You hired a spokesman you never interviewed. He works around the clock, sounds certain, and quotes 2025 prices.
It Isn’t Lying. It’s Repeating.
1. The damage is measurable. One audit found 72% of brands had at least one flat factual error in AI answers about them, usually stale pricing lifted from a review site that quit updating. The Faro Index scanned 244 companies across four assistants: 88.8% accuracy, one fact in nine wrong, fintech last at 82.4%.
2. Arguing with the model is arguing with Paulie Walnuts. Paulie originates nothing. He hears a thing at the Bing, repeats it with conviction, and by Tuesday it’s gospel. You don’t correct Paulie – you correct the room. NP Digital ran 600 prompts across six platforms: ChatGPT was fully correct about brands 59.7% of the time; Grok, 39.6%.
3. A citation is decoration, not proof. Only 51.5% of AI sentences were fully supported by their cited sources. The BBC tested 100 news stories: 51% of answers had significant problems, 13% of quotes altered or invented.
4. The playbook, five moves. (a) Build a claim ledger: every material fact next to its source of record. (b) Run a fixed prompt set monthly across the major assistants, same wording, logged. (c) Record the cited URLs; the URL is the crime scene. (d) Repair upstream: the dead directory listing, the pricing page nobody redirected, the PDF rotting on your server. (e) Retest in 30 days.
5. The leak starts at home. Livia never raised her voice. She just fed Junior a slightly wrong version of events, and people died over it. Most bad answers trace to something you published and never killed: an old deck, a zombie landing page, a partner site running 2019 boilerplate.
Nobody Owns This, So Name Your Consigliere
6. Right now it’s a no-show job. Tony was on the books at Barone Sanitation: title known, work optional. No job title exists for correcting what an assistant says about you, so it lands on whoever noticed. That is not an org chart.
7. Silvio had the real assignment. Know what’s being said, know if it’s true, get to it early. Name it – AI Answer Accuracy Lead – and give it to one person in marketing ops or comms. Committees notice things; people fix them.
8. The whole job, monthly, four to six hours. Run the prompt set. Update the ledger. Triage by severity times frequency (price beats founding year). Open source-repair tickets with an owner and a due date. Send one page upstairs: what changed, what’s still wrong, what it costs.
9. Skipping it is the Adriana problem. A quiet problem, known to one person, buried because nobody wanted the conversation. It compounded, and then came the meeting in the woods.
The Bottom Line
You can’t argue with the machine. You can change what it reads. Put a human name on that this month – until somebody owns it, it stays nobody’s job until it’s everybody’s emergency.