A generic website chatbot reads pages. A Magento 2 product discovery chatbot has to read the catalog: categories, filterable attributes, configurable parents, store views, and stock. Those are not the same architecture, even when both UIs are a bubble in the corner. Claspwell is the catalog-aware Magento 2 / Hyvä sales assistant for that second job. This article is the contrast piece for merchants who have already been shown a crawl demo.
Pair it with /magento-2-ai-chatbot and /product. If you are still choosing among named vendors, use /blog/how-to-choose-magento-ai-chatbot and the dated roundup at /blog/best-ai-chatbots-magento-2.
What “product discovery” means on Magento
Shoppers rarely type a SKU. They describe a need: a 3-metre strap in a load class, a jacket in a size, a spare part they can only name by symptom. Magento already solved half of that with layered navigation. A product discovery chatbot should reuse that graph — resolve category first, clarify attributes, return in-scope SKUs — not paste the homepage into a prompt and hope.
Claspwell’s loop is category-first, typically no more than two catalog queries per turn. Nested taxonomy stays intact. Unmapped attributes will not clarify well; mapping is part of the guided install, not a self-serve “enable all attributes” toggle. That is why v1 install is assisted and Magento auto-install is out of scope.
Why crawlers fail Magento catalogs
HTML does not encode parent/child relationships cleanly. A crawler will index child PDPs if they are reachable and then recommend three cards that are the same jacket. It will not know which store view owns the price. It often puts an API key in the browser or asks you to BYO OpenAI. Conversation “unlimited” usually means your token bill is unlimited.
Catalog sync plus parent collapse is the Magento-native fix. Claspwell hides simple variants that belong to a configurable parent; child SKU lookup walks to the parent; cards use a From price when variants differ. Mechanics: /blog/magento-ai-chatbot-catalog-sync and /blog/magento-ai-chatbot-configurable-products.
Freshness without pretending every turn is live
Generic bots cache whatever they crawled on Tuesday. Catalog-aware chat answers from the last successful sync for that store_id. Claspwell cadence is daily (Essential), hourly (Growth), or 15 minutes (Scale). Special prices sync with the product. We do not claim a live Magento stock API on every turn in v1. Stale sync is the usual incident: upgrade cadence or fix the Magento job, do not yell at the model. See /blog/magento-chatbot-price-stock-validation.
Customer-group and Shared Catalog prices are not a published v1 claim. Guest catalog prices are not entitlements. A generic bot that scrapes a logged-in PDP on a staging cookie will lie to guests with more confidence than a sync job that only publishes public catalog rows.
Storefront cost: Hyvä is not a generic theme
Generic SaaS widgets ship a React island because that is what their dashboard compiles. Hyvä removed RequireJS so first-party JavaScript stays small. Dropping a heavy chat runtime on every category page taxes LCP and INP. Claspwell’s storefront surface is a deferred script: launcher first, panel on demand, no layout-shifting iframe, theme-token friendly UI. Secrets stay off the browser. Notes: /blog/hyva-alpine-storefront-chat-widget.
Luma installs are still in scope for the same guided onboarding. The constraint is Magento-thin, backend-thick — not a second storefront framework.
Isolation and EU hosting
Generic website chat treats “your site” as one blob. Magento multi-store is several catalogs that must never mix. Claspwell binds every session to a verified store_id after origin checks. Credentials are encrypted per tenant. The backend target is Hetzner in Frankfurt. IPs are hashed. Question text may still go to an inference provider for Tier-3 answers. We do not claim SOC 2. Read /blog/magento-ai-chatbot-gdpr-eu-hosting and /blog/magento-ai-chatbot-multi-store.
Three-tier routing versus stuffing GPT
A generic chatbot shoves catalog text into every model call. That is slow and expensive. Claspwell keeps most turns on scripted lookups and light-intent attribute filters; contextual GPT is the minority path. That is how Essential and Growth stay honest for Magento merchants instead of surprising finance with a token invoice. Conversation allowances and overage live on /pricing: 750 / 3,000 / 10,000 AI-path conversations on Essential / Growth / Scale, throttle at 200% to deterministic catalog answers.
What a product discovery chatbot is not
It is not a WhatsApp order agent. It is not Magento-admin live chat for refunds. It is not a replacement for Clerk.io-style search merchandising or Luigi’s Box conversational search, though all three can help a shopper find a product. Claspwell’s job is storefront sales chat that knows Magento structure. Pick search vendors for search KPIs; pick helpdesks for tickets; pick Claspwell when the Magento catalog is the sales floor.
Implementation is still high-touch: request, meeting, proposal, guided install (1–2 weeks after access is ready), go live. Magento install is included. No Stripe. Request that path at /request-integration when a crawler demo is no longer a sufficient answer.