How this test works

I rebuilt this in 2026 because the old version was making things up. Not deliberately, but the arithmetic had a hole in it, and once I measured how big the hole was there wasn’t much point patching it. This page is what replaced it and how I checked.

The hole in the old one

Every percentage it showed you was a raw score divided by the most you could theoretically have scored on that category. Sounds reasonable. The problem is that those maximums were hand-written and wildly uneven: Bondage could be maxed out from a single question, Dominant needed the top answer on nine.

So I fed it twenty thousand respondents clicking at random. Bondage came out top for 42% of them. Dominant came out top for 0.5%. None of those people had preferences at all. The ranking was reading the answer key, not the person.

Then I checked whether it agreed with itself. Simulating three thousand people answering twice, the top category changed between sittings for 59.5% of them. Nineteen of its thirty-four categories had a retest correlation below 0.70, which is roughly the floor for showing anyone an individual score.

You can run both of those yourself: node tools/simulate.mjs.

The rule I built the replacement on

Every number you see is one of three things. The average of a rating scale whose words you actually read. A normalised comparison between two such averages. Or a z-score against a published mean. If a number can’t be one of those, it doesn’t get shown.

Nothing is divided by a theoretical maximum ever again.

Where the questions come from

Twenty-four of the activities are lifted from the Sadomasochism Checklist, published by Weierstall and Giebel in 2017 under CC BY 4.0. It was built on 652 German-speaking adults, and it is the only reason this test can say anything about how you compare with other people rather than only with yourself.

Everything else is mine. Restraint, personas, service, psychological play, being watched, materials, body focus, ritual, edge play. My questions, my groupings, and neither has been validated against anything. Each area on your results carries a chip saying which of the two it is, because the difference is the whole point and burying it would defeat the exercise.

Four honest deviations from the published original. It’s an English rendering of a German instrument, so what you answered is not the validated instrument. I ask how appealing something sounds where the original asked how much pleasure it gave. By default you see 12 of the 24. And I reworded one item: the published bondage question ends “for total devotion”, which asks about rope and about a motive in the same breath. Somebody who likes rope without the devotional framing would mark it down for the wrong reason, so that clause is gone.

Intensity and breadth

Everything is rated 0 to 4 for how appealing it sounds. Intensity is your mean across the 33 questions everybody sees, divided by four. Breadth is how many of those you rated at least moderately appealing.

That fixed set matters more than it looks. Later questions only appear because you scored high earlier, so counting them would inflate the score for exactly the people who unlock more of them. Choosing the longer form would do it in reverse, because the extra items are the rarer ones. Both are blocked in code and both have a test sitting on them.

Direction

Anything with a direction gets asked twice, once as doing it and once as having it done to you, spread apart in the running order with the two halves in random order. You never pick a side. I work it out afterwards from which one you rated higher.

With a as the active rating and s as the passive, D = Σ(a−s) and M = Σ(a+s):

π = D / (M + 4)

Dividing by M is what stops direction tracking intensity. Multiply every rating by a constant and π doesn’t move. Leave it out and somebody who rates everything 1 with a slight lean is arithmetically identical to somebody rating 4 against 3, which is obviously wrong. The +4 is one imaginary balanced pair, sitting there so a single lopsided answer can’t read as absolute.

Then a range gets computed around π, and the label comes off the range rather than the estimate. Cross the middle and no side gets named. Fewer than two complete pairs, or too little engaged interest, and there’s no number at all.

I also spent a while on question order, which turned out to matter more than expected. Grouping all the doing-it questions together makes the test noticeably easier to answer, and it inflated direction by 46% in simulation, because sitting in a long block of one voice lets a self-image form and you start answering as that person. Short runs of three get most of the comfort and almost none of the inflation.

Areas

Each area shows the plain mean of its questions with an interval around it. What you get depends on how many questions fed it:

People read numbers and skip error bars. I tried a few ways around that and the only one that works is not printing a number you shouldn’t trust. If your answers inside an area scattered badly, the average gets flagged as a poor summary and the individual questions are shown instead.

The comparison with published data

Two quantities get compared with other people, and only two: your totals on the reusable published questions, in each voice, out of 96.

527 of those 652 people were recruited through BDSM communities. It is not a general population norm, and there isn’t one for this scale. The group labels were chosen by the participants at the start with nobody checking. Landing near a group average means your answers looked like theirs. That is the entire claim.

Two of those eight cells sit less than a standard deviation off the floor, where a percentile stops meaning anything, so the page says “at the floor” rather than inventing a number.

