Customer experience optimization
Tone advice for chatbots is everywhere, and most of it doesn't apply here. What actually drives ShopGuide's experience quality is whether its answers are grounded in your real catalog and content, and whether it knows when to stop and hand off instead of guessing. This page sticks to what's genuinely configurable.
Grounded answers beat generic ones
Product and shop-knowledge answers come out of Qdrant vector search over your catalog and RAG over your store's content — the AI is retrieving and summarizing what's actually indexed, not drawing on general knowledge. That means the fastest way to improve how good the answers sound is improving what's indexed: complete descriptions, current policy content, populated metafields. A confident wrong answer is almost always a coverage gap, not a personality problem, and no amount of prompt tone-tuning fixes it.
If your store needs something beyond the standard flow — a bespoke funnel, a restricted tool set, forced tool-calling — that's handled through custom system prompts. See Custom system prompts.
Knowing when to step back
SUPPORT and ORDER_MANAGEMENT are deliberately separate intents. Order tracking, status checks, address and quantity changes stay with the AI; complaints, cancellations, and refund requests route to your support handling instead of the AI trying to resolve them itself. The reasoning is straightforward: those situations usually call for a human decision — a refund, a policy exception — that the AI shouldn't be making on its own. If you're seeing it attempt cancellations or complaints directly, flag that as unexpected rather than assuming it's by design.
A bounded decision, not an open conversation
For shoppers who describe a need rather than naming a product, GUIDED_DISCOVERY asks up to three short questions, three options each, before recommending anything. It's deliberately narrow — meant to converge on a recommendation fast, not simulate a long back-and-forth. If those conversations tend to end in generic browsing instead of a specific pick, that's usually because your product descriptions don't carry enough distinguishing detail (use case, fit, skill level) for the questions to matter. See Chat optimization.
Language coverage only covers part of the experience
The AI's actual replies adapt to whatever language the shopper is using, regardless of store settings. What's fixed is the widget's static UI — buttons, guided-discovery labels, footer line, disclaimer — which only has translated strings for English, German, and Dutch. A shopper in a fourth-language market gets fluent AI prose sitting next to English UI chrome. If you do operate in one of the three supported languages, it's worth double-checking your shop's content-language metafield is actually set right; a mismatch there is a common source of an experience that feels off for no obvious reason. See Multi-language support.
Closing the loop on what's actually confusing people
Information Gaps Detection runs after every response and sorts what the AI couldn't resolve — including cases where the app itself misbehaved (system_error: a failed tool call, a reply in the wrong language) separately from real content gaps (knowledge_gap) or store issues (bug). Check the Knowledge Gaps tab regularly, and "the chat feels off sometimes" turns into a specific, ownable list instead of a vague impression. See Chat management.