Search first, then decide. Classify the intent before choosing a response format: return results for discovery and comparison, generate an answer for single-fact questions, show a form or account data for transactions, and escalate to a human when confidence is low. Answering everything in prose is the common failure.
| Intent | Example | Right response |
|---|---|---|
| Discovery | black dress for a party | Results grid |
| Known item | silk kurta 42 | Direct result |
| Comparison | wool vs cashmere | Comparison view |
| Fact | what is the return window | Generated answer with source |
| Transaction | where is order 4521 | Account data, not prose |
| Support | this arrived damaged | Human handoff |
Describing six jackets in a paragraph is strictly worse than showing six jackets. Visitors comparing options need parallel presentation — image, price, availability — and prose destroys it. The interface should match the shape of the task, not the shape of the technology.
When the system does generate an answer, link the page it came from. It lets the visitor verify, it protects trust when the model is wrong, and it turns the answer into a route into the site rather than a terminus.