Best practices

The Conversations server is a prototype environment for building AI shopping experiences on top of the Pacific Apparel sample catalog. The tips below help you design effective prototypes, evaluate results, and prepare for what changes when you move to a production integration.

Understand the catalog

Your prototype is only as convincing as the queries it can serve. Before designing demos or test cases, get familiar with what's in the Pacific Apparel catalog:

• Coverage: Strong in women's and men's apparel, especially jeans, trousers, tops, and shirts. Smaller selections in footwear, accessories, and jewelry.
• Brand mix: House brands Roadster and Pacific have the deepest coverage. Adidas, Levis, MANGO, Tokyo Talkies, and Van Heusen also have multiple products.
• Price range: Most items fall between $20 and $60. Premium products go up to $300+, mostly in jewelry.
• Promotions: Six seasonal tags are available: blackfriday, back2school, mothersday, christmas, valentine, easter. Use these to test promotion-based discovery flows.
• Gaps: The catalog has no kids' clothing, no home goods, no electronics, and no groceries. Prototypes outside fashion won't have realistic data to work with.

Asking your AI assistant to give you an overview of the catalog is a useful first step. See Test the Conversations server for verification prompts.

Design realistic prototype queries

The best prototypes feel like real shopping conversations. To build queries that match real shopper behavior:

• Mix specific and exploratory prompts. Real shoppers move between modes — sometimes they know exactly what they want, sometimes they're browsing. Test both.
• Use natural phrasing. Avoid catalog-speak like "filter on data.color = blue" Use what a shopper would actually say — "show me blue jeans."
• Include refinements. Real conversations rarely end after one search. Test follow-ups like "those are nice but cheaper" or "do you have any in black?"
• Test edge cases. Out-of-stock queries, oddly specific requests, and queries with no obvious match all reveal how your prototype handles the messy reality of shopping.

Guide shoppers when integrating in a chat widget

If you're embedding the prototype in a chat widget for demos, a short prompt sets expectations:

Tell me what you're looking for and I'll find it in our catalog. The more detail you share — color, size, budget, or what it's for — the better the results.

This signals that plain language works and that more detail produces better results.

Evaluate prototype quality

A good prototype produces results that look and feel like a real shopping conversation. As you test, watch for:

• Accurate filtering. Does the AI assistant pick up on price, color, size, and brand mentions correctly?
• Helpful follow-ups. When the AI assistant asks clarifying questions, are they useful or generic?
• Smooth transitions. When the shopper moves from "I'm browsing" to "show me X," does the AI assistant transition cleanly?
• No dead ends. When filters are too strict, does the AI assistant relax constraints and explain the change?
• Realistic suggestions. Refinement suggestions should reflect real catalog distribution, not made-up ranges.

If your AI assistant consistently misses on any of these, the issue is usually in how the AI assistant interprets shopper messages, not in the catalog data.

Share feedback

The Conversations server is in early access. If you hit cases where the AI assistant misinterprets shoppers, returns the wrong tool, or produces unhelpful results, share examples with your Bloomreach customer success manager. Feedback during this phase shapes how the production version behaves.

Related pages

• Test the Conversations server: prompts you can use to validate prototype changes.
• Tools: the tools your AI assistant calls during a conversation.