How to correct what ChatGPT and other AI assistants say about your company

Ask the same questions of ChatGPT, Claude, Gemini and Perplexity, and add Grok if your audiences use it. Google AI Overviews is a separate surface: it appears on search results pages and draws on Google’s index, so record it in its own column rather than pooling it with the assistants.

Write a fixed set of benchmark questions and keep the wording stable so later runs can be compared. Cover several kinds of question:

• Direct questions: what the company does, who leads it, where it is based.
• Issue questions: past controversies, lawsuits, recalls, breaches, and whether they are resolved.
• Comparison questions against named peers.
• Unbranded category questions, where an appearance means the assistant volunteered your name.
• Questions about your executives by name.

Ask each question more than once, in a fresh chat, and on different days. Live answers vary between runs, so a single answer is a snapshot. Note whether web search was on and which model the product showed, since both change the answer.

For a dated reference point, the AI Perceptions free check publishes snapshots of which brands five assistants name for unbranded questions in a range of sectors, and which domains they cite.

Your own pages. Publish a plain fact page or newsroom section with the details assistants get wrong: official name, founding date, headquarters, leadership, ownership, and the current status of past issues. Keep leadership bios and the about page current, and update or redirect old pages that still state superseded facts. Write facts as plain sentences, dated where they change.

Structured data. Organization markup (schema.org) on your site states your name, logo, founding date and official profiles in a machine-readable form. It makes the facts unambiguous for crawlers. It does not guarantee how any assistant uses them.

Wikipedia. Wikipedia’s conflict-of-interest guideline strongly discourages editing an article about your own organization. Propose changes on the article’s talk page, cite reliable published sources, and disclose your affiliation. Paid editors must also disclose who pays them, under Wikimedia’s Terms of Use.

Third-party profiles. Business directories, data providers, review sites and employer profiles are often cited and often stale. Claim the profiles you can and update them.

Earned media. If a news story is wrong, ask the publisher for a correction under its corrections policy, with evidence. A correction on the original article helps more than a new story elsewhere, because the original is what gets cited. If your own press release was wrong, issue a correction through the same channel. Where the record is simply thin, credible coverage of the accurate facts gives assistants something better to cite; our glossary entry on earned media covers the basics.

In-product feedback. ChatGPT, Claude, Gemini and Perplexity each have thumbs-up and thumbs-down buttons under an answer, and Perplexity also lets you report an inaccurate answer. Google’s AI Overviews has “Good overview” and “Bad overview” buttons and a “Report a problem” option. Use these on each logged error and say what is wrong and where the correct fact is published. Feedback goes to the provider; it is not a request with a reply.

Privacy requests for individuals. OpenAI accepts privacy requests, through its privacy portal, from people who find inaccurate personal information about themselves in ChatGPT output, and says it considers them based on applicable law and what its models can technically do. That can apply to a named executive. It is a personal-data process and does not cover company facts. Other providers publish their own privacy request routes in their privacy policies.

Legal matters. Potentially defamatory claims belong with counsel, who can advise on the options in your jurisdiction.

Re-run the same benchmark questions, with the same wording, on a fixed schedule: monthly as a baseline, and again after each fix. Answers that use live web search can reflect a corrected page once it has been re-crawled. That can take weeks or longer. Answers from model memory may not change until the provider updates the model, which you do not control.

Report each error as the share of sampled answers that contained it, with the dates, the assistants and the number of answers. Show before and after as two snapshots, and avoid calling a change between two runs a trend, because answers vary from one run to the next.

Running this by hand works for a handful of questions. AI Perceptions runs benchmark questions daily across ChatGPT, Claude, Gemini, Perplexity and Grok and records the domains each answer cites. For how the teams involved split this work, see our guide to GEO for comms teams.

Sometimes the false claim an assistant repeats was put there on purpose. Signs include the same wording appearing across many new or low-quality sites at once, cited pages created shortly before the claim surfaced, and bursts of posts from accounts with little history. Fixing your own sources still matters, but it will not stop a campaign that keeps producing new pages.

PeakMetrics is a narrative intelligence platform that shows communications, public affairs and risk teams how stories about their organization spread across news, social media and AI answers, and flags when those narratives are coordinated or manipulated. A campaign calls for monitoring where the claim is spreading and who is amplifying it. Our glossary entry on coordinated inauthentic behavior explains the signals, and coordinated attack detection describes how PeakMetrics approaches it.