Reputation in AI Answers: What to Do When ChatGPT or Google AI Overviews Get It Wrong

Just three years ago, the path to information about a person or company looked the same everywhere: a query in Google or Yandex, ten blue links, and the reader deciding for themselves what to believe. Increasingly, that's not how it works anymore — people ask ChatGPT directly, Google serves up a ready-made AI Overview right above the results, and some users never make it to a website at all, settling for a single generated paragraph. For reputation, this isn't a cosmetic interface change. It's a new authority that shapes opinion about you in a single answer — and, as lawsuits that have already happened show, that authority gets things wrong regularly, and sometimes very seriously.

When AI Makes Things Up: This Is No Longer Hypothetical

Over the past three years, enough documented cases have piled up to talk about "defamation by hallucination" as its own, now quite real, category of dispute.

In the US, a radio host sued OpenAI after ChatGPT, responding to a journalist's query, claimed he had "embezzled funds" from a nonprofit — something that never happened. The case reached court and was dismissed in May 2025: the judge found that no reasonable reader would take a chatbot's answer as a statement of fact, and the plaintiff couldn't prove actual damages. The ruling is telling in its own right: the court didn't say "ChatGPT was right" — it said "lawsuits like this are hard to win right now," which isn't the same thing.

An activist in the US sued Meta after its AI claimed he had taken part in the Capitol riot and had ties to a movement he had nothing to do with — the suit was settled out of court in August 2025. A Norwegian man filed a complaint with a supervisory authority after ChatGPT generated a story claiming he was serving a sentence for murdering his own children — no such conviction ever existed, of course. Microsoft Copilot went as far as confusing a German court reporter with the subjects of his own articles about criminal trials, portraying him as an escaped psychiatric-hospital criminal.

Finally, 2026 brought the first genuinely precedent-setting court ruling: two German publishers won a decision from a Munich court that Google bears direct liability for the content of AI Overviews as its own statement — and can't hide behind its status as a "neutral search engine" when the text in question is generated. Google is appealing the ruling, but the mere existence of this first-of-its-kind precedent is a signal that regulatory uncertainty around AI answers is coming to an end.

Why You Can't Just Ask AI to "Fix Itself"

Intuitively, it seems that if a chatbot says something false, filing a complaint should be enough — the same way you can request removal of an inaccurate webpage. In practice, it works differently, and the difference is fundamental.

A regular webpage has an author, a hosting location, and text that physically exists in one place — it can be edited or deleted. A language model's answer has no "page": it's a statistical output of the model's weights, sometimes supplemented with real-time search results. OpenAI's own privacy policy admits as much: "Given the complexity of how our models work, we may not be able to correct the inaccuracy in every instance" — and offers, instead of a targeted factual correction, the option to request removal of personal-data mentions from the model's output entirely. Those are different things: "stop mentioning" is one request, "fix this specific falsehood and keep answering on the merits" is a completely different one.

The picture is similar at the other major players:

The takeaway is simple: everyone has a formal complaint channel, but none of them functions as a guaranteed "removal" or "correction" — at best, it's a signal to a company that may or may not act on it.

Regulators Are Already on This — But No Rulings Yet

The GDPR formally provides grounds to demand accuracy of personal data (the accuracy principle, Article 5) and its correction (Article 16) — and the advocacy group noyb has been filing complaints against OpenAI on this basis for several years now. In Poland, an investigation into a complaint over a false "biography" generated by ChatGPT has been running since fall 2023. In Austria, a similar complaint over a public figure's incorrect birth date was filed in April 2024 and referred to the Irish regulator — OpenAI's lead supervisory authority in the EU. A Norwegian complaint over a fabricated "child murderer" story was filed in March 2025. As of mid-2026, none of these has produced a final decision.

The only actual fine OpenAI has received from a European regulator — €15 million from Italy's Garante in late 2024 — was imposed for lacking a legal basis for using data in training and other violations, not for hallucinations as such. In March 2026, a Rome court overturned that fine entirely, on procedural grounds (jurisdiction should have passed to Ireland), without ruling on the merits.

The upshot for anyone managing a reputation: the legal lever exists in theory but hasn't actually worked even once yet. Counting on a regulator to "fix everything" within a reasonable timeframe is premature — de facto, we're talking years, not weeks.

Where AI Actually Gets Its Facts — and What SEO Has to Do With It

There's a technical fork in the road here that determines what can actually be done.

Some answers are "live": Google's AI Overviews and ChatGPT's search feature query the search index in real time and cite up-to-date pages for a specific query. Google itself, in its official guidance on optimizing for AI search features, states outright that AI Overviews "rely on our core ranking and quality systems used in Search" and that "there's nothing specific you need to do to appear in AI Overviews" — in other words, Google's official position is that classic SEO is the optimization for AI answers.

But there's an important detail here that isn't the most comfortable one for the SEO reputation industry: the link between a top-10 position and getting cited in an AI Overview has weakened noticeably over the past year. According to Ahrefs data, back in summer 2025, 76% of AI Overview sources came from Google's regular top-10 results — by March 2026, that share had dropped to 38%, with more than a third of cited pages not even ranking in the top 100. Technically, this is explained by the fact that modern models generate several parallel sub-queries ("query fan-out") and cast a much wider net than before, while systems like Perplexity build their own independent search index entirely, without relying directly on Google's results.

The practical takeaway: ranking at the top of search results is still necessary for managing what AI says — but it's no longer sufficient on its own. Work on search visibility needs to expand to a broader set of sources than classic SEO reputation management used to require.

