The model said something wrong about your company. Now what?
The wrong instinct is often to file a complaint with the engine. The right instinct is typically to edit the citation graph the engine pulls from.
Here's the playbook.
Step 1 — Verify the error is reproducible
Run the prompt across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. Run it five times each. Note which engines produce the error and how consistently.
An error that appears in one engine on one run is often noise. An error that appears across multiple engines or consistently across runs is typically a citation graph problem. Treat them differently.
Step 2 — Find the source
When the engine cites, capture the source. When it doesn't cite, ask follow-up prompts: "Where did you get that?" "What sources support that claim?" The model will often surface the underlying reference — or admit the gap.
The source list becomes the repair map. You can't easily fix what you can't trace.
Step 3 — Score the source's authority
— Tier-1 outlet getting the fact wrong. Hardest fix. Typically requires a correction request to the publisher and follow-on coverage that supplies the correct fact at equal or higher authority. — Mid-tier outlet or Wikipedia. Faster fix. Wikipedia edits with strong sourcing often land in weeks. Mid-tier corrections through PR outreach can land in days. — Low-authority blog or scraped content. Hardest to remove, easiest to displace. Stop trying to delete it. Outrank it with authoritative content the engine will weight higher.
Step 4 — Issue the correct fact at higher authority
The model is a weighted retrieval system. The correct fact doesn't have to delete the wrong fact. It has to outweigh it.
That tends to mean tier-1 earned media stating the correct fact. A Wikipedia article with the correct fact and strong citations. An owned authoritative source — fact sheet, press kit, leadership bio — that supplies the correct fact in a format the engine retrieves cleanly. [Read: The Authority Stack]
Step 5 — Re-test
Most engines update retrieval on shorter cycles than they used to. New authoritative sources often enter the citation graph within days. Re-run the audit. Re-score. The error will typically move first in Perplexity (heaviest live retrieval), then ChatGPT with browsing, then the engines more reliant on frozen corpora. [Read: Updating What AI Knows About You: A Realistic Timeline]
What does not work
Contacting OpenAI, Anthropic, Google, or Perplexity for individual corrections at scale — the engines don't generally operate manual review queues for brand factual disputes outside of clear policy violations. Suing the model — the legal frontier is unsettled; pursue legal remedy when the harm justifies it, not as the primary repair channel. [Read: When AI Defames You] Demanding the source publisher delete the article — even when the source agrees, the model has already trained on it. The work is typically to displace it, not erase it.
What tends to work
Authority outranks authority. The fix typically runs upstream. Build the correct fact into sources the engine already weights — and the engine will often surface the correct fact on the next retrieval.
See also: Signals That Move AI Reputation · AI Reputation Glossary
No communications firm can guarantee specific outputs inside third-party AI systems. The discipline is shaping the inputs the engines retrieve from — not directing the engines themselves.
