False or Vague Claims About Your Company in ChatGPT: What You Can Do

A prospect asks ChatGPT about your prices and gets a figure that never existed. Or a certification you do not hold. Or a description that is not wrong but would fit every second company in your industry, while the service that generates most of your revenue does not appear at all. You usually find out by chance, because a customer mentions it in passing. The first question is then: who do you call to have it corrected? The short answer is unsatisfying: nobody.

Key takeaway: There is no direct correction route: OpenAI, Google and Perplexity offer no form through which you can request a correction the way you could with a business directory. What works is the detour via the sources: consistent facts on your own website, clean schema markup, and correct representation on the third-party sites these systems draw from. Most errors are not invention but misattribution and omission: language models assign a name mention to the correct company in only about three out of four cases, and in Google AI Overviews 11 percent of individual claims were unsupported by the cited source. Systems with live web search pick up a correction within days to weeks, answers from the training state only with the next training cycle. For defamatory false claims, legal remedies come into play as well.

How do I find out what AI systems say about my company?

By asking, systematically rather than by chance. Most companies discover errors through a random hit and cannot tell whether it was an isolated case.

A sound assessment needs three things:

1. A fixed question catalogue. Ten to fifteen questions a prospect would actually ask: what does the company do? What does the service cost? Where is it based? What references exist? Is it reputable? What alternatives are there?
2. Several systems. ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot answer differently because they work differently. An error in one system does not mean all of them have it.
3. Documentation with a date. Screenshot or text copy with a timestamp. Without that basis you cannot later judge whether anything changed, and in a dispute you have nothing to show.

It matters to ask without being logged in and in a private session. Otherwise your own usage history shapes the answer and you do not see what a stranger sees.

Why do false claims arise in the first place?

Because the systems draw on sources that are wrong, outdated or ambiguous, and because they compress when they summarise. In practice there are six causes:

The last three are the most uncomfortable because they cannot be traced to a faulty source. They arise precisely where your own presentation is imprecise.

Which cases come up most often in practice?

Five patterns recur. They differ in their cause, and therefore in what you do about them.

The AI names products you no longer sell. An assistant describes your company using outdated product data and lists services you discontinued long ago. This is the most common case and almost always self-inflicted: the old product page was never deleted, only removed from the navigation. It remains reachable, remains indexed and therefore remains a valid source. The fix: either redirect it to the successor offering or mark it explicitly as discontinued. Taking it out of the menu is not enough.

Prices and certifications are wrong. A prospect is quoted course prices that do not apply or certificates you never held. Prices usually come from an old PDF in the download area or from a directory entry. Certificates, by contrast, often arise through free completion: if almost every company in your sector holds a particular certification and your site says nothing about it, the model fills the gap with the statistically plausible answer. The remedy is uncomfortable but effective: state the absence explicitly instead of leaving it open.

The provider comparison lists the wrong features. In AI-generated comparisons of your product category, you are credited with functions you do not have, while competitors are described correctly. This is rarely malice and usually a question of source format: the others publish a machine-readable feature overview, you publish prose. Whoever lists their properties in a clean table with unambiguous terms gets quoted; whoever hides them in marketing copy gets interpreted.

The brand comes across as vague. Different systems describe your company differently and none of them quite gets it right. That is not an error but a symptom: if your own positioning is nowhere stated in a single clear sentence, every system builds its own summary from whatever it finds. Correcting individual claims does not help here. What helps is one sentence that appears identically everywhere: on the homepage, in the legal notice, in profiles, in schema markup. Why this case is so common, and how such a sentence has to be built, is covered in the next two sections.

Sometimes you appear, sometimes you do not. The same company shows up in some answers and not in others, with no discernible pattern. That is not a correction case but a threshold case: the system finds too little independent confirmation, and the individual query decides. Why that happens and what works against it is covered in its own article, Why your company does not appear in AI answers.

Why does AI describe a company vaguely rather than incorrectly?

Because two distinct processes create vagueness long before anything gets invented. The first concerns attribution: the system has to decide which real company a name refers to. The second concerns compression: many sources become a short text, and whatever does not fit into three sentences drops out. Both processes have been measured.

