Journal / AI
AI · 7 min read

How we keep AI-written FAQs from making things up

Grounding, confidence thresholds and the approval queue: the boring engineering behind trustworthy answers.

Dev Sharma
Engineering · 15 Jul 2026
article hero image 2000×1000 · drop chart or store screenshot here

The fastest way to lose a merchant's trust in an AI feature is one confidently wrong answer on a live product page. Most of Genie's engineering effort goes into preventing exactly that, not into writing better prose.

Every answer needs a source

Genie will not draft an answer unless it can point at a specific product field, a review, or a resolved support ticket. If no source exists for a common question, it surfaces the gap to the merchant instead of guessing.

The approval queue is the safety net

Nothing publishes without a merchant reviewing the first batch. Confidence thresholds mean well-sourced, low-ambiguity answers can be set to auto-publish later, but that setting is opt-in, never a default.

Trustworthy AI copy on a storefront is less about the model and more about refusing to answer when the evidence is not there.

See the sourcing view yourself

Every Genie draft shows exactly which review or ticket it came from.

See the app