Skip to main content

Chat optimization playbook

ShopGuide's chat quality depends less on wordsmithing welcome messages and more on three things you directly control: how completely your catalog is indexed for semantic search, whether your product data matches how customers actually phrase requests, and how quickly you act on what the AI couldn't resolve. This guide covers each, using ShopGuide's actual architecture rather than generic chatbot advice.

How ShopGuide decides what to say​

Every incoming message is classified into one of the intent types — PRODUCT_SEARCH, PRODUCT_QUESTION, PRODUCT_COMPARISON, GUIDED_DISCOVERY, BROWSE_COLLECTIONS, INFORMATION, ORDER_MANAGEMENT, SUPPORT, ADD_TO_CART, CHECKOUT, SPECIAL, CLARIFICATION (plus DISCOUNT_CODE if you've turned on the AI Discount Code Generator) — before the AI responds. Each intent activates a different tool set and prompt section, not just a different tone.

The practical implication: if customers describe products in phrasing your catalog doesn't contain, they land in BROWSE_COLLECTIONS or CLARIFICATION instead of PRODUCT_SEARCH, and get a generic category response instead of a direct answer. The fix isn't a better welcome message — it's tightening product titles, collection handles, and descriptions to match how shoppers actually talk about what you sell.

Catalog coverage is your real lever​

Product search and product questions are answered using Qdrant vector search over your catalog, plus RAG over your shop's knowledge content — semantic matching, not keyword matching. A shopper typing "warm winter hat" should reach a product titled "Merino Wool Beanie" because the match is on meaning, not shared words.

That also means the AI can only recommend or answer questions about what's actually indexed. If a product line isn't showing up in chat results, check catalog sync and content completeness (full descriptions, current variants, populated metafields) before touching prompt wording — an unindexed or thinly-described product can't be found no matter how the AI is instructed.

Guided Discovery: when it earns its keep​

GUIDED_DISCOVERY covers vague-need requests — "help me choose," "what do I need for X," "based on my riding style" — where a keyword search would return too broad a list to be useful. Instead of searching immediately, the AI asks up to three short questions, each with exactly three answer options, presented as an in-chat overlay, then searches using the shopper's answers.

This flow is only as good as the distinctions your catalog supports. If two products differ in ways your descriptions don't capture (skill level, use case, fit, material), the follow-up questions can't meaningfully change the result set. If guided discovery conversations tend to end in generic browsing rather than a specific recommendation, that's usually a catalog-detail problem, not a prompting problem.

Closing the loop with Information Gaps​

Every chat response is checked afterward by Information Gaps Detection, which classifies anything the AI couldn't fully resolve into one of five categories: knowledge_gap (your content doesn't cover it), bug (something is broken in your store), live_data (needs real-time information the AI has no tool for), system_error (the app itself misbehaved — a failed tool call, a wrong-language reply), or other. Recurring phrasings of the same underlying question are stacked into a single row instead of fragmenting into duplicates, so the count you see is a genuine frequency signal, not noise.

Check the Knowledge Gaps tab in your transcript viewer regularly — see Chat management. It's the most direct answer to "what is my AI actually failing at," and a shop-specific spike in system_error is worth treating as a regression to investigate, not a content gap to write around.

Test launch messages and prompts, don't guess​

Rather than assuming which welcome message or system-prompt variant performs better, run it as an A/B test. ShopGuide's testing framework splits shoppers into variants and attributes both chat engagement and completed Shopify orders back to each variant, so you're comparing real conversion outcomes rather than proxy engagement numbers. Set one up under A/B testing and read results via A/B test results — see Common mistakes for how long to let a test run before acting on it.

Language coverage​

The widget's static UI chrome — buttons, guided-discovery labels, footer text, disclaimer — ships in English, German, and Dutch only; those are the three shop_content_language values with translated strings. The AI's actual conversational replies aren't limited to these three languages: it detects and responds in the shopper's language regardless of the UI chrome setting. If you serve a market outside en/de/nl, expect fluent AI prose alongside English-language buttons and labels. See Multi-language support for what's configurable and what isn't.

Next steps​

Chat quality is a function of catalog data and configuration, not copywriting. Fix coverage and routing first — tone changes matter far less than whether the AI can find and understand what you sell.