The other kind of answer is "frozen": a fact baked into the model's weights at training time, with no connection to the current state of the internet. That kind of fact won't update until a new model version ships — and fixing anything on websites is pointless here: the source of the distortion isn't on the internet, it's inside the model itself.

One of the key sources that shapes a structured picture of an entity in these systems is Google's Knowledge Graph — the same graph that also powers the Google Knowledge Panel for a person or company in the right-hand column of search results. The panel and a generative answer are different formats of the same digital entity, and the knowledge panel has its own logic for obtaining and correcting data, separate from general SEO reputation management: more in the article on Google Knowledge Panel for a Person or Company.

What Actually Works

All of the above adds up to a workable plan of action — not an idealized one.

First, classic SEO reputation logic still works — correcting and suppressing the sources that live search draws on — but adjusted for the fact that the circle of relevant sources is now wider than the top 10 and includes more independent outlets, profiles, and authoritative mentions. Second, it's worth building a consistent, verifiable profile across the internet: when the same accurate information is confirmed by several independent, authoritative sources, the model has less reason to cite an outdated version. Third, monitoring can no longer be limited to Google and Yandex rankings — it's worth regularly checking what ChatGPT, Gemini, Copilot, and Perplexity say in response to your name or your company's name: these answers can diverge from each other and from classic search results. A narrower, separate version of that divergence happens when a model doesn't fabricate a fact about you at all, but statistically confuses you with a different, real person; we cover the mechanics of that mix-up and its correction channels in our article on AI entity confusion.

One industry case is worth noting separately: according to an account published in Forbes Business Council (October 2025), after correcting information on a website and in profiles, changes in AI Overview text became visible within roughly a week — but this is a single, anonymized example, not a confirmed pattern, and no one should rely on that specific timeline. If the incorrect fact is baked into the model's frozen knowledge rather than pulled from live search, no amount of work on websites will change it directly — only the next model update will, and no outside party controls when that ships.

Finally, for genuinely serious cases — like the ones mentioned above — a legal claim against the platform itself becomes a real, if difficult, path: most such suits so far are either dismissed in the company's favor or settled without an admission of fault, but the number of these cases is growing, and the Munich ruling against Google shows courts are willing to revisit conventional assumptions about search engine liability.

See Also · Compliance Databases World-Check and PEP Status: How to Remove an Inaccurate Record The GDPR right to rectification already works in practice in an adjacent domain — one where there's an operator on the other end obligated to respond within a deadline. Read the article →

The monitoring described above — tracking what ChatGPT, Gemini, Copilot, and Perplexity say — isn't just routine brand hygiene. In an acute crisis, when the clock runs in hours rather than weeks, AI answers are precisely what can lock in an inaccurate summary of events faster than classic search results stabilize — and watching them should start at hour zero, not after the first wave of publications.

See Also · Crisis Management Reputation Crisis Management: The SEO Reputation Protocol for the First 72 Hours After a Data Leak or Scandal Monitoring AI answers in the first hours after a leak or scandal — running alongside the legal and communications tracks. Read the article →

Common Mistakes

• The first mistake — applying classic SEO logic to AI answers one-to-one, expecting that the #1 position in Google will automatically become the chatbot's correct answer.
• The second — treating the "Report a problem" button as a working removal channel rather than general product-quality feedback.
• The third — focusing only on Google while ignoring what ChatGPT, Gemini, and other systems say about you: they increasingly diverge from one another.
• The fourth — confusing "opting out of having your data used for training" with "correcting a specific fact in the existing model": these are different requests with different odds of success.

How Long This Actually Takes

If a fact comes from live search rather than being baked into training data, it's reasonable to expect changes at the source to show up within a few days to a few weeks — but with no guarantee and no single accepted methodology that yields a precise timeline. If the fact is part of the model's frozen knowledge, there's no defined correction timeline at all — it depends on the release cycle for new model versions, which no outside party influences. Judging by the current unresolved cases in Poland, Austria, and Norway, the regulatory path is measured in years, not months. A legal claim against the platform is a separate, lengthy, and so far largely unsuccessful strategy, applicable only to genuinely serious, provable cases.

Frequently Asked Questions

What to Do Next

Managing reputation in the age of AI search isn't a replacement for SEO reputation management — it's its logical extension: the same principle of managing sources, applied to a wider range of outlets and supplemented with dedicated monitoring of what ChatGPT, Gemini, Copilot, and Perplexity specifically say about you or your company.

Aegis Reputation Lab tracks not just Google and Yandex rankings but also how the major AI systems answer about clients, and structures its work so that source-level corrections show up for both live search and classic search results. Contact us to discuss your situation.

AI answers are just one of seven independent layers that make up digital reputation: alongside it live search results, entity data, compliance databases, deal due diligence, deepfakes, and crisis events — each checked by a different party under its own rules. For the full map of all seven layers, with who checks each one and when, see "Reputational Risk Map for Individuals and Companies."

Service Online Mention Monitoring Discuss →

See Also

GEO Audit

AI Due Diligence: What ChatGPT Tells Investors About You

Read the article → SEO Reputation

SEO Reputation Management: How to Suppress Negative Content from Google and Yandex Top Results

Read the article → Wikipedia

Removing a Wikipedia Article: How It Actually Works

Read the article → Compliance

Digital Profile and International Compliance: How Your Online Reputation Affects Bank Accounts, Visas, and Deals Abroad

Read the article → Compliance Databases

World-Check and PEP Status: How to Remove an Inaccurate Record and Restore Bank Trust

Read the article → Crisis Management

Reputation Crisis Management: The SEO Reputation Protocol for the First 72 Hours After a Data Leak or Scandal

Read the article →

Show more