Every shop you visit now claims to know you. Feeds reshuffle themselves around your last click, emails arrive addressed to a version of you assembled from browsing history, and somewhere a model is quietly estimating how much you would tolerate paying. The technology is genuinely impressive. What it has not solved is loyalty.
We think that gap matters, because the two are often confused. Prediction is about the next transaction. Loyalty is about the hundredth one, and about whether a person feels respected in between.
Prediction is not the same as understanding
A recommendation engine is very good at noticing patterns and very bad at knowing why they exist. It sees that you looked at three coats and infers that you want a fourth. It does not know that you were buying for someone else, that you already found what you needed, or that you were simply curious on a slow evening.
So the model keeps pushing coats. Multiply that across a whole storefront and personalisation stops feeling like service and starts feeling like being followed. The failure is not the maths, it is the assumption that more signal always means more relevance.
The quiet things that actually earn trust
When we ask ourselves why anyone would come back, the answers are unglamorous. Prices that do not shift depending on how badly you seem to want something. Delivery estimates that hold. Returns that take one step, not five. A product description that mentions the drawbacks, so you can decide instead of guess.
None of that requires a model. All of it requires a decision to be consistent even when a short-term nudge would earn more this quarter. Loyalty is mostly the accumulated evidence that a shop behaves the same way when it is inconvenient.
Personalisation done with restraint
Used carefully, knowing something about a shopper is a kindness. Remembering a size so the wrong one never gets suggested. Surfacing the three things worth seeing in a category of three hundred. Staying quiet when there is nothing genuinely new. Research on interface design has argued for decades that recognition should beat recall, and that principle applies as much to a storefront as to software.
The test we try to hold ourselves to is simple: would this feel helpful if the shopper could see exactly why we did it? Reordering a category by relevance passes. Manufacturing urgency does not.
What we are choosing to optimise
Attention is finite and increasingly contested, which is a problem we have written about in why calm tends to beat loud online. Optimising for it aggressively works, briefly, and then people leave. Optimising for whether someone found what they came for is slower and compounds.
In practice that means a smaller edit rather than an endless catalogue, and honest reasons for why something is in it. Whether you are looking through everyday wardrobe pieces or the accessories that finish them, the promise is the same: fewer, better, explained.
The unfashionable conclusion
The most durable version of loyalty is not a points balance or a tier badge. It is the small, boring certainty that a shop will not waste your time or exploit your impatience. Personalisation can support that. It cannot substitute for it.
If we get this right, the model in the background becomes invisible, and what you notice instead is that shopping took ten minutes and you are happy with what arrived. That is the whole ambition.
