Why AI Assistants Get Your Products Wrong (and How to Catch It Before Shoppers Do)

• Stale snapshots — the answer reflects your store as of the last crawl, not today. Price changes, sales, and stockouts arrive late; the model states them as present-tense fact.
• Variant blindness — data flattened to the parent product. The assistant recommends the sold-out size, quotes the base price for the premium variant, or misses that the color exists at all.
• Facts trapped in pixels — size charts, ingredient lists, and dimension diagrams that live in images. Invisible to most retrieval, so the model guesses from category averages.
• Policy drift — your policy page, FAQ, and structured data disagree after an update touched only one of them. The assistant picks a truth; often the old one.
• Borrowed identity — thin brand data forces the model to lean on third-party content: an old review, a Reddit thread, a competitor comparison. Your story gets told by whoever wrote about you last.

These failures are silent by construction. The conversation happens on OpenAI's or Perplexity's servers; no analytics event fires; the shopper who was told the wrong thing doesn't file a ticket — they just don't arrive. The first detectable symptom is usually a support ticket that starts "but ChatGPT said…", which is the channel telling you it's been wrong for months.

The fix starts with detection. Build a panel of your fifteen most consequential questions — top products' fit/compatibility questions, shipping and returns, brand basics, your top three head-to-head comparisons. Run it monthly in fresh sessions across ChatGPT, Perplexity, and Gemini. Grade three things: present or absent, right or wrong, and whose sources got cited. Wrong-and-confident items are your priority queue.

Then fix upstream: push trapped facts into text and structured fields, make one policy surface canonical, complete variant data in feeds. The answers follow the data within crawl cycles.

This loop — probe, grade, fix, verify — is Kinect's AI Readiness Audit productized. We run it with your real shopper questions, verify every wrong answer against your live site, and hand you the ranked fix list. The same corrected catalog truth then grounds the AI sales rep on your storefront, so shoppers on-site get right answers even while the external crawlers catch up.