Maintenance & ongoing optimization
ShopGuide doesn't degrade on its own, but the store around it changes constantly — new products, updated policies, seasonal discounts, new languages. Maintenance here means keeping those changes reflected in what the AI can see and act on. This is a short, practical checklist rather than a generic ops framework.
Keep catalog content synced
Product search, product questions, and guided discovery all answer from a vector index of your catalog plus RAG over your shop's content. A product that's discontinued, repriced, or rewritten in Shopify but not reflected in the indexed content will produce stale or wrong answers — not because the AI degraded, but because it's answering from what it can see. When you make significant catalog or policy changes, confirm they're reflected in chat answers rather than assuming sync is instant and automatic.
Review Information Gaps on a set cadence
Information Gaps Detection is silent by design — it never interrupts the chat, which also means it's easy to forget to check. Set an actual recurring time (weekly is reasonable for most stores) to review the Knowledge Gaps tab in your transcript viewer, work through open knowledge_gap and bug entries, and watch system_error specifically as an app-health signal rather than a content one. Recurring entries are already stacked by topic, so a high occurrence count is a real prioritization signal, not noise to filter out. See Chat management.
Keep the Discount Generator's eligible list current
If you use the AI Discount Code Generator, its eligible products/collections list is a manually maintained setting, not something that updates itself as your catalog changes. A new product line launched last month isn't eligible for AI-generated discounts until it's added to that list in Settings → AI Behavior. Review it alongside any catalog refresh, and confirm the discount value, duration, and once-per-customer settings still match your current promotional intent. See Settings panel.
Don't let A/B tests run indefinitely — or end too soon
A test that's been running for months without a decision is wasted signal; a test called after two days is likely noise. Set an expectation for how long a launch-message or prompt test runs (a full business cycle, at minimum) before reading A/B test results, and archive or conclude tests once you've acted on them rather than letting old variants keep splitting traffic indefinitely.
Re-check language configuration after theme or market changes
The chat widget's UI chrome only has translated strings for English, German, and Dutch, keyed off the shop_content_language metafield. If you change themes, add a market, or a metafield gets reset during a theme migration, confirm this value is still set correctly — a silent mismatch here produces English-language buttons sitting next to AI replies in the shopper's actual language, which reads as a bug even though the AI itself is working correctly. See Multi-language support.
Revisit custom prompt configurations periodically
If your store runs a fully custom system prompt or restricted tool set instead of the standard flow, remember that it's maintained independently of the standard configuration — new intents, tools, or features added to the standard flow don't automatically apply to a custom setup. Periodically confirm a custom configuration still reflects your current goals rather than a snapshot from when it was first built. See Custom system prompts.
Next steps
Maintenance here is mostly about keeping configuration and catalog content current — the AI's quality is a direct reflection of what it's allowed to see and act on.