Customer experience optimization
The biggest driver of customer experience quality in ShopGuide isn't tone or personality — it's whether the AI's answers are actually grounded in your real catalog and store content, and whether it knows when to hand off instead of guessing. This guide covers the parts of the experience that are genuinely configurable, rather than generic tone advice that applies to any chatbot.
Grounded answers over generic ones
Product and shop-knowledge answers come from Qdrant vector search over your catalog plus RAG over your store's content — the AI is retrieving and summarizing what's actually indexed, not generating claims from general knowledge. That means the fastest way to improve perceived answer quality is improving what's indexed (complete product descriptions, current shipping/return policy content, populated metafields), not adjusting the system prompt's tone instructions. A confident-sounding wrong answer is almost always a catalog/content coverage problem, not a personality problem.
If your store has custom prompt requirements beyond the standard flow — a fully bespoke funnel, restricted tool set, or forced tool-calling behavior — that's supported via custom system prompts. See Custom system prompts.
Knowing when to hand off
SUPPORT is a distinct intent from ORDER_MANAGEMENT: order tracking, status checks, and address/quantity changes stay with the AI, while complaints, cancellations, and refund requests route to support handling rather than the AI attempting to resolve them itself. This split exists because those situations usually need a human decision (a refund, a policy exception) that the AI shouldn't be making unilaterally. If your team is seeing AI attempts to handle cancellations or complaints directly, that's worth flagging as unexpected behavior rather than assuming it's by design.
Consistent guided decisions, not open-ended browsing
For shoppers who describe a need rather than a product, GUIDED_DISCOVERY asks up to three short, structured questions (three answer options each) before recommending — a consistent, bounded interaction rather than an open-ended back-and-forth that can go anywhere. This is deliberately narrow: it's meant to converge on a recommendation quickly, not to simulate an extended conversation. If you want that structured decision experience to trigger more often, make sure your product descriptions carry enough distinguishing detail (use case, fit, skill level) for the follow-up questions to matter — see Chat optimization.
Language coverage matches only part of the experience
The AI's actual replies adapt to the shopper's detected language regardless of store configuration. What's fixed is the widget's static UI chrome — buttons, guided-discovery labels, footer attribution, and the disclaimer — which only has translated strings for English, German, and Dutch. A shopper in a fourth-language market gets fluent AI prose next to English-language UI chrome. If you operate in one of the three supported languages, confirm your shop content language metafield is actually set correctly — a mismatch here is a common source of an inconsistent-feeling experience. See Multi-language support.
Closing the loop on what's actually confusing customers
Information Gaps Detection runs after every response and classifies what the AI couldn't resolve — including cases where the app itself misbehaved (system_error: a failed tool call, a wrong-language reply) separately from genuine content gaps (knowledge_gap) or issues with your store (bug). Reviewing the Knowledge Gaps tab in your transcript viewer regularly turns "the chat feels off sometimes" into a specific, prioritized list of what to fix and who owns it. See Chat management.
Next steps
A trustworthy chat experience comes from grounded, catalog-accurate answers and clear escalation boundaries — not from a friendlier-sounding prompt.