Sooner or later it will happen to you. A prospect screenshots a ChatGPT answer that says your product has a feature you’ve never built, costs three times what you actually charge, or — worst case — combines your brand name with a competitor’s pricing. They forward it to their procurement team. You find out about it on Tuesday morning.
This is the playbook we run with our clients in the first 24 hours.
Hour 0–2: contain, don’t spiral
- Don’t tweet about it. Public correction posts feed the model the wrong narrative as much as the original hallucination did.
- Reproduce it cleanly. Get the exact prompt, exact model, exact date. We log every detail because patterns matter — single hallucinations are noise, repeated ones are a signal.
- Check whether it’s just one model or all of them. If only ChatGPT has this wrong, the fix is different than if every model agrees.
Hour 2–24: trace and patch
- Find the source the model is leaning on. Almost every hallucination has a real-world origin — usually an outdated press release, a confused Reddit thread, or a competitor’s comparison page. Find it. That’s where the patch lives.
- Update your own canonical sources first. Your website, Wikipedia entry, schema markup. Make sure the correct answer is well-supported in places models trust.
- Don’t request a “correction” from the model. You can’t. Models don’t have a feedback loop you can ping. Time and corpus refresh are the only fixes.
Tags Playbook Hallucinations Crisis Response