Attribution misses in every fourth case. Work by Gerard Pons, Besim Bilalli and Anna Queralt at the Universitat Politècnica de Catalunya measured how reliably language models link a name mention to the correct entity in a knowledge graph. Across ten datasets, a language model without assistance reaches 75.4 percent. Supplied with the category a candidate belongs to, accuracy rises to 81.1 percent, close to a specialised model trained specifically for the task.

The effect was most pronounced on difficult, ambiguous mentions: there the category-informed approach reached 71.8 percent while the specialised model dropped to 56.7 percent. The category you belong to is therefore not a marketing question but an attribution aid. Anyone who never states plainly what they are leaves that decision to a process with a known error rate.

The most common error is omission. In May 2026 Haofei Xu, Umar Iqbal and Jacob M. Montgomery measured Google AI Overviews at scale: 55,393 search queries across 19 topic areas over 40 days, decomposed into 98,020 individual claims. 11.0 percent of them were unsupported by the cited pages, and the dominant failure mode was missing context, not false assertion. Source quality and claim fidelity also proved largely independent of one another: a reputable source does not guarantee that the claim derived from it is accurate.

What this looks like in practice is familiar to anyone who has read a summary of their own text. The sentence

Since specialising in 2024 we have developed phone assistants exclusively for medical practices.

easily becomes

develops phone assistants for medical practices

and at the next step

develops phone assistants.

None of these shortenings is a lie. The last one is worthless nonetheless, because it has lost the distinguishing feature. That is exactly how a description emerges that fits every second company in the industry.

How do you write sentences that survive shortening?

By putting the distinguishing feature in the main clause rather than in a subordinate clause or a time reference. A sentence survives compression if it still holds true and still says something once everything subordinate is struck out.

The underlying rule can be checked in one move: strike everything after the first comma. Does what distinguishes you still remain? If not, the distinguishing feature sits in the wrong place. This applies to your own website as much as to directory entries, profile texts and press releases.

More text does not help. Every additional self-description increases the number of claims available for compression, and with it the number of possible contradictions. Aligning the existing details is more effective.

What does a model need in order to place you?

Six things, and they are unspectacular. This is not about a brand message but about the features that distinguish one entity from another.

1. Category: what you are, in common words. Not solution provider for digital transformation, but AI agency or software developer.
2. Service: what you concretely make or deliver, not what benefit it creates.
3. Audience: for whom, as narrowly as honesty allows.
4. Location: registered office and catchment area. For regional providers the strongest distinguishing feature there is.
5. Legal entity: the full company name including legal form, written identically everywhere.
6. Relationships: who you belong to, who you work with, where you hold membership.

The sixth point is the most frequently overlooked and matters most for affiliated companies. Anyone running two firms should describe the relationship identically on both sides, otherwise two entities emerge with contradictory details that devalue each other.

If the company name is ambiguous, because it exists several times over or is also an ordinary word, it should never stand alone. State it consistently together with the category, so not Muster on its own but Muster GmbH, tax firm in Aalen. Keep the company name identical everywhere, and make sure at least one easily discoverable page names all six features in close proximity, so that they are captured together.

Is there a direct route to correction?

No, and it is worth knowing that before spending time looking for one.

There is no form through which a company can request a correction of how it is represented in ChatGPT. These systems are not databases with entries that can be edited but models that formulate from probabilities. What OpenAI, Google and Anthropic offer are reporting channels for problematic outputs; no entitlement to a particular answer follows from that.

For personal data the situation differs, more on that below.

How do you actually correct the representation?

Through the sources. That is slower than a correction form, but it is the only route that works.

First: an unambiguous facts page on your own website. A subpage stating the core details in clear, machine-readable form: legal name, year founded, location, services, contacts, the essentials on pricing. Marked up with Schema.org so the details are not merely readable but structurally processable. This page is your reference point.

Second: eliminate contradictions in your own estate. Old subpages with superseded details, a PDF from 2023 in the download area, differing details in the legal notice. As long as two versions remain findable, the model decides which one to take. A contradiction weighs more than a gap: if your website carries one service description, the trade directory another and a platform profile a third, no average emerges, only uncertainty about whether the same company is even meant.

