This is the setup guide we wish we had the first time we put an AI sales agent on a Shopify store. It assumes you already have a Shopify store with real traffic, real products, and a desire to lift conversion. If you're pre-launch, come back later — the math doesn't work yet.
Before you start
You'll need: a Shopify admin login (or a collaborator invite with read_products, read_orders, read_customers, read_content scopes), an ESP API key if you use one (Klaviyo, Mailchimp, HubSpot), and 30 minutes uninterrupted. That's it. No developer required.
We also recommend pointing the install at a dev store first if you have one. The free trial means you can validate the experience without risking the production catalog.
1. Install the embedded app
Go to the Shopify App Store and click Install Zubby. Shopify will route you through the OAuth flow and ask you to approve the scopes listed above. Approve, and you'll land back inside the Zubby embedded admin — already logged in, no second password.
Behind the scenes, the install hooks every webhook the agent needs (products/create, products/update, orders/create, customers/redact, etc.) and starts the first catalog crawl. You don't have to manage any of this manually.
2. Sync your catalog
The first catalog crawl uses Shopify's bulk-operation GraphQL API, so even a 25,000-SKU catalog finishes in under 20 minutes. We index:
- Every product, variant, and collection.
- Metafields you've marked as agent-readable (allowlisted, never the full namespace).
- Pages tagged as policy or FAQ — shipping, returns, sizing.
- The last 30 days of customer orders for context.
While the crawl runs, the Zubby dashboard shows a live progress meter and the first products show up immediately. You can start tuning voice rules before the catalog finishes.
3. Tune voice + guardrails
Pick a tone preset (Professional, Warm, Playful, Minimal) or write a custom voice block. Then set the guardrails: any banned phrases, any forbidden claim categories (medical, legal, competitor), any handoff triggers (low confidence, pricing override request, etc.).
Spend extra time on guardrails. The most common reason an AI agent embarrasses a brand is a missing claim block, not a bad prompt.
4. Drop the widget
Zubby offers two ways to render the agent on your storefront. The default is the embedded widget — a single <script> tag added via the theme app extension, no theme.liquid edits required. The other path is the checkout extension, which surfaces the agent inline on the cart and checkout pages.
Both paths respect Core Web Vitals: the widget bundle is ~28KB gzipped, loads after Largest Contentful Paint, and never blocks render.
5. Test in a sandbox
Open your store in an incognito window and chat with the agent like a real customer. Try the hardest questions — sizing across variants, "does this work with X", an out-of-stock item, a return question. The agent should answer accurately, escalate gracefully when it can't, and never improvise.
Every conversation is logged in the Zubby dashboard with the retrieved sources, the tool calls, and the final response. That's your audit trail — review the first 50 conversations carefully before letting traffic flow at full volume.
6. Measure what shipped
Within 7 days you'll have a per-store ROI dashboard showing: AI-influenced conversion rate, attributable revenue, cart-recovery dollars, and a breakdown by acquisition channel. Pricing stays simple while you evaluate: one flat monthly fee per tier, billed by conversation volume, with no per-resolution AI charges.
Most stores see meaningful lift in week 2 once the catalog is fully embedded and the agent has handled enough conversations to tune itself.