AI Scent Profiling: What It Actually Does (And Why Ours Isn't Magic)

In short

AI cannot smell. It can, however, learn which scent descriptors tend to cluster together in customer preferences and which fragrance families a person is likely to enjoy based on adjacent choices. That's pattern matching, not olfaction. Casa Nochi's scent quiz is a calibrated heuristic built on five questions, mapped to our 10-SKU range, and tuned against actual sales data - not a neural network reading your aura. We'd rather call it a clever questionnaire than oversell it.

The pitch most fragrance startups are making

In the last 18 months, somewhere between eight and twelve startups have launched "AI-powered fragrance matching" platforms. The pitch is almost always the same: answer a few questions, upload a selfie or your music taste, and an algorithm picks your perfect scent. Some promise bespoke blends, some recommend from a catalogue, and one (no names) claims to "decode your olfactive DNA."

It's a good pitch. It is also, almost entirely, marketing dressed as engineering.

Here's what's actually happening under the hood at most of these services, in plain English.

What "AI" actually means in this context

When a fragrance brand says "AI-powered scent matching," they usually mean one of three things, in descending order of how real the AI is:

1. A decision tree. A few branching questions where each answer narrows down a recommendation. This is the most common. It's the same thing a perfume sales assistant has been doing for 80 years, except a website does it now. Calling this AI is generous.

2. Collaborative filtering. The system looks at what other people who answered similarly bought or liked, and recommends from that pool. This is the same algorithm Amazon and Netflix use. It works reasonably well at scale. Below a few thousand customers, the signal is too sparse to be meaningful.

3. An actual model. A small handful of brands have trained machine-learning models on large fragrance databases - combinations of chemical composition, perfumer notes, and consumer ratings - to predict scent preference from input features. This is real ML. It also produces recommendations that are, on testing, often only slightly better than a well-built decision tree.

There is, to our knowledge, no commercial system that can take a photograph of you or a sample of your music and meaningfully predict which fragrance you'll love. Anyone claiming this is selling you a magic trick. The trick can still be fun. It is still a trick.

What AI cannot do, no matter how good the model

The fundamental limit is this: scent preference is shaped by memory, culture, and chemistry - and at least two of those three are inaccessible to any model that hasn't lived your life.

A model can know that you said you like "warm, woody, slightly sweet." It cannot know that the first cedar candle you ever smelled was in your aunt's house in Yekaterinburg in 1998 and you've been chasing that exact note for 25 years. It cannot know that your partner finds tuberose nauseating and so any candle with tuberose will end up in a cupboard. It cannot know that you bought a "warm woody" candle last month and hated it because the "warm" was tobacco and tobacco gives you a headache.

The result is that even the best AI scent system is operating on maybe 30% of the information that would actually predict your satisfaction. The other 70% is in your head, your home, and your nose - and the only way to surface it is to ask better questions, not run a bigger model.

What Casa Nochi's quiz is actually doing

Our scent quiz is a five-question calibrated heuristic. There is no neural network. There is no model. There is a decision tree built by humans (Pavel and a perfumer we work with), tuned against our actual returns and reorder data, and mapped to the 10 SKUs we make.

Here's the structure, briefly:

Q1: Time of day

When do you most often light a candle - morning, afternoon, evening, or late night? This is a stronger signal than people realise. Citrus candles get bought by morning people. Smoky candles get bought by 10pm people. If you tell us morning, we will not steer you to Amber Nochi.

Q2: Fragrance family preference

Floral, woody, gourmand, smoky, or "honestly, I have no idea." The last option is treated as a real answer - it routes to a starter trio of crowd-pleasers rather than forcing a guess.

Q3: Existing scent reference

We ask which perfume or candle you already love, with a few prompts (Diptyque Baies, Le Labo Santal 33, Jo Malone Wood Sage, etc.). This single question is worth more than all the others combined - if you already love Black Orchid, we're going to suggest Noir Orchid, and we're going to be right about 80% of the time.

Q4: What the room is

Living room, bedroom, bathroom, study. Different volumes and use cases warrant different throws. A study candle and a bedroom candle behave differently because you want different things from them.

Q5: A negative

Which scent family do you dislike? This question is rare in scent quizzes and we think it's the most useful one we ask. Avoiding a hate is more important than chasing a love when the catalogue is only 10 SKUs.

That's it. Five questions. Output: one primary recommendation and two adjacent suggestions. We tested longer versions and they performed worse - people abandoned at question seven, and the additional data didn't materially improve the recommendation against our sales-tuned tree.

Where this beats AI, and where it doesn't

The quiz beats a generic AI fragrance recommender on three axes:

• It's tuned to our actual catalogue. Generic AI tools recommend from a database of thousands of fragrances, most of which aren't available to you. Ours recommends from 10 candles you can actually buy this afternoon.
• It's honest about what it doesn't know. We don't claim to read your soul. We claim to keep you off the wrong candle.
• It accounts for the negative. Most AI recommenders only optimise for "you'll like this." Excluding what you'll hate is half the job.

Where a more sophisticated AI system would do better: if we had 100,000 customers and rich purchase histories, a collaborative filter could likely find non-obvious matches our tree misses. We don't have 100,000 customers. When we do, we'll consider it. For now, a five-question tree maps a 10-SKU range better than any model would.

How to get the most out of the quiz

1. Answer the "what do you already love" question honestly - even if the reference is from a brand we don't compete with.
2. Take the "what do you hate" question seriously. Don't be polite about florals if florals give you a headache.
3. If the result feels wrong, ignore it. The quiz is a starting point, not a verdict.

What this means for Casa Nochi

We think a lot of the AI-in-fragrance pitch is overclaim, and we'd rather be the brand that doesn't overclaim. The scent quiz is a tool. It's a useful tool. It is not magic, and it doesn't need to be - because the underlying offer (10 SKUs, £29.99 each, 50-hour burn, free UK shipping over £40, 30-day returns) means the cost of guessing wrong is genuinely low.

If you'd rather skip the quiz entirely, the discovery bundle is the other way to find your match - three small-batch pours, much smaller commitment, and your own nose doing the work the algorithm can't.

FAQ

Can AI actually pick the right candle for me?

It can narrow the field. It cannot predict love. The best AI fragrance recommender in the world is still working with limited signal - your memory, your home, your partner's preferences are all invisible to it. Treat any "AI scent match" as a starting point, not an answer.

Is Casa Nochi's quiz powered by AI?

No, and we don't pretend otherwise. It's a five-question decision tree built by humans, tuned against our sales and returns data, mapped to our 10-SKU range. We'd rather call it a clever questionnaire than oversell it.

Why only five questions?

Because longer quizzes lose people, and the additional data didn't measurably improve the recommendation in our testing. Five questions is the sweet spot where you give us enough to be useful and we don't waste your evening.

What if the quiz recommends a candle I end up hating?

You're covered by our 30-day return policy on unopened items. We also recommend the discovery bundle for first-time buyers - three small pours so your own nose does the work the algorithm can't.

Will you launch a real AI fragrance tool?

Probably not. The honest version of that tool would still be a decision tree dressed up. The dishonest version would be marketing. Neither feels right for a 10-SKU brand built around two worlds and one match.

---

If you want to try the quiz, it's at /quiz. If you'd rather just be told where to start, Amber Nochi - honey, tobacco, smoky cedar - is the candle most people stay with.