Most teams react to a false AI answer by looking for someone to call. There is no such desk. What you can act on is what feeds the answer, and what you can build is the file that supports a legal step against the vendor if one is ever needed. Reports do not bind vendors, and the law protects a company less than most marketing leaders assume.
So work in five moves: rate the severity, freeze the evidence, find the origin, fix the source, and escalate only what warrants it. This guide picks up where the accuracy audit stops: that piece finds errors, this one deals with them.
How common are errors about brands?
The market does not know. As of 30 October 2026, the research for this article found no large independent study of factual errors about brands: prices, warranties, features. The figures available answer other questions, and most of their spread comes from what they test.
| Study | What it tests | Result |
|---|---|---|
| Tow Center, March 2025 | Tracing press excerpts to their source, 8 tools, 1,600 queries | Over 60% incorrect answers |
| EBU and BBC, October 2025 | Assistant answers about the news | 45% with at least one significant issue, about 20% with major accuracy issues |
| Oumi for the New York Times, April 2026 | AI Overviews, 1,609 SimpleQA questions | About 91% correct, 52% with at least one unsupported claim |
| Xu, Iqbal and Montgomery, May 2026 | AI Overviews, 98,020 claims, trending queries | 88.97% consistent with cited sources, 2.66% contradicted |
Sources listed at the end. Google disputes how representative SimpleQA is.
The takeaway is a consequence rather than a rate: an answer can be right and poorly supported, or lean on a page that does not say what it is credited with.
How should you triage errors by severity?
By the decision the error distorts, not by how often it appears. One question does most of the work: would a customer, candidate or partner act differently if they believed this? Count an error once even when ten answers repeat it, because there is one fact to fix. And separate false from unflattering: a harsh opinion is not a factual error.
| Severity | Examples | Owner | First move |
|---|---|---|---|
| Critical | Invented fraud or lawsuit, safety or health risk, accusation against an executive | Legal and comms | Evidence, then counsel |
| High | Price, warranty, return terms, availability | Marketing and the relevant team | Evidence, then fix sources |
| Medium | Discontinued range, confusion with a competitor, wrong location | Marketing, SEO | Fix sources |
| Low | Approximation that changes no decision | SEO | Monitor |
Grid proposed by this article, illustrative examples.
What should you record before acting?
Everything a screenshot leaves out. An evidence log is the dated record of each false answer: what was asked, where, what came back, and what you did next. It measures whether a fix works, supports a report and feeds any later file.
- Context. Date and time, engine, logged in or not, country, language, exact prompt.
- Answer. Full text, capture, share link if there is one, cited URLs.
- Effect. Related tickets, calls or lost sales, if any.
- Steps taken. Every report, with its date and any reply.
Mind the wording. A loaded question makes weak evidence: Roger Montti of Search Engine Journal notes that 49 of the 56 questions in Ahrefs' made-up brand experiment were leading, with the assumption built into the question itself. Reproduce the error with a neutral question and keep both versions. In a serious case, ask counsel whether a formal record by a bailiff (a French commissaire de justice) is needed.
Where does the error come from?
Not necessarily from the cited page. Xu and coauthors count 6.98% of claims that none of the cited sources they analysed supports. In the Wolf River Electric case, brought by a Minnesota solar installer, the complaint alleges that none of the materials the AI Overview referenced contained the accusation. In its March 2025 study, the Tow Center found fabricated or broken URLs in over half of Gemini and Grok 3 responses. So read each page before contacting anyone. Four situations follow, each with its own remedy.
| What you find | Diagnosis | Remedy |
|---|---|---|
| Your page, a PDF or an old version states the error | Internal source | Correct, redirect, unpublish |
| A third-party site states the error | External source | Correction request, routes below |
| The cited page says something else | Generation error | Clarify your pages, report to the vendor |
| No source, or a URL that does not exist | Unknown origin | Strengthen published facts, report, monitor |
Do not conclude from a single answer. Ahrefs tracked over 43,000 keywords in autumn 2025: between observations, 45.5% of the URLs AI Overviews cited were new. To see which domains keep returning, the piece on the sources each engine reads gives the method.
What should you fix on your own pages?
Anything that contradicts your current position, including what you forgot: old terms, PDFs still online, foreign editions, outdated press releases. It is the one workstream that depends on nobody else.
- One fact, one reference page. Price, warranty, delivery and materials stated in plain text, dated, in the served HTML.
- An explicit denial when the error has taken hold. In Ahrefs' experiment (December 2025), ChatGPT cited the official FAQ in 84% of answers and stayed under 7% misinformation, while Gemini and Perplexity repeated planted stories in 37 to 39%. One fictional brand and leading questions: an illustration, not a rate.
