Why ‘Zero Hallucinations’ Is the Wrong Promise for Commerce AI

A shopping assistant that invents a material, a compatibility claim or a return condition can erode trust quickly. The answer is not to promise “zero hallucinations”. No language model, prompt or retrieval layer makes that a credible absolute guarantee.

A more useful goal is to make every consequential product answer accountable to knowledge the merchant has approved. Good answer governance draws a clear line between supported claims and uncertainty. It is not a certificate of infallibility.

Start with the right product knowledge

The assistant needs more than a title and a list of specifications. It needs the intended use, constraints, comparisons and a clear source for each important claim. Product Engine builds this knowledge from the existing catalog and relevant evidence under merchant control.

An external source may suggest a useful addition. It does not become a merchant-approved fact merely because a model found it. An unanswered question is a reason to investigate, not an invitation to fill the gap with plausible prose.

Separate a recommendation from an unsupported claim

“Both jackets are waterproof, but this one is lighter” requires evidence for each factual part. “I would choose this one for your trip” also depends on the shopper's stated needs and the product's limitations. A useful answer should make those conditions visible rather than disguise them as certainty.

When a source is missing or contradictory, the assistant should say what it knows, state the limit and direct the shopper to an appropriate next step. That may be a product page, a clarification question or a human adviser, depending on the merchant's service model.

Controls worth testing

• Is the answer grounded in the correct product and variant?
• Can a reviewer find the source behind a material, safety or compatibility claim?
• Does the assistant distinguish product fact from style advice or inference?
• Does it acknowledge a missing answer instead of inventing one?
• Do price, stock and service terms come from their current authoritative systems?

These controls reduce risk, but they must be tested against real questions and monitored as the catalog changes. They do not turn every generated sentence into a verified fact.

What this means for a merchant

Brand Agent uses approved product knowledge to guide shoppers on the merchant's site. The commercial benefit is more useful product advice with clearer boundaries, not a claim that errors have been eliminated or that regulatory compliance has been conferred by an AI feature.

Read the deeper governance framework or request a demo to explore how your catalog supports real buying questions.