Noma by Nurdd (NomaHQ)Noma
Sign up

Free runbook · PDF

ChatGPT says something wrong about your company? Here is how to get it corrected

A runbook for when ChatGPT, Gemini, Google AI Overviews, Perplexity, Claude or Copilot get your company wrong: triage by severity, trace the source, fix it, use each engine's verified reporting route, and recheck.

  • Severity triage grid with owners and response times
  • Source-tracing worksheet for any wrong answer
  • Publisher, listing and Wikipedia correction templates
  • Per-engine reporting card, checked against vendor help pages
  • Evidence log and recheck schedule

Get your free copy

Instant download, plus a copy by email. No spam: one email with your link, and a reply from a person if you write back.

The short answer

You correct a wrong AI answer by fixing what it is built from, not only by reporting it. Reproduce the answer and record the prompt, engine and date. Find the source it relies on. Correct your own pages, ask third-party publishers for corrections, then use each engine's feedback or legal route. Recheck until it changes.

For PR and comms, product marketing, legal and founders

What you will learn

  1. Chapter 1

    Triage: how wrong, how bad, whose job

    • Classify every wrong answer as legal or safety, pricing or availability, factual product claim, positioning or tone, or outdated before deciding what to do.
    • Severity is the error type adjusted for reach: how often real buyers would see it.
    • Every incident gets one named owner and a response clock you have written down in advance.
    • Capture evidence in the first hour, because AI answers change without notice.
    • Bring legal in for statements about people, allegations of wrongdoing, regulated claims or formal requests, and let them choose the route.

  2. Chapter 2

    Diagnosis: where the wrong answer comes from

    • Wrong answers come from training data, retrieval of a wrong page, a third-party source, or the model filling a gap, and each has a different fix.
    • Reproduce the answer yourself and record the exact prompt, engine, mode, account state, date and cited sources before acting.
    • Run the prompt several times and record how often the error appears, because frequency tells you how serious it is.
    • Search your own domain for the wrong claim first; stale pages you still host are a common source.
    • If no source contains the claim, publish the true fact clearly rather than relying on a feedback report.

  3. Chapter 3

    Fix the source you control: your own site

    • Write each corrected fact as a standalone sentence that names the brand, the product, the fact and its conditions.
    • Date the pages that carry changeable facts and publish explicit statements when something changes, such as a rebrand or a retired plan.
    • Redirect, annotate or remove your own pages that still state the wrong fact, including PDFs and old press releases.
    • Use Organization and product structured data that matches the visible page exactly, with sameAs links to profiles that carry correct facts.
    • A plain facts page with a corrections address gives engines, journalists and publishers one source to check.

  4. Chapter 4

    Fix the sources you do not control

    • Publisher corrections succeed when the request quotes the wrong sentence, links to proof and proposes replacement wording, and asks for nothing else.
    • Under Wikipedia's conflict-of-interest guideline, company staff disclose their affiliation and request changes on the talk page with {{Edit COI}} rather than editing the article.
    • Wikidata has no standalone COI policy page today, but its paid-editing disclosure policy applies the Wikimedia rules: disclose, and propose changes with references.
    • Claim and maintain business listings and review profiles, and keep their facts identical to your own facts page.
    • Fake reviews, undisclosed editing, threats and spam make the problem worse and can break the law.

  5. Chapter 5

    Engine by engine: feedback and formal routes

    • Feedback buttons tell a vendor an answer was poor; no vendor page we read promises they will change a specific answer.
    • OpenAI's content reporting form can lead to review of reported domains by its Model Quality team, which is the route to name a source of harmful falsehoods.
    • Gemini offers a 'Report legal issue' option under each response, and Google Search has feedback controls on AI Overviews and AI Mode.
    • Privacy correction routes at OpenAI and Anthropic are for personal information about individuals, not company facts, and both vendors say they may not be able or obliged to act.
    • For engines that ground answers in web search, fixing the cited page is usually faster than any report.