- Access. A corrected page closed to bots does nothing. The choice is covered in the guide to AI bots.
The same experiment might suggest flooding the web with counter-narratives. Rule it out: you would be manufacturing exactly what you object to engines repeating.
How do you get a third-party source corrected?
Ask first, cite a legal text second. Muck Rack's What Is AI Reading? study (May 2026, over 25 million links) attributes 84% of citations to "earned" content, a broad category and a figure from a PR vendor. The direction is still clear: most of the raw material is not yours.
- Media. Start with a factual correction request and supporting documents. In France, online publications are also subject to a right of reply: the request needs proof of receipt, and under the 2007 decree the publishing director need not run the reply if the passage is removed or corrected within three days. Have the filing deadline checked as soon as you find the article.
- Wikipedia. Do not edit it yourself. The conflict-of-interest page asks you to propose changes on the talk page, source them and disclose any paid link.
- Forums, reviews, marketplaces. The DSA requires hosting services to offer a notice mechanism to any individual or entity (Article 16), but for illegal content. Mere inaccuracy does not qualify. Platforms must then offer an internal complaint route (Article 20) and access to out-of-court settlement (Article 21). That leaves a reply in the thread itself, from an identified account: the rules are set out in the guide to Reddit, YouTube and AI sources.
What are vendor feedback channels worth?
They leave a trace without promising a correction. Google says the thumbs-down and "Report a problem" under an AI Overview help improve the product, and acknowledges that these summaries "can and will make mistakes". Its legal form does list AI Overviews and AI Mode, and states that a product-policy report does not serve as legal notice: the two paths are filed separately. Anthropic documents a thumbs-down and a support address.
On OpenAI, noyb reports that the company said in 2024 it could not correct data, only block it on certain prompts. That is the complainant's account and should be read as such.
File the reports and archive them. None of the pages consulted sets a timeframe for a brand.
What legal routes exist in France and the EU?
Fewer than expected, and classification matters as much as substance. As of 30 October 2026, the research for this article found no French decision on an AI answer defaming a brand, which does not prove none exists.
| Route | What it covers | Period written in the text | Limit |
|---|---|---|---|
| GDPR, Articles 5 and 16 | Inaccurate data about an individual | Response within one month, extendable by two | Recital 14 excludes legal persons |
| Defamation, French law of 1881, Article 29 | A fact harming honor or reputation | Three-month limitation (Article 65) | No French decision identified on a generated answer |
| Disparagement, French Civil Code, Article 1240 | Statements about products and services | General five-year rule (Article 2224), to confirm | Can be reclassified |
| AI Act, Articles 50 and 85 | Transparency and marking, complaint to the surveillance authority | No handling period in Article 85 | Article 50 does not address accuracy |
| DSA | Illegal content at hosts, systemic risks of very large services | ChatGPT: obligations by January 2027 | Application to generated answers needs analysis |
GDPR helps only when a person is targeted. A named executive or founder can act; the company cannot. CNIL, the French regulator (February 2025), considers retraining the most effective measure and output filters a fallback, with the required response depending on proportionality. noyb's complaint in Norway (March 2025) relies on Article 5(1)(d); its page reports no decision.
Defamation or disparagement: the choice sets the clock. France's Cour de cassation upheld disparagement for statements about a company's services and products (15 January 2020). It also upheld the defamation classification for an action that went beyond product quality to the honor and reputation of the company and its manager (17 March 2021). A misclassified claim can therefore run into the three months. When that period starts for an answer generated on demand is a question for counsel: as a precaution, count from the first capture.
The AI Act and the DSA are not correction tools. Article 50 requires disclosure and marking, with a grace period to 2 December 2026 for marking by systems already on the market (Regulation 2026/1744). The Commission designated ChatGPT a very large online search engine in August 2026: its page speaks of assessing and mitigating systemic risks, not of correcting an answer for a brand.
US case law is moving but does not transfer. OpenAI won Walters in Georgia in May 2025, with the court pointing to its warnings, and the lack of a retraction request barred punitive damages. In July 2026 in Delaware, at an earlier stage of proceedings, Judge Adams declined to dismiss Robby Starbuck's suit against Google, describing "a new frontier for defamation law". One precaution holds everywhere: request the correction in writing.
How long before a fix takes effect?
There is no public measure. The only sourced figures concern neighboring steps: Google says recrawling can take a few days to a few weeks, with no guarantee of inclusion, and GDPR gives a controller one month to respond.
