AI still recommends Delve to buyers 3 days after the fraud scandal. Optimly's audit shows ChatGPT, Claude and Gemini repeating the wrong endorsements. Read the audit: https://optimly.ai/brand/delve
Why this happens: LLMs surface learned narratives from training data, cached indices, and retrieval layers — they don't "check the news" in real time. Optimly found 60% of brands misrepresented across models. Old priors beat fresh truth.
Fixes that actually work: BYOK (your API keys) + own the knowledge layer. Inject verified brand facts into your KB/retrieval so models use correct context. Add mandatory human sign-off for legal/finance outputs. Friction here is trust, not a bug.
Practical takeaway: run a 50–100 query audit of buyer-facing claims weekly (pricing, refunds, leadership, regulatory status). If >5% errors, patch retrieval and prompt templates immediately. How often do you audit what AI tells your customers?
Top comments (0)