Spec-hallucination audit

Invented product specifics — wrong battery life, storage, tier contents, compatibility — are the highest-severity AI visibility incident type for a technology catalog, and the tools are weak at catching them: in nine-platform testing, only two consistently flagged errors about a brand’s own pricing or features. This audit is the capability no vendor ships, because only you hold the ground truth.

Why this, mechanically. A model that confidently invents a spec does real commercial damage and shows up in no error log. Diffing assistant answers against your PIM truth, monthly and per market, converts that silent exposure into a routed incident list — and the routing matters: wrong facts usually trace back to an entity or content gap.

Deliverable. A monthly job asking each assistant the top factual questions per flagship product, per market; a per-surface accuracy score; and errors routed to entity or content owners, with pricing and invented-capability errors at the top.

Finish line. Every high-severity wrong answer has a routed owner, and the accuracy score trends down over quarters.

This encodes the truth set only you hold. It pairs with the prompt corpus baseline — the corpus measures presence, this measures accuracy.