A clean capture proves nothing either. Ahrefs found that an AI Overview's text differed from the previous observation 70% of the time, with mean persistence of 2.15 days. Measure how often the error appears on the same prompts, across several engines and dates, before and after. The pieces on how many prompts you need and false signals cover what you can conclude.
Where should you start?
- List ten facts that carry a sale or your reputation. Prices, warranties, terms, executives, disputes. That is your reference.
- Open the evidence log before sending a single message, and note the date of the first capture.
- Read the cited pages. Put each error in one of the four situations.
- Fix your pages, then approach third parties. In that order.
- Report, archive, and escalate anything critical to counsel without waiting for the other workstreams.
Steps 1 and 3 become unmanageable once answers run into the dozens. They are what the Accuracy tab of the AI Audit, in beta, covers in Nanga's GEO module. You build a source of truth from your site, reference PDFs and URLs, and facts added by hand. Engine answers are compared with that base, and distinct errors, one per contradicted fact, read by severity. The "Sites behind the errors" view lists sites consulted for an answer holding a wrong claim, and the "Identify the faulty site" action reads those pages to show whether one of them states the claim.
To see which errors surface about your brand, and where they come from: request a demo.
Frequently asked questions
Can I ask ChatGPT or Google to correct an error about my brand?
You can report it, with no assurance of a result. Google's help offers a report under each AI Overview and a separate legal form. It announces neither case-by-case correction nor a timeframe. Keep a record of everything you send.
Can a company use GDPR to get an AI answer corrected?
Not for the company itself. Recital 14 of the GDPR excludes data about legal persons. The route stays open when the answer concerns an identifiable individual, such as a named executive or founder.
How long do I have to bring a defamation claim in France?
Three months, under Article 65 of the French press law of 29 July 1881. When that period starts for an answer generated on demand is a question for counsel. As a precaution, count from your first dated capture.
How long until an AI stops repeating an error?
There is no public measure of this as of 30 October 2026. Google says only that recrawling can take a few days to a few weeks, with no guarantee of inclusion. Measure how often the error appears before and after your fix.
Is the page an AI cites always the origin of the error?
No. In the study by Xu and coauthors, 6.98% of AI Overview claims were supported by none of the cited sources analysed. Read the page before contacting anyone: it may be accurate, with the error arising in generation.
Sources:
- Columbia Journalism Review, Tow Center, AI Search Has A Citation Problem, March 2025.
- EBU, AI governance must put trust, information integrity and child safety front and centre, July 2026, restating the News Integrity in AI Assistants study of October 2025.
- Oumi, Oumi's study finds 50% of AI Overviews, April 2026.
- Xu, Iqbal and Montgomery, Measuring Google AI Overviews, May 2026.
- Ahrefs, AI vs made-up brand experiment, December 2025.
- PPC Land, What Ahrefs' fake brand experiment actually proved about AI search, restating Roger Montti's analysis (Search Engine Journal) of 28 December 2025.
- Ahrefs, AI Overview change, November 2025.
- Muck Rack, Earned media still drives 84% of AI citations, May 2026.
- Google, AI Overviews, Search Help, accessed 30 October 2026.
- Google, Report content on Google, Legal Help, accessed 30 October 2026.
- Google Search Central, Ask Google to recrawl your URLs, accessed 30 October 2026.
- Anthropic, Claude is providing incorrect or misleading responses, accessed 30 October 2026.
- Wikipedia (French edition), Conflit d'intérêts, accessed 30 October 2026.
- noyb, AI hallucinations: ChatGPT created a fake child murderer, March 2025.
- CNIL, IA : respecter et faciliter l'exercice des droits des personnes concernées, February 2025.
- GDPR, Recital 14, Article 5, Articles 12 and 16.
- Légifrance, law of 29 July 1881, Article 29 and Article 65.
- Légifrance, Code civil, Article 2224.
- Légifrance, Decree no. 2007-1527 on the online right of reply.
- Cour de cassation, commercial chamber, 15 January 2020, no. 17-27.778 and 17 March 2021, no. 19-20.459.
- AI Act, Article 50, Article 85 and Digital Omnibus, Regulation (EU) 2026/1744.
- DSA, Regulation (EU) 2022/2065, Articles 3, 16, 20 and 21.
- European Commission, Commission designates ChatGPT, Reddit, Roblox under the Digital Services Act, August 2026.
- Loeb & Loeb, Walters v. OpenAI LLC, May 2025.
- Claims Journal, Starbuck v. Google, July 2026.
- Reason, The Volokh Conspiracy, Google missed key deadline in suit alleging Google's AI libeled business, January 2026.