Working demo · updated 23 October 2026
A recommender over 87 fragrances. Name the notes you love, type in something you already own, or just describe what you're after in your own words. It matches on composition — and it shows you why it picked what it picked.
Answer as many or as few as you like — nothing is required. The two that move the needle most are the notes you already love and the fragrances you already own, because what you’ve chosen before beats any adjective about your personality. (We tested that: in 32 survey responses, personality words barely separated people at all.)
Every fragrance suspended in three dimensions by its composition. Drag to rotate. Tap or hover any ball for its full note pyramid. Switch between similarity, where distance means "smells alike", and character, where the three axes are fresh–sweet, light–heavy and woody–floral. Tap a family in the key to isolate it.
The clusters here are emergent, not designed. Nothing told the model that the three Acqua di Giò flankers belong together, or that Club de Nuit should sit on top of Aventus — that falls out of the note vectors. Where a cluster looks wrong, suspect the note data before the algorithm.
Known clone pairs act as ground truth: publicly agreed similar fragrances that exist independently of this dataset. If the model can't place each pair next to its partner, it's broken. The test runs live on page load.
Three of the six family colours fall below 3:1 contrast on the light surface, so the palette's relief rule applies — this table is that relief, and it's sortable.