How-to · Train Shopify AI on products
How to train your Shopify AI chatbot on your products — 10-step playbook.
The AI is only as good as the data you give it. Most stores under-invest in product-data quality + knowledge base content, then complain the AI "doesn’t answer well." 2-4 hours of focused work + the right structure transforms quality dramatically.
Free tier covers all of this. The work is in your Shopify product data + Clearly knowledge base, not in extra paid features.
The 10-step playbook
- 01
Audit your existing product data
Open Shopify Admin → Products. Pick your top 20 SKUs by revenue. Check each: does it have descriptive title? Detailed description? Filled-in tags (material, size, color, use-case)? Metafields populated? Most stores have 50% of products at "needs work" quality. The AI is only as good as this data.
- 02
Structure product descriptions for chat
Customers ask "is this waterproof? what fabric? machine washable?". Write descriptions that answer those questions explicitly. Avoid marketing flowery language; prefer concrete specs. Format: 2-3 sentences of brand voice + bullet list of factual specs.
- 03
Use Shopify metafields for structured data
Metafields > tags for structured info. Example metafields: care_instructions, dimensions, material_composition, country_of_origin, lab_certifications, brand_voice_notes. Agent reads metafields natively. Spend 30 min on metafield structure for top SKUs; agent quality lifts dramatically.
- 04
Populate the Clearly knowledge base
Settings → Knowledge. Paste in: return policy (verbatim, not summarized), shipping table (zones + costs + lead times), size chart (including fit notes "runs small"), ingredient lists, brand story (1-2 paragraphs), FAQ — the 20 questions you answer most in email. Format: narrative paragraphs. Don't convert to strict Q&A pairs.
- 05
Add brand voice instructions
Settings → Voice. Be specific: "use lowercase only", "never use exclamation marks", "always offer free shipping when asked", "in Spanish use tú not usted", "keep replies under 3 sentences". 5-10 explicit rules > 1 generic adjective ("be friendly"). The agent applies these to every response.
- 06
Configure "always mention" and "never mention"
Always-mention examples: current promo code, free shipping threshold, brand origin story. Never-mention examples: competitor names, refund amounts (route to human), unapproved health claims (supplements/food). Critical for compliance categories.
- 07
Set up product collections for recommendations
For "build me a routine" or "outfit for occasion X" requests, agent surfaces from your Shopify collections. Configure collections by use case ("Beach vacation kit", "Skincare for sensitive skin", "Beginner runner essentials") and the agent applies them when the matching prompt comes up.
- 08
Test with real customer questions
Settings → Preview shows the widget on a sandbox. Ask the 10 questions your customers ask most. Each answer should pass three tests: (1) factually correct based on your data; (2) sounds like your brand voice; (3) ends with a clear next step (add to cart, see related, ask follow-up).
- 09
Iterate the knowledge base from real conversations
Settings → Inbox → filter "AI low confidence" or "escalated". Each entry is a hole in your knowledge base. Add the answer; agent improves immediately. First month: do this 2-3x per week. After: weekly. The knowledge base improves on its own loop.
- 10
Tune escalation rules based on KB gaps
If specific question types keep escalating despite good KB content, the trigger rules need tightening. Maybe "refund" was over-escalating routine policy questions; relax that. Or specific keyword should escalate but isn't. Quarterly review of escalation patterns.
FAQ
- Do I need to write a separate FAQ for every product?
- No. The agent reads your full product catalog (descriptions, metafields, tags) automatically. Per-product FAQs are only needed for highly specific questions (one product has a unique care instruction the others don't). For most stores, well-structured product data + a general knowledge base covers 80% of questions.
- How much time does training take?
- First-time setup: 2-4 hours of focused work for a 100-SKU catalog. Most of the time goes into populating knowledge base + restructuring product descriptions to be chat-friendly. Subsequent maintenance: 30 minutes a week reviewing escalations and adding answers.
- What if I add new products — does the agent learn automatically?
- Yes. Catalog sync runs in real-time via Shopify webhooks. New product added in Shopify → available to agent within seconds. No re-training needed; the agent reads current catalog state, not a snapshot.
- Can I exclude products from the agent?
- Yes — tag products as "exclude_from_agent" and configure the agent to skip them. Useful for: archived SKUs not yet hidden, B2B-only products in a DTC-mostly catalog, products under review for compliance.
- How do I handle products with custom options (engraving, monogramming)?
- Configure the customization options as metafields on the product. Agent walks customer through the choices. For complex custom orders (truly bespoke), agent gathers requirements + escalates to live takeover with a structured summary. You're not re-asking; you're continuing.
- What's the biggest training mistake?
- Sparse knowledge base + good product data. Agent knows your catalog cold but doesn't know your policies, brand story, or shipping rules. Result: agent fakes answers or escalates trivial policy questions. Fix: spend 30-60 min on the knowledge base; quality jumps dramatically.