I’ve been a digital designer for about twenty years. Stylin came out of something I kept watching happen: people with completely full wardrobes, standing in front of them, saying they have nothing to wear — and then going out and buying more clothes. The problem isn’t that they don’t own enough. It’s that they can’t see what they own as combinations. That’s a styling problem being solved as a shopping problem.

What’s different is what happens when the app doesn’t know what to suggest. Nearly everything in this category is a funnel to a purchase — the wardrobe features exist to make the shopping feel personalised. Ours runs the other way round. Stylin starts with what’s already hanging in your closet, and only points you at something to buy when there’s a real gap, like a neutral layer that would unlock ten outfits you can’t currently make.

The other difference is format. Most apps give you a search box and wait. We publish you a daily issue — a personal fashion magazine styled from your own clothes. You open it the way you’d open a magazine, not the way you’d open a database.

Two places: App Store discovery, and AI assistants. The second one still surprises people when I say it out loud. We get a continuous flow of users arriving from LLM-powered apps — someone asks an assistant what to do about a wardrobe full of clothes they don’t wear, and we’re in the answer. That’s not a trickle we’re hoping turns into something. It’s a channel.

Where I wish it came from: other users. Getting dressed is social — people show each other outfits constantly, that’s the entire basis of the fashion internet. An app that styles your own clothes should spread the way a good recommendation spreads, and we haven’t earned that yet. Discovery traffic is someone with a problem looking for a solution. A referral is someone who liked it enough to say so unprompted. The second is worth far more, and we don’t have enough of it.

Handed over: watching the industry. Brand tracking, reading trends, working out where the major fashion houses are actually heading — that’s a machine job now. It’s breadth work. Nobody on a small team can go through that many collections, that many brands, that consistently, and stay current. A model does it without getting bored, and it’s genuinely better than we were.

Took back: the principles. What actually defines a style. What suits whom, and why. Which body shape wants which silhouette, what level of contrast works against someone’s colouring — the rules underneath a good outfit rather than the surface of it. We tried to let the model derive those and it produces something that sounds authoritative and is subtly wrong. It has seen every outfit ever photographed and never watched one person get dressed.

So the split ended up being clean: AI is very good at what — what’s happening, what’s selling, what everyone is wearing. It’s unreliable at why. And the why is the part that decides whether a person opens their wardrobe in the morning and feels fine.

Yes — we see a continuous flow from LLM-powered applications. Not a spike, not a novelty, a steady stream. For us this isn’t a thing that might happen in a couple of years. It’s already one of the two ways people find us.

What’s changed just as much is what people know before they arrive. They turn up having already had the category explained to them by an assistant, knowing roughly what the app does and how it differs from a shopping app. We used to spend the first screen explaining the concept. Now we’re mostly confirming something they’ve already been told.

We built for this deliberately. Our site publishes a machine-readable licence saying AI systems may quote and cite our content with attribution, and may not use it to train models — terms written for assistants rather than for people. And a lot of our writing is structured as the direct answer to the question someone actually asks, because that’s the shape of thing an assistant can lift and cite. The goal is to be the sentence a model quotes, not the tenth blue link.

The honest caveat is that attribution is murky. Someone who asks an assistant, then searches our name, arrives as branded search — so the real number is almost certainly higher than what we can see.

Do: get findable by machines, not just by people. Publish the direct answer to the question your customer actually asks, in plain language, on a page a model can read and cite. Put your facts somewhere structured. Most people in fashion are still writing for a search ranking, and the thing worth optimising for now is being quoted. I’d treat that as urgent rather than interesting — it’s already one of our two main channels, and I don’t think most of our category has noticed yet.

Also do: work out which part of your job is taste and refuse to automate it. Ours is styling judgment. Yours is something else. The stuff around it — cataloguing, tagging, recognition, the grinding data work — hand all of that over without sentiment.

Ignore: the chat box reflex. Not everything is a conversation, and in a visual category it’s actively worse. Ignore novelty as a metric — novelty buys a first session and nothing after it. And ignore the pressure to put "AI" in front of what you sell. Our customers don’t want AI. They want to be dressed and out of the door in four minutes. The model is our problem, not theirs.