Good AI recommendations can still look arbitrary

We had a shopping chat that returned good products. The problem appeared when someone asked a simple question: why these? The interface had no answer. It placed everything in one flat list, so a considered recommendation looked much like a search result.

Show the judgement

The strongest three results became a ranked group. Each received one plain reason: it packed light, it was the least expensive option that would last, or it matched a constraint mentioned earlier. The remaining results stayed available below.

The model output had not improved. The interface had started exposing the selection that was already happening. That was enough to make the recommendations easier to compare and easier to challenge.

Let the assistant do some of the work

A blank prompt saying “What are you looking for?” appears open, but it asks the user to structure the problem before the assistant has helped. A sharper first question, such as who the item is for, narrows the space without demanding a complete brief. In some situations, a couple of initial picks work even better because the person can react to something concrete.

Keep the action next to the recommendation

Originally, the chat could name a product and explain why it was right without making that product tappable. The person then had to leave the conversation and find it again elsewhere in the app.

We put the product card inside the message. When the answer creates intent, the next action should not require a second search.

Remember less

A purchase for someone else should not quietly become part of the user’s long-term taste profile. Gifts, seasonal searches and one-off needs can improve the current conversation and then expire. Memory is useful only when the product is selective about what deserves to remain.

None of these changes required a new model. They required the product to show why an answer was chosen, lead with an appropriate question, keep the next action close and know what not to remember. The quality of an AI feature is partly the quality of that framing.