A recommendation should survive being explained at the till

Every product team in retail is now asked to personalise. The tools have become cheap, and the case for them is persuasive: a customer who sees products suited to her size, her climate and her past purchases is more likely to buy. At Sablemere we personalise too. But I have come to judge each personalised element by a test that has nothing to do with its contribution to basket size. Could a good store associate say the same thing to the customer's face, and explain why? If not, it does not belong on the screen.

The test came from watching our stores. A good associate in our Calgary store remembers that a customer bought a medium last winter and found the sleeves short, and suggests the tall fit this time. She notices a pack too heavy for the trip described and says so. She does not tell a customer that only two of a jacket remain when the stockroom holds twenty, and she does not follow a customer to the door suggesting that people who care about the outdoors buy the more expensive model. Customers trust her because what she says is true, relevant and offered for their benefit as well as ours.

The associate test

Applied to a website, the test removes more than one might expect. A recommendation based on size and fit history passes: an associate would say it and could explain it. A suggestion of a compatible insulating layer for a rain jacket passes. A note that a boot runs narrow, drawn from our return reasons, passes, and it has reduced returns. A banner telling a returning customer that a product she viewed has 'sold quickly' fails, unless we can show the rate and the remaining stock. Re-ordering search results so that higher-margin products rise, without telling the customer, fails, because no associate would say 'I am showing you this first because we earn more on it'.

The test also governs how the model reasons, not only what it shows. Our personalisation team uses purchase history, sizes, the region a customer gives us and what she has browsed in the current visit. We decided not to infer income, family circumstances or health from purchases, even where a model might find the patterns useful, because no associate would say aloud 'you seem to have had a baby' or 'you can afford more'. Québec's amended privacy law requires us to tell customers when we use technology capable of profiling them and how those functions are activated, and we apply that standard across Canada. The legal obligation is a floor. The associate test is the standard we chose above it.

Each recommendation now carries a short reason, written in plain language in English and French: 'Because you bought this in a medium', 'Often worn under this jacket', 'Our repair centre suggests this for boots like yours'. Writing those reasons was more work than building the model. It forced the team to decide which signals it would be willing to name, and several were abandoned at that point.

We also ask associates to judge the reasons. Twice a year a group of twelve associates from stores in four provinces reads a sample of recommendations and marks any they would be embarrassed to say aloud. In the first review they flagged a suggestion that paired children's rain gear with almost any adult purchase made in spring, which looked to them like guessing at a family. We removed it the following week, and the review has since become part of how a new model is approved.

Urgency we cannot prove

The commonest dark patterns in retail are claims of scarcity and urgency: countdown timers, low-stock warnings, notices that other people are looking at the same item. They work, which is why they persist. They also teach customers to distrust every message on the page, including the true ones.

We removed them in 2022. Countdown timers went entirely, because our sales end when we say they end and a timer adds nothing honest. Low-stock warnings remain only where the figure comes from a location whose counts meet our accuracy threshold, and they state the number. Notices about other shoppers went, because we could not show that they told the customer anything useful. Messages that made declining an offer sound foolish, such as 'No thanks, I prefer wet feet', were rewritten as plain choices.

The commercial team measured the effect closely, as it should. Conversion on promotional pages fell by about 2 per cent in the first month and recovered over the following quarter. We cannot prove that the recovery came from trust rather than from other changes we made at the same time, and I would not claim that it did. What we can say is that complaints about misleading claims fell to almost nothing, and that our associates stopped meeting customers who had been told online that something was nearly gone.

Switches a customer can find

Personalisation should be something a customer can see and adjust without hunting for it. Every recommendation panel on our site has a control beside it: 'Show me fewer suggestions like this', or 'Stop using my purchase history'. The controls sit on the page where the recommendation appears, not three levels down in an account menu. About 6 per cent of signed-in customers have used them. Some switched personalisation off altogether; more adjusted one kind of suggestion they found unhelpful, which told us more about the model's weaknesses than any test we ran.

The same principle governs our paid care membership. A customer can join from a product page with one choice, and since 2022 she can leave from her account page in two steps. We do not ask her to telephone, and we do not dangle a discount at the moment of cancellation unless she asks for one. The membership is smaller than it was when people were enrolled by default, but those who remain use it.

None of this makes personalisation less useful. It makes it legible. A recommendation that can be explained is one a customer can correct, and a correction is better evidence than a click. A retailer that asks its models to meet the standard of its best associate will build fewer clever features and more trustworthy ones, and its stores will not have to apologise for what its website said.