On the short form the totals get scaled up cluster by cluster. Against assumed item difficulties that lands about two points high out of 96, almost entirely because the short form leaves out the two most extreme questions. The full 24 removes it. Skip enough reference questions, or miss a whole cluster, and the comparison doesn’t appear.

Things it won’t do

It won’t give you a type. That is the most shareable output available and the one I’d have the hardest time defending, given what the retest numbers looked like on the old version.

It won’t score the six published clusters as subscales, because the study’s own authors say not to. The factor structure differs between the two voices.

It won’t give you a compatibility verdict. The old version printed “Potential Role Overlap/Clash” off a hardcoded threshold of 55, applied to percentages that meant different things in different categories. Two people leaning the same way is a thing to plan around, and a number doesn’t help anybody plan.

It won’t treat a hard limit as a negative score. A boundary isn’t a preference pointing the other way, and putting it in the arithmetic would say it was.

And it won’t show a bar for something it never asked about. “Not asked” is a state you can see, so an absence never reads as a refusal.

What I check before shipping a change

The scoring code touches no page and holds no state, which means it runs against simulated respondents offline. Three things have to pass.

Random clicking mustn’t produce a confident result. Twenty thousand random respondents: no area is the standout more than about a tenth of the time, mean direction is zero to three decimal places, and 97% are told their direction is balanced. None are told they have a strong side.

The same person has to get the same answer twice. Three thousand simulated people answering twice with independent noise: mean retest correlation across areas is 0.92, none below 0.88, direction at 0.96. Not one was told they leaned one way on one sitting and the other way on the next.

And the structural claims have to actually hold. Direction invariant to intensity. Later questions not moving the intensity score. The longer form not moving it either. Nothing under three questions emitting a number. Hard limits never producing a negative score.

Where it still falls short

Direction reads activity rather than identity, so a service top, or a brat running the scene from the bottom, shows up by what they do. Ten of the seventeen areas are unvalidated and are only good for spotting patterns in your own answers. The retest figures above come from simulation rather than real people sitting down twice, so real stability will be lower. And self-report measures what you’re willing to say about yourself today.

The last one is structural. I collect nothing, so none of this can be validated further without changing what the test is. That was a deliberate trade and I’d make it again, but it does mean the unvalidated areas stay unvalidated.

Common questions

Is this test accurate?

Two of the scores are compared against a published sample of 652 adults, so they mean something outside your own answers. The rest compares your answers only with each other, and every area tells you how many questions fed it. No online kink test is clinically validated, and any that claims to be is overstating. What is different here is that the uncertainty is shown rather than hidden.

Is anything I answer stored or sent anywhere?

No. There is no server, no account and no analytics on the page. Answers stay in your browser and are saved only on the device you used. A result is shared only if you generate a link yourself, and the default share link carries scores rather than individual answers.

How long does it take?

Seven to eight minutes for most people. Everyone answers a fixed core of 32 questions; further blocks are asked only where your answers suggest an interest.

Why does it not give me a single type?

Because a single named type is the most shareable output and the least defensible one. In simulation, the previous version of this test changed its top-ranked category between two sittings for 59.5% of people while still printing it as a verdict.

Why does an area say “not enough items to score”?

Fewer than three questions fed it. People read numbers and ignore error bars, so the only reliable way to stop a two-question area being over-read is not to print a number for it.

Can I compare results with a partner?

Yes, by pasting each other’s share links. It happens in your browser and produces a list of where your answers meet. There is no compatibility score.

Does an interest in BDSM indicate a psychological problem?

The research does not support that. Wismeijer and van Assen (2013), comparing BDSM practitioners with a control group, found practitioners scored less neurotic, more extraverted, more open to experience, more conscientious, less rejection-sensitive and higher on subjective wellbeing. Their conclusion was that BDSM is better understood as recreational leisure than as an expression of psychopathology.

Source

Weierstall, R. & Giebel, G. (2017). The Sadomasochism Checklist: A Tool for the Assessment of Sadomasochistic Behavior. Archives of Sexual Behavior, 46(3), 735–745. doi:10.1007/s10508-016-0789-0. Licensed CC BY 4.0. Items were rendered into English, administered as a subset, and the response scale reworded from pleasure gain to appeal. Those modifications are ours, not the original authors'.

Background reading that shaped the design: Alison, Santtila, Sandnabba & Nordling (2001) on the structure of SM behaviour; Joyal & Carpentier (2017) on the gap between interest and behaviour, which is why fantasy and experience are recorded separately; Wismeijer & van Assen (2013) on why none of this is framed as pathology.

Back to the test