Your AI chat isn't bad at recommendations. It's bad at confidence.

The same information, presented differently, produces different decisions. Wine tastes better poured from a heavy bottle, and a painkiller works better when you think it was expensive. The product never changed; the framing did. Behavioural economists call it the presentation effect.

Most AI shopping chats give the effect away for free. The model finds genuinely good products, then the interface drops them into a flat paragraph or an even flatter grid: twelve items, no order, no hint of why any of them made the list. You’re left to rank a dozen things the model already ranked.

Give every pick a reason

The fix for a chat that looks dumb is almost never a smarter model. Take the same results and show the strongest three as a ranked list, each with a one-line reason: this one because it packs light, this one because it’s the cheapest thing that’ll last. Nothing about the output changed, but it now reads as a judgement instead of a dump.

Ellen Langer’s photocopier study is why this works. People let a stranger skip the queue far more often when the request includes the word “because”, even when the reason that follows is hollow (“because I have to make copies”). The justification carries the weight, not its content. A recommendation with a reason attached gets the same benefit of the doubt.

Everything else goes in a secondary grid beneath the three. The placement is the argument: the three reasoned picks become the reference point, and the grid quietly reads as also good, but not what I’d choose. You never have to write that line. The hierarchy says it for you.

The opening line decides who is the expert

A shopping assistant usually opens with a blank box and a friendly prompt: “What are you looking for today?” It feels welcoming, and it hands the hard part straight back to you. You came to be advised, and the first thing the chat does is ask you for the answer.

A good salesperson does the opposite. They ask one sharp question, or they put something in your hands and watch your face. The opening move assigns the roles. “What are you looking for” casts you as the one who already knows; “Who’s it for?”, or two confident opening picks, casts the assistant as the one who knows. Same screen, very different relationship.

Rory Sutherland calls this reframing: changing the meaning without touching the substance. A first-class lounge is the same building as the departure gate, with much the same chairs; what you pay for is the feeling of being someone who belongs in it. An assistant that leads instead of interrogating isn’t any smarter. It’s just sitting in the expert’s chair.

Make every pick buyable in place

Here’s a quieter mistake that costs more than it looks. The chat names a winner, explains why it’s right, gets you genuinely ready to buy, and then doesn’t link to it. Now you have to leave the conversation, search the app, and find the product yourself.

Every step between intent and action leaks, and the moment of highest conviction is also the moment of lowest patience. Break the flow there and you don’t get a slightly worse conversion rate; you get an abandoned session. So every product the chat names has to be tappable inside the message itself, not waiting on the next screen.

Forget the one-off purchases

Extend a fashion chat to cover gifts and homeware and a trap opens up. Someone buys a single birthday candle, the system dutifully learns the preference, and for the next month it keeps suggesting candles. The algorithm is doing exactly what it was told to, which is the problem.

Good personalisation needs selective amnesia. A gift should shape the current session and then drop out of the long-term profile, and the same goes for one-off searches, seasonal buys, and anything bought for someone else. Products feel creepy when they hold on to the wrong things and won’t let go. People clear their own browser history for a reason; a considerate system clears it for them.

Trust builds slowly and breaks fast

Trust in an AI product builds up over dozens of small, good interactions and gets spent in a single bad one. A voice assistant that keeps listening after you’ve hung up. A recommendation that won’t take a tap. Neither is a failure of the model.

They’re failures of the experience, and they happen when the team building the model and the team building the interface don’t agree on what “good” means at the point where a person actually uses the thing. The model was never the hard part. Making its answers feel usable is.