Scent Space
A fragrance finder that matches on what is actually in the bottle, with a hundred fragrances arranged in a space you can rotate, where distance means "smells alike".
Most fragrance quizzes ask what kind of person you are. Confident? Adventurous? Then they hand back a bottle as if the two were connected. Having run the numbers on a real survey, they mostly are not. “Confident” turned up in nearly two thirds of the answers, which means it separates almost nobody from anybody.
So this one asks a different question: what have you liked before. That answer is a fact rather than a flattering adjective, and it is something you can actually compute with.
How it works
Every fragrance is broken into its notes and its accords, then turned into a list of numbers. Base notes count for roughly three times what the opening does, because the opening is gone in twenty minutes and the base is what somebody notices when they hug you at the end of the night.
Compare two of those lists and you get a single number for how alike they smell. Do that for all hundred, across forty-seven houses, and you have the distance between every possible pair.
The map is where it gets fun. There is an old technique called multidimensional scaling that takes a big pile of “how far apart is everything” and works out the arrangement in three dimensions that best preserves it. Feed the distances in, and the fragrances arrange themselves. Then you can grab it and spin it.
Nothing was told where to put anything. The clusters that show up, the fresh citrus corner, the sweet ambery huddle, arrived on their own, which is the main reason to believe the thing is measuring something real.
The honest part
There are two viewing modes, and only one of them is measured. Similarity is built from the note data. The character view, which maps fresh against sweet and light against heavy, is estimated from base-note weight and accords. Nobody sat down with a bottle and a stopwatch and recorded how far it throws. That view is labelled “estimated” on the page, and it should stay labelled until somebody does the smelling.
The note data itself is corroborated across public sources. Fragrantica was deliberately left out: its terms prohibit building a dataset from it, and its robots.txt blocks ClaudeBot by name, so using it would have meant ignoring both.
To check the model rather than trust it, it runs a test on every page load. Certain pairs of fragrances are independently known to smell close to each other, so if the maths is doing its job they should land near each other without ever being told to. Five of the six pairs pass.
The one that fails is the more interesting half of that sentence. It used to sit just outside the cutoff, and then drifted further out when thirteen more fragrances were added, because the new arrivals crowded the same corner of the space. More data made that one answer worse, not better. Worth knowing before assuming a bigger pile of rows fixes anything.
Built the boring way
No framework, no dependencies, no build step. One HTML file, and all the arithmetic runs in the browser in under fifty milliseconds. The whole thing is small enough to open from a folder and read top to bottom, which felt like the right call for something whose main job is to be trustworthy.
There is a test harness too, because every scoring bug found so far was
perfectly valid JavaScript and completely invisible by eye. It pulls the script
out of the page, runs it against a stub browser, and shouts if a NaN or an
unfilled template slot reaches anything a visitor would read.