Brand Reputation in AI - What to Do When ChatGPT Says Something Untrue About Your Company

In March 2025 a customer of Wolf River Electric, a Minnesota solar company, cancelled a contract worth $150,000. The reason: Google's AI Overview claimed the company was being sued by the state attorney general for deceiving customers. It wasn't true - the attorney general had sued four other companies in the industry, and the model stitched the facts together. Wolf River filed a defamation suit seeking $110 million to $210 million from Google. You don't need a lawsuit to have a problem, though: according to a July 2026 Searchable study, 93% of 1,704 brands checked had at least one basic fact wrong or missing in AI answers.

In a July 2026 Searchable study, 93% of 1,704 brands had at least one fact wrong or missing in AI answers, and small businesses got fabrications far more often. Where models get untrue facts about a brand, how to audit them, five correction paths and what the law says when an error costs a contract.

This post is about what to do when ChatGPT, Gemini or AI Overviews say something untrue about your company. I deliberately separate it from the AI visibility audit, which checks whether a brand shows up in answers at all. The question here is different: what the models say about you, where they get it from, and how to correct it - from fixing the source, through making your own site the source of truth, to reporting channels and the legal route when the error is serious.

How often AI gets brands wrong

The Searchable study is valuable for one reason: each of its more than 32,000 facts was graded by an employee of the company the fact described, not by an automated check. The result: 9.2% of facts were false, and three in four brands that checked thoroughly found at least one falsehood about themselves. By platform: ChatGPT 8.1% false facts, Perplexity 11.5%, AI Overviews 11.4%. Small businesses are the most exposed - one in two got at least one fabricated fact, versus 32% of large companies, simply because there's less data about a small company online for the model to draw on.

Wider context comes from the EBU and BBC report from October 2025: 22 public service broadcasters from 18 countries assessed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity to questions about current events. 45% of answers had at least one significant issue, most often poor sourcing. If assistants get almost half of answers wrong about news that is well documented online, the risk of error only grows for a lesser-known brand.

Where models get wrong facts about your brand

To correct effectively, you need to know which mechanism produced the error. A model draws on two sources of knowledge, and each is fixed differently.

• Training knowledge - what the model "remembered" from the data it was trained on. It's frozen at the data cutoff and changes only with the next model version, so a fix takes effect with a delay measured in months.
• Real-time search (grounding) - when ChatGPT, Perplexity or AI Overviews search the web and summarize the pages they find. Here the error usually has a specific source visible in the citations, and fixing it at the source can work within days or weeks.

In practice, errors fall into four types. Outdated facts (an old price list, a former CEO, a closed branch) come from old pages still in the index. Confused entities are the Wolf River case - a similar name, the same industry, the same region, and the model attaches someone else's history to the company. Fabrications are classic hallucinations, when there's too little data and the model answers anyway. Inherited negatives are old forum threads, single reviews or outdated articles the model treats as the company's current picture. Each type has a different source, so a different repair path - starting with a report to OpenAI when the error sits in an old business directory is a waste of time.

Audit: what AI says about your brand

Before fixing anything, gather the facts. Ask the models the same set of questions and record the answers with the date, model name and mode (with or without search). The questions fall into three groups: facts, reputation and comparisons.

FACTS- What does [brand] do? Since when, and who runs it?- How much do [brand]'s services/products cost? Where is it based and how do I contact it?REPUTATION- What do people say about [brand]? Have there been any problems, lawsuits or scandals?- Is [brand] trustworthy? What are its weaknesses?COMPARISONS- [Brand] or [competitor] - which should I choose and why?- Who do you recommend for [service] in [city/country]?

Rate each falsehood you find by severity. It decides how fast you react and through which channel:

For every answer produced with search, also record the cited sources - they show what needs fixing. Answers without search have no sources, but the same error repeating across several models suggests it sits in the training data.

Correction paths - most effective first

1. Fix the source. If the answer cites a specific page, start there: an old article, a business directory, a review aggregator, a profile on an industry site. Asking the publisher for an update or correction often works faster than any intervention with the AI provider, because once the page is re-indexed, the model simply finds the new version when it searches.

2. Make your own site the source of truth. Models reach for sources that answer the question unambiguously. If your site has no clear answer to "how much does it cost" or "who runs the company", the model will take it from somewhere that does - even if it's five years old. In concrete terms: an "About" page with facts (founding year, management, locations, scope of services), an FAQ answering the audit questions, a pressroom with current information, and Organization structured data with sameAs links to official profiles. For a specific false allegation, a separate, factual page with an explanation works well - written to answer directly the question users are asking.

3. Make profiles and knowledge bases consistent. Google Business Profile, Merchant Center, LinkedIn, Wikidata and, if the company meets notability criteria, Wikipedia. I described how to build and maintain a Wikidata record in the post on Wikipedia, Wikidata and the Knowledge Panel. Conflicting data in different places tells the model that none of them can be fully trusted.

