run feeds the simulator: HB or AI-Driven Profile Analysis
- Which segment you look at, filtered by any survey variable
- The question: share, price, willingness to pay or sensitivity
- What you hand over, and in which format
Choicery takes care of
- Choice share by Share of Preference, Randomized First Choice or First Choice, with bootstrap confidence intervals.
- Willingness to pay three ways: Simulation WTP, Robust WTP across many sampled market scenarios, and Pseudo-WTP.
- Price tools: reservation price with a demand curve, own-price and cross-price elasticity.
- Sensitivity: one attribute changed level by level, with every product's share.
- Persona simulation after an AI-Driven Profile Analysis run: predicted preferences for a profile that never took the survey.
- A MaxDiff simulator with item rankings, assortment shares and TURF reach.
Hand-over formats
- Reports
- PDF reports, Excel workbooks, PNG charts, APA-style Word tables
- Data
- Responses as Excel, CSV or SPSS, preference scores next to the answers, long-format choice data
- Documentation
- Codebook, the questionnaire as a Word file, a full project backup
- Market shares
- Willingness to pay
- Price curves
- Reports for your team
Questions about the workflow
Short answers. The FAQ page covers pricing and accounts.
Which design method should I use?
Balanced Overlap, the default, suits most studies: it keeps levels and level pairs balanced and allows some overlap between products in a task. D-optimal is available when you want to maximize statistical efficiency, and a random design is available as well. The diagnostics show the result of each choice before you publish.
Can I change the survey after it is published?
Yes. Edits stay as unpublished changes until you publish them, and the publish history lists every release. The conjoint design itself cannot be regenerated once responses exist, so every respondent answers the same experiment.
Can I test the survey without collecting data?
Yes. Preview & Test opens the draft or the published survey at desktop, tablet and phone sizes, can start at any question or in any experimental group, and saves nothing you enter.
How do panel respondents get into the survey?
Through personalized panel links. You name the URL parameter, Choicery stores the participant ID with the response, and you can accept only one completed response per ID. Without the parameter, the general survey link keeps working.
Which analysis should I start with?
Hierarchical Bayes with the default settings. It gives individual utilities and importance, feeds the simulator and WTP, and reports its own convergence. Latent Class, AI-Driven Profile Analysis and Interaction Search are for questions that go beyond individual preferences: segments, the influence of respondent characteristics, and combinations of attributes.
Start with step 1
Create a project, enter your attributes and levels, and read the design diagnostics before anyone answers. The free trial includes every step.