Conjoint Experiment Simulator
Take a real choice-based conjoint study, then analyse the data it produces — with your own answers sitting inside it.
This is the product
A reusable water bottle. Drag it to turn it, scroll to zoom — this is exactly what you will be choosing between.
Every bottle in the study is this same object with its features changed: a different lid, a different body, a different material, a different price. Nothing else about it moves.
What you are about to do
You will see 18 choice tasks. Each one shows three reusable water bottles and asks a single question: which would you buy? You can also say you would buy none of them, and sometimes that will be the honest answer.
Nobody will ask you how important price is, or how much you like a straw lid. That is the entire point. Conjoint analysis works out what each feature was worth to you by watching the trade-offs you make when you cannot have everything.
When you finish you become respondent R100 in a dataset of 100 people. The other 99 are simulated. You will download that file and run it through the Conjoint Analysis tool, where you can find your own row and compare yourself to everyone else.
The four attributes
Three levels each. Everything else is held constant — the same brand, the same camera angle, the same lighting on every bottle — so the only things that vary between two options are the four features being tested.
| Attribute | Level 1 | Level 2 | Level 3 |
|---|
The level marked reference is the baseline the analysis measures the other two against. That choice is arbitrary and does not change what the study finds — it only changes which number is zero.
How the cards were built
Four attributes at three levels each gives 34 = 81 possible bottles. Judging all 81 would take far longer than anyone will sit still for, and most of the work would be wasted. A designed experiment gets the same answer from a fraction of it.
The first bottle in each task comes from an L9 orthogonal array — nine rows chosen so that every pair of attributes appears in every combination equally often. The second and third bottles are built by rotating every attribute one notch: level 1 becomes 2, 2 becomes 3, 3 becomes 1. Do it twice and you have three alternatives that never share a level.
That single trick buys three things at once: every level is shown equally often, no attribute repeats inside a task, and no attribute is entangled with any other. Run it again with prices stepped one level along and you get the second block of nine.
Check the balance yourself
Counted from the actual design you are about to see, not typed in by hand:
Price rode along for free. An L9 array carries four three-level factors in nine rows, and lid, body and material only used three of its columns — the fourth was sitting empty. Adding price cost no extra cards at all.
Why “I would not buy any of these” is on every task
A study that forces a choice can only ever tell you which bottle wins. It cannot tell you whether anyone would have bought a bottle at all. Real customers walk away, and a forecast that cannot represent walking away will overstate demand for everything.
So use it honestly. If all three look overpriced or unappealing, say so — that is information, not a failure to answer.
Ready
Advanced settings
Eighteen choices, about five minutes. There are no right answers — answer as yourself, and turn the bottles if it helps you decide.
Choice 1
Which of these would you buy?
Drag any bottle to turn it. All three turn together, so you are always comparing them from the same angle.
Keyboard: 1, 2 or 3 to choose a bottle, 0 for none. Tab to a bottle and use the ←→↑↓ keys to turn it.
Done — you are respondent R100
Here is a plain count of what you picked. It is not your part-worths: counting how often you chose a feature is not the same as working out what that feature was worth, because the features you picked came bundled with others. Separating them is the job of the analysis tool.
You chose not to buy in — tasks.
Your part-worths, against the other 99
Now the real thing. A choice model has been fitted to your 18 answers, and separately to each of the 99 simulated respondents, then averaged. These are part-worths: what each feature was worth to you, worked out from the trade-offs you made rather than from anything you were asked directly.
Both sets are scaled to 100 points of total swing. Raw choice-model numbers carry an arbitrary scale that reflects how consistently someone answered, so without that rescaling this would partly be measuring decisiveness rather than taste.
Treat your own numbers gently. Eighteen choices is a thin basis for eight part-worths, so yours are far noisier than the market average, which pools 99 people. If one of your bars looks extreme, that is as likely to be sampling noise as a strong opinion — which is exactly why real studies recruit hundreds of respondents instead of asking one person hundreds of questions.
This is the small part
What you see above is one panel of a conjoint analysis. The tool it came from also runs market simulations against an outside option, converts part-worths into willingness to pay in dollars, finds latent segments, and lets you change a product's price and watch demand move. None of that fits here — so take the data across.
Take your data to the analysis tool
-
Download the dataset. It holds your 18 answers plus 99 simulated respondents —
7,200 rows in the long format the tool expects.
- Open the Conjoint Analysis tool and upload the file.
-
Set the “None” alternative. In the attribute setup step,
choose
Nonein the None alternative dropdown. The tool then gives the walk-away option its own constant instead of treating it as a bottle with missing features. - Mark Price as numeric if you want willingness-to-pay in dollars; leave it categorical if you would rather see whether the jump from $22 to $30 hurts more than the jump from $14 to $22.
- Find R100 in the individual-level results. That row is you.
Then check the segmentation against the truth
The tool can group respondents by running k-means on their estimated part-worths. It will find some segments. The obvious question — and the one a cluster chart can never answer on its own — is whether they are the right ones.
The 99 simulated respondents were drawn from three deliberately opposed groups, and this file says which group each of them actually came from. Cluster first, then open it and compare.
Keep this file out of the analysis tool. Every column it does not recognise is treated as a feature of the bottle, so a segment column loaded alongside the choice data would be read as a product attribute.
Expect the recovery to be imperfect, and do not treat that as a failure. Eighteen choices give a noisy estimate of one person's preferences, so some respondents will land in the wrong cluster. How many, and which ones, is the interesting part: the people who are hardest to classify are usually the ones whose preferences sit between two segments rather than squarely inside one.
A fair warning about your own row: 18 choices is not many for estimating eight part-worths. Your individual estimates will be wobblier than the averages across all 100 respondents. That gap is itself worth understanding — it is exactly why commercial studies recruit hundreds of people rather than asking one person a hundred questions.