The assistant can only be as specific as your data
If a product page says “premium materials” and nothing else, no assistant can tell a shopper whether it survives a dishwasher. A vague answer is a faithful reflection of a vague source, and the fix is upstream of the conversation.
Attributes beat adjectives
Shoppers make decisions on facts: dimensions, weight, capacity, power, fabric, compatibility, care. Marketing language belongs in the description; the decision-making facts belong in structured fields where they can be compared, filtered and quoted accurately.
Write the variant names for humans
“Option 1 / Option 2” is a data-entry shortcut that costs you conversions for years. A variant a shopper cannot identify is a variant an assistant cannot recommend with confidence, and a support message waiting to happen.
Policies are product data too
Shipping, returns and warranty questions arrive inside product conversations, not in a separate policy conversation. If the assistant reads your catalog but not your policies, half of what shoppers ask is out of reach.
Fix the source, then retest the same question
The loop that actually improves things: find a weak answer in a transcript, identify the missing fact, add it where the assistant reads from, then ask the identical question again. Two or three of these a week compounds faster than any prompt tuning.
Explore this workflow in Dejor.
See how shopify product recommendations fits into your store, including setup steps and common questions.