Visualising fragrance online
In 2025 Kateřina Milatová asked real people which visual cues help them imagine a perfume they cannot smell, and measured the answers. We took five of her findings and built them across every perfume we hold. We refused two more, on her own evidence. This page shows which is which, how each one is built, and the warning of hers we took most seriously.
On this page
The research
What Does a Scent Look Like? Visualising Fragrances On E-commerce Websites is Kateřina Milatová's bachelor's thesis in New Media Design at Jönköping University, June 2025. It combines three things. A systematic literature review. A focus group of 6 people who drew and mood-boarded blind-sampled perfumes, then ranked 15 possible scent cues. And an A/B usability test with 8 participants, comparing a standard shop layout against a visualisation-rich prototype.
Two things make it unusually useful. First, it measures rather than asserts: her enhanced prototype scored 43.5% higher on helpfulness, made imagining the scent 43.6% less difficult, improved accuracy of choice by 33.3%, and lifted overall user experience by 60.7%. Six of the eight said they would prefer it when buying an unknown perfume, and the other two would take either. Second, it is honest about the risk in its own result, which we come back to below.
What we implemented
| Her finding | What we built | Our data |
|---|---|---|
| The top, heart and base pyramid is outdated. She replaced it with a horizontal evolution timeline of fading curves - voted the 2nd most helpful cue overall, and valued especially by novices | A timeline where each note is a bar whose colour fades in and out as the note rises and dies, with its image at the moment it peaks | Each note's shape from the substantivity we hold for it in hours, falling back to its volatility class where we hold no value, scaled by the note's weight |
| Notes shown as an image plus a word tag rated 1.7 times more helpful than the same notes as a text list | Every note carries its ingredient image, sized by how much it dominates, sitting on the timeline at its peak | The ingredient images we generate for each note, sized by the note's weight in the perfume |
| Evocative elements - simplified familiar things the scent reminds you of - topped the ranking in both the focus group and the usability test. Her focus group scored 15 attributes and put this first on 85 points; gender and notes written as text came last on 43 | A "what it smells like" row of the four dominant notes, shown large and simply | The four highest-weighted notes that carry an image |
| Asked to draw a scent, people drew scenes - 16 of the 24 focus-group drawings were places, not abstract shapes | A short "feels like" scene cue - a season and an occasion | The perfume's strongest season and occasion scores |
| People reason in opposites - fresh against sweet, light against heavy, summer against winter - and their confidence matched reality only when identifying what a scent was not | Three axes with a plain-English reading that leads with what the perfume is not | Our stored olfactory axes for fresh-sweet and light-heavy. Our season scores for summer-winter |
Scroll the table sideways to see every column
What we deliberately did not build
She found gender an unreliable signal, because her participants routinely mislabelled it. She also found that fragrance family is a poor proxy for similarity: when asked to pick the most similar pairs, only 1 of 3 pairs shared a family.
So there is no masculine-to-feminine axis on our page. There is no natural-to-artificial axis either, because we hold no honest measure of it. And where we group the timeline by family we use it only to order the bars, never as a claim that same-family perfumes smell alike. Implementing a finding means honouring the negative results too.
Her participants also reached for scenes, objects and personas rather than abstraction - not one of the 24 drawings in her focus group was abstract, and 16 of them were scenes. So these concrete cues stand on their own here, with nothing abstract standing in for them.
The warning we took most seriously
Her most important finding is the one that cuts against her own result. Confidence rose in almost every task, and accuracy did not always follow: in the ranking task the plain layout was slightly the more accurate of the two, while hers raised confidence by 27.9%.
The design contributed to perceived understanding rather than factual correctness
One participant put it precisely. With the plain layout she knew her mental image might be wrong, and braced for it. The vivid version gave her such a clear idea that a mismatch with the real bottle would be more misleading. Milatová warns this can buy a short-term sales bump at the cost of long-term trust.
We sell nothing directly, but the risk is ours all the same. So the timeline is drawn from data we can point at rather than art direction, and the opposites readings lead with what the perfume is not. We would rather be trusted than convincing.
One more thing we were careful about. Every curve here is drawn from stored numbers we can point at, and nothing on the timeline is a model's guess at what a smell looks like. The ingredient images are made by an image model, but only ever to depict the named ingredient itself. We ask it for a lemon, never for the smell of one. There is more on that distinction on our page about ingredient pictures. It matters because her participants found concrete, checkable cues more helpful than decorative ones, and a generated image sits close to that line.
The hours behind every curve are estimates, not measurements. Nothing in our catalogue was timed in a laboratory. Each figure was produced by the model that classifies our notes, working from the material class. So it restates whether a note is a top, a heart or a base, in finer grain. It is not an independent reading. About a third of the values land on one of the round numbers used as worked examples in that prompt. Roughly a quarter of our notes carry no value of their own at all, and fall back to the median for their class.
We are saying so here because this is the page that warns about confidence running ahead of accuracy. A smooth curve is exactly the kind of picture that looks measured, and a reader has no way of telling from the drawing that it is not.
See it
Six perfumes spanning fleeting to long-lasting, each rendered live from its own data.
- Mugler Angel - a fruity gourmand that lasts all day
- Tom Ford Black Orchid - a dark floral with staying power
- Yves Saint Laurent Opium - a spicy, musky oriental
- Loewe Aire - a green floral over a moderate wear
- Acqua di Parma Colonia - a bright citrus cologne, gone by mid-afternoon
- Guerlain Aqua Allegoria Rosa Rossa - a fleeting fruity floral
Every perfume on ScentVerdict has one - append /evolution to any perfume URL.