It is lazy-loaded and the API only fires when the buyer interacts with it. The widget itself adds almost nothing to the initial page weight, and Lighthouse scores stay clean. The model request goes straight to your provider rather than back through WordPress, so checkout latency is unaffected even at peak load.

Yes. Cart items, totals, the customer's billing and shipping address, and any custom session data can all flow into the prompt. That's what lets the bot answer "will this arrive before Friday" with the actual cutoff for the buyer's zip rather than a generic shipping table.

Yes. Display conditions cover the block-based checkout and the classic shortcode, plus the cart page and any custom checkout template. The widget renders the same way regardless of which checkout flavour your theme uses, so a migration to the new blocks doesn't break anything.

Yes. Display conditions include device targeting, so you can run the bot on desktop only or on mobile only or on both. Some stores find that mobile checkout is where the bot helps most because the policy pages are harder to navigate on a phone, but you can A/B both ways from the same install.

Yes. Drop a webhook to your live chat tool when the bot detects a hand-off intent - phrases like "can I talk to someone" or repeated unresolved attempts. The full transcript travels with the handoff so the human picks up where the bot left off, without making the buyer repeat themselves.

Run two bots with different prompts and split with a query parameter or a feature flag. Conversion rate is the obvious metric; the more interesting one is "chats that resulted in a completed order", which the bot logs and you can correlate with order data. Iterate the prompt, not the placement.

Ground in the policy page and inject the customer's country into the prompt. The system prompt should tell the bot to refuse to compute taxes itself and to defer to the cart breakdown WooCommerce displays. That keeps the bot honest about a question that's easy to get subtly wrong.

It can, but most stores don't and shouldn't. Training buyers to expect a discount in the cart erodes margin permanently. A better pattern is to use the bot to answer the actual question that triggered the hesitation, and reserve discount logic for an exit-intent or post-abandon email where the precedent is contained.