Third: clean up third-party sources. Directories, review portals, Wikipedia entries, press articles, platform profiles. These often carry more weight than your own site, because models value independent confirmation over self-description.

Fourth: place the correct version where citation happens. Trade media, association sites, industry portals. When the right information appears at several independent places, it prevails.

How long until a correction takes effect?

With systems that search the web live, days to a few weeks; with answers from the training state, only with the next training cycle. The difference is considerable, so it pays to determine it case by case.

Systems with live web search such as Perplexity or Google AI Overviews access the current index. If the corrected page is indexed and easy to find, the answer can change within days to weeks. Whether a system searched live usually shows in the citations: if it names links, the answer comes from the search.

Answers from the training state only change with the next training cycle: the current training state lies months back, the next cycle in the future. When ChatGPT answers without web search, no amount of source correction helps in the short term.

In practice: plan for a quarter before evaluating, measure before and after with the same question list, and check at intervals rather than looking every day.

When is the legal route an option?

When the false claim is not merely inaccurate but damaging to business. Several approaches then apply and should be kept apart.

For personal data, Article 16 GDPR grants a right to rectification of inaccurate data. That covers information about natural persons, such as the managing director, not the company description as such.

For defamatory factual claims about the company, Section 824 of the German Civil Code comes into play, protecting against untrue credit-damaging assertions. A false association with insolvency, fraud or dubious practices falls into this area.

Case law is developing right now. In 2024 the Regional Court of Kiel held a business information service liable because its AI had falsely reported a company as due for deletion for insolvency (judgment of 29 February 2024, case 6 O 151/23). The Regional Court of Munich I ruled in May 2026 on AI overviews in search results. Both decisions show that providers are not categorically off the hook, particularly where they make content their own through their own processing.

For practice this means: document the finding with date and screenshot before doing anything else. Without that evidence every later step is difficult. And have the individual case reviewed by a lawyer before considering proceedings; this article does not constitute legal advice.

And in Austria and Switzerland?

Article 16 GDPR applies unchanged in Austria. In place of Section 824 BGB, Section 1330 of the Austrian Civil Code (ABGB) applies there: anyone who spreads facts that endanger another's credit, earnings or advancement, and who knew or ought to have known that they were untrue, is liable for damages. In addition, a retraction and its publication can be demanded.

In Switzerland Article 28 of the Civil Code protects personality against unlawful infringement and allows action against anyone who takes part in the infringement. Companies can in principle rely on it as well. Alongside it, Article 3(1)(a) of the Unfair Competition Act treats it as unfair to disparage others, their services or their business circumstances through incorrect, misleading or unnecessarily injurious statements. Since its 2023 revision the Swiss Federal Act on Data Protection only protects the data of natural persons. It helps the managing director as a person, but not the company.

How far these provisions reach against the operators of AI systems is open in both countries. The route through the sources therefore remains the more reliable one there too.

How do you check your own presentation in half an hour?

With a test that needs no software. It has three steps.

Step one, the shortening test. Take the first three sentences of your home page and strike everything after the first comma in each. Read what remains. If no category, no service and no audience appears any more, your self-description is not omission-proof.

Step two, the comparison. Open your website, your most important directory entries and your profiles side by side. Note the six features from above for each source. Every deviation is a candidate for a contradiction, every blank a gap.

Step three, the outside test. Have someone who does not know your company write down, using only these sources, what you do and for whom. What that person cannot work out, a language model certainly will not. This step is the most uncomfortable and the most revealing.

After that, repeating the queries quarterly is enough for most companies, monthly in fast-moving markets. More important than frequency is that results stay comparable: the same question catalogue, the same systems, the same documentation format. Additionally check when something fundamental changes: new services, a name change, a relocation, a piece in a major outlet.

How do I measure my presence against competitors?