  6. Chapter 6

    Recheck, close and prevent

    • One evidence log, one row per incident and a dated note for every action, makes legal, leadership and pattern-spotting possible.
    • Recheck on a fixed schedule and report runs with the error before and after, since no vendor commits to a timeline.
    • Write a closing rule in advance and separate “resolved” from “mitigated, monitoring”.
    • A monthly and quarterly accuracy audit against your facts page finds wrong answers before buyers do.
    • Put AI sources on the launch and change checklist so retired facts do not linger.

6 worksheets inside

The chapters explain the method. The worksheets are what you fill in to run it: print them, or copy the tables into a document of your own.

  • W1Severity triage grid
  • W2Source-tracing worksheet
  • W3Correction request templates
  • W4Per-engine reporting card
  • W5Evidence log and recheck schedule
  • W6Quarterly accuracy audit
Get the runbook

Questions

ChatGPT says wrong things about my company. What should I do first?
Capture the evidence: a screenshot, the answer text, the exact prompt, the date and any cited sources. Then rerun the prompt in a fresh session to see whether the error repeats, and classify it by type and severity. Only then decide on a fix, because the fix depends on where the error comes from.
Does the thumbs-down button fix a wrong AI answer?
Not on its own. Vendors describe feedback as a way to tell them a response was poor, and some say it is reviewed or used to improve the product, but none of the help pages we checked promises that a report will change a specific answer. Use it once per confirmed error, with the correct fact and a link, and put most of your effort into correcting the sources.
How do I fix wrong information in Google AI Overviews?
Use the thumbs-down and 'Report a problem' option on the overview, then open the pages it cites and find the one that carries the error. Correct it if it is yours, or ask its publisher to correct it if not. Google says AI Overviews have no special requirements beyond being indexed and eligible for a snippet, so a corrected, crawlable page is the durable fix.
Why does an AI hallucinate facts about my company?
Language models predict likely text, and when they have no clear source for a fact they can produce a plausible guess. OpenAI and Anthropic both say their assistants can produce incorrect or misleading answers. The practical response is to publish the missing fact plainly on your own site so there is something to retrieve instead of a guess.
Can I edit my company's Wikipedia article to correct it?
Wikipedia's conflict-of-interest guideline strongly discourages employees and paid representatives from editing affected articles directly. Disclose your affiliation, then propose the specific change on the article's talk page with the {{Edit COI}} template and a reliable source. There is a narrow exception for defamation or serious errors, which should be followed up through Wikipedia's volunteer response team.
Is there a legal route to correct ChatGPT, Gemini or Copilot?
Each vendor has one. OpenAI has a content reporting form and IP forms, Gemini has 'Report legal issue' under each response, and Microsoft's Copilot terms point to its Report a Concern page. For inaccurate personal information about an individual, OpenAI and Anthropic offer privacy request routes. Ask legal to choose the route; most company errors are better fixed at the source.
How long does it take for an AI answer to change after a correction?
No vendor we checked commits to a timeline. Engines that search the web can reflect a corrected page once it is recrawled, while errors rooted in training data may persist until a later model release. Recheck on a fixed schedule and report the share of reruns that still contain the error.
How do I stop AI engines getting my company wrong in the first place?
Keep a dated facts page, retire or annotate old pages when something changes, maintain your listings and review profiles, and run an accuracy audit on a set of questions about your company each month and quarter. Add AI sources to your launch checklist so a price change or rebrand updates every source the same week.
Is When AI Gets You Wrong: The Correction Runbook free?
Yes. Give a first name and a work email and the download starts at once. We also email you the link, and one person from the Noma team follows up a day later; reply and say so if you would rather not hear from us.
Why do you ask for a work email?
The playbooks are written for teams at companies, and a work address tells us which company is reading, so a follow-up can be about your brand rather than generic. Personal and temporary inboxes are not accepted.
What format is it?
A PDF of 38 A4 pages, laid out to print, with worksheets you can fill in by hand or copy into a document of your own.

More playbooks

Updated 20 October 2026 · Written by the Noma team · All resources