How to stop an ecommerce chatbot from making things up

When an ecommerce chatbot invents availability, a product detail, or a returns rule, the problem is more than awkward wording. A shopper may make a decision based on information the store never approved. Reducing that risk requires reliable sources, explicit boundaries, and continuous checks.

Start with an authoritative source

The chatbot should answer from recognizable, maintainable data: the catalog, product pages, store policies, help pages, and approved internal documentation. Each question type needs an accountable source. The catalog can describe materials and variants; the shipping page should govern delivery areas and conditions.

Valen ingests the Shopify catalog as a product information source. It can also enrich catalog attributes so the available information is easier to use. This does not replace good source data. Ambiguous titles, empty fields, and outdated policies still need to be corrected at their origin.

Separate facts, inferences, and advice

A sound answer distinguishes what the catalog states from what the assistant infers. “The product page lists cotton” is a fact. “It may suit a warm day” is advice and should sound like advice. A system should not turn probability into certainty or silently complete missing details.

Define sensitive categories such as availability, price, compatibility, delivery, returns, ingredients, and warranties. If the source does not clearly answer a question in one of these areas, the safer behavior is to state the limitation and point to the appropriate page or channel.

Design a safe response to uncertainty

“I cannot find that information” is often more useful than a plausible falsehood. Prepare fallback responses that explain what is missing, show where to verify it, and invite the shopper to rephrase. A fallback does not need to end every conversation. It may still suggest products when enough criteria are available.

Valen lets shoppers search the catalog and ask product questions through conversation. The result should therefore be evaluated against real store questions, not only polished demonstration prompts.

Build a control set

Collect frequent requests, confusing wording, and questions the system should decline to answer. For each one, record the expected source, an acceptable response, and the conditions that require a refusal. Include similar products, unavailable variants, misspelled names, and requests that combine several constraints.

Repeat the checks whenever the catalog, policies, or configuration changes. Record incorrect responses and identify whether the cause is source data, retrieval, or wording. Fixing the cause is more robust than adding a special sentence for every mistake.

Where Valen fits

The limitation is straightforward: a chat can phrase an answer even when its source is incomplete. The required capabilities are identifiable product data and a path back to product details the shopper can inspect. Valen imports the Shopify catalog, as demonstrated on a demo store. It completes and enriches product listing attributes, as demonstrated on a demo store. It also lets shoppers search the catalog and ask product questions within the conversation, as demonstrated on a demo store. This makes Valen more than an isolated chat, but it does not repair missing data or disconnected policies: the merchant must maintain sources and define when the assistant should stop answering.

A checklist before publishing

Confirm that every sensitive statement can be traced to a source, that uncertainty is visible, and that absent details are not invented. Check that old content is removed, updates reach the assistant, and the team knows how to review problematic conversations.

Book a demo to evaluate how a catalog-based conversation handles real questions from your store.