How it works

Every reply runs the same disciplined loop. It keeps recommendations accurate, on-brand, and rooted in what you actually sell — the assistant never free-styles a product into existence.

1. Understand. It reads the shopper’s intent and the page they’re on to grasp what they really need — not just the literal words, but the goal behind them.

2. Search. It queries your live catalogue across categories in parallel. Search is keyword-precise and grounded in your inventory, never guessed.

3. Verify. It pulls real product details — prices, stock, attributes — so it never recommends something unavailable or misstates a price.

4. Answer. It streams a short, on-brand reply with the products that fit, and links only to your own product pages. It points; it never adds to the basket.

5. Remember. It keeps the conversation’s context so follow-ups (“compare those two”, “any vegan options?”) feel like one continuous chat.

Grounding

The defining property of the assistant is grounding: it only ever recommends products that are present in your catalogue at answer time, with the prices and stock it just verified. If something isn’t in your catalogue, it isn’t in the answer. This is what separates it from a general chatbot bolted onto a storefront.

Tip

Because answers are grounded in live data, they stay correct as your catalogue changes — new lines appear, sold-out items drop out — with nothing to re-train.

Memory

Each conversation has a session with memory, so the assistant can resolve references like “the first one” or “something cheaper” without the shopper repeating themselves. Sessions are scoped to a single shopper on a single storefront and never bleed across brands — see Trust & safety.

Where it runs

The assistant is a Vertex AI agent running on Google Cloud, delivered as a fully managed service. Your storefront only loads a lightweight launcher; the model, catalogue search, and session memory all run server-side under your storefront key. There is nothing for your team to host, scale, or patch.