4. Report the error to the provider. Every platform has a reporting channel, but you need to know what each can actually do. In AI Overviews, use the thumbs-down icon and the report-a-problem option, and for content that breaks the law, a separate legal removal form. OpenAI accepts requests about personal data through its privacy portal, but its help pages state plainly that it handles them according to applicable law and the technical capabilities of its models; instead of correcting a fact, it may block information about a person from appearing. That channel protects people, such as a CEO, not the company as a whole - for errors about the business, paths 1-3 are more effective.

5. Crowd out bad sources with good ones. If a negative picture rests on a few old texts, the most durable answer is new, current and credible mentions: trade media coverage, interviews, rankings, honest reviews. I described how to earn such mentions in the post on digital PR, and community presence in the post on Reddit and forums.

It's worth setting honest expectations: an error based on search usually disappears within days or weeks of fixing the sources, while an error stored in the model's knowledge can persist until the next model version. That's why you do paths 1-3 right away and measure the effect separately for answers with and without search.

When the error is serious - what the law says

Courts are only now drawing the lines of liability for AI output, and the first decisions point in different directions. In Walters v. OpenAI, a Georgia court dismissed the defamation claim in May 2025, finding among other things that a reasonable reader wouldn't take the output as a statement of fact, that OpenAI hadn't acted with actual malice because it warns about errors, and that the plaintiff showed no damages. In Robby Starbuck's cases, Meta settled in 2025, and his suit against Google survived a motion to dismiss in July 2026 and continues. The takeaway for a business: documented harm - like Wolf River's lost contract - and documented attempts at correction are what count.

In Europe there's the GDPR dimension. In March 2025 the organization noyb filed a complaint against OpenAI in Norway on behalf of a man ChatGPT had described as the murderer of his own children, citing the data accuracy principle (Art. 5(1)(d) GDPR). The GDPR protects natural persons, though - the owner, the CEO, the company's expert - not the company itself. A company in Poland can rely on the protection of the personal rights of legal persons (Art. 43 in conjunction with Arts. 23 and 24 of the Civil Code), including its good name; other jurisdictions have their own defamation and trade libel rules. A lawyer will assess the chances in a specific case; my part ends at making sure the evidence is collected properly:

• a screenshot with the date, time and exact question,
• the model name and mode, plus a link to the shared conversation if the platform allows it,
• the test repeated across several sessions to show the error wasn't a one-off,
• copies of reports to the provider and their replies,
• evidence of harm: lost inquiries, cancelled contracts, messages from customers citing the AI answer.

Monitoring - so you find out before your customers do

A one-off audit shows today's state, and model answers change with every update and every new text on the web. A fixed set of questions asked monthly across several models, with automated grading of facts and sentiment, turns AI reputation from a surprise into a measurable metric. I described the measurement method in the post on Share of Voice in AI, and compared ready-made tools in the post on AI visibility monitoring tools.

That's how AnswerLyzer works, a system I built on the LLM-as-a-judge pattern: it asks the models a set of questions about the brand, and a separate model grades each answer against defined criteria - whether the brand appears, in what context and with what sentiment. For reputation, one function matters most: an alert when a new negative claim appears in the answers, or a fact that contradicts the company's fact list. That list, a few dozen verified sentences about the brand, is the same material that goes onto the "About" page and into the FAQ.

Step-by-step action plan

1. 1.Write the company fact list - 20-40 verified sentences: what you do, since when, who runs it, where you operate, what key services cost.
2. 2.Run the audit with the diagnostic questions in ChatGPT, Gemini, Perplexity and AI Overviews, with and without search.
3. 3.Rate the severity of each falsehood and set the order of response.
4. 4.Identify the sources of errors from the citations in the answers and ask publishers for corrections.
5. 5.Expand your own site: facts, FAQ, pressroom, Organization structured data with sameAs.
6. 6.Make profiles consistent: Google Business Profile, LinkedIn, Wikidata, industry directories.
7. 7.Report critical errors to providers and keep full documentation.
8. 8.Build new, credible mentions that crowd out old, negative sources.
9. 9.Set up monthly monitoring with alerts for new negative claims and contradicting facts.

---

I help companies check what AI models say about them and correct the errors - from the audit and fact list, through fixing sources and structured data, to ongoing monitoring with alerts. I do this as part of AI optimization (GEO). Get in touch - I'll start with an audit of the answers about your brand across four models and a list of errors ranked by severity.

Worth reading next:

• AI Visibility Audit - Does ChatGPT Recommend Your Brand?
• Wikipedia, Wikidata and the Knowledge Panel - How a Brand Becomes an Entity AI Trusts
• Digital PR and Brand Mentions - The New Currency of Visibility in Google and AI Answers
• How to Measure Brand Share of Voice in AI Models