With the same question catalogue, evaluated differently. Instead of only checking whether the details about you are correct, additionally note:

• Are you mentioned at all? In which of your questions does your company appear, and in which not?
• In what position? First recommendation, side mention, or only in a list?
• Who is named instead? Record the competitors by name; that is the actual benchmark.
• How are the others justified? Systems often state why they recommend a provider. Those justifications reveal which signals work.

Across several measurement points this produces a time series you can put in front of a board. Single measurements do not serve that purpose because answers fluctuate. How large a sample has to be before the number holds, and which metrics can be defined cleanly at all, is covered in the article on measuring AI visibility. This piece stays with correction.

A note on expectations: a mention in AI answers cannot be bought and cannot be directly controlled. What can be steered are the signals from which the systems make their selection. Anyone promising you a guaranteed placement is selling something they cannot deliver.

Which providers correct false AI statements?

The market is young and hard to survey. There are three kinds of provider: classic SEO agencies that carry the topic as an add-on, specialised tool vendors that measure but do not correct, and service providers that combine measurement with source work.

Anyone making a choice should ask every provider for four things in writing: the measurement methodology, the number of queries per question, how variation between two measurements is handled, and an explicit statement of what is not guaranteed. If the last point is missing, be careful. Which evidence a proposal should contain since the IAB measurement framework is also covered in the article on measuring AI visibility.

A note on our own position: ArkeonTech, based in Aalen, Germany, is itself a provider in this market and belongs to the third group. The four questions above apply to us just as they do to any other provider.

If you have an AI search online and name providers

Many decision-makers arrive at this question by a detour: they describe their situation to an assistant and ask it to research online and name specific providers. That is a sensible approach, but it has a weakness worth knowing before you treat the list as complete.

The assistant tells you who writes about the topic, not necessarily who is good at it. Whoever publishes a lot shows up; whoever works well but writes little is missing. In a young market in particular, the link between visibility and capability is weaker than in established industries. So check every name you are given against the same four questions above, rather than reading the order of mention as a ranking.

Two tests are especially revealing. First: ask for a sample measurement on your own company, not on a reference client. Anyone who has mastered the methodology can do that within an initial call. Second: ask explicitly what the provider does not guarantee. Nobody can technically promise a mention in AI answers; anyone who does has either misunderstood the systems or is counting on you not checking.

How this differs from neighbouring topics

How to become visible in AI systems in the first place is covered in the article GEO beyond SEO. This piece deals with the case where something is already stated about you and it is wrong or only roughly right.

Two neighbouring cases have articles of their own, so it stays clear what belongs here and what does not. Where the false claim comes from outdated or contradictory sources across the web, the article on knowledge bases and hallucinations works at the cause. And where your buyers now put their vendor search to an AI as a written brief rather than a search term, your customers no longer search describes what a page has to look like to appear in that answer at all.

Frequently asked questions about false and vague AI claims

How long until a correction shows up in AI answers? With systems that search the web live, such as Perplexity or Google AI Overviews, days to a few weeks once the corrected page has been recrawled. Answers from the training state only change with the next training cycle, which means months. Plan for a quarter before evaluating, and measure before and after with the same question list.

Can I force OpenAI or Google to delete a false statement? There is no general entitlement to a particular answer. For personal data, Article 16 GDPR grants a right to rectification; for defamatory factual claims about the company, Section 824 of the German Civil Code comes into play. Case law on this is emerging: in 2024 the Regional Court of Kiel held a business information service liable for AI-generated false statements, and the Regional Court of Munich I ruled in 2026 on AI overviews in search results. The individual case should be assessed by a lawyer.

Why does ChatGPT describe our company so generically? Because nothing more specific emerges from the available sources, or because the distinguishing feature fell away during compression. Omission is the most common failure mode in AI answers; in the study of Google AI Overviews, 11 percent of individual claims were unsupported by the cited page, mostly through missing context.

Why does ChatGPT not know my company at all? Because too few independent sources report on you. Models reflect what exists online and is confirmed repeatedly. A well-made website alone often does not suffice if no directories, trade media or platforms pick up the details. For young or very local businesses that is the norm, not a fault.

Does it help to write the correction into the chat? No. What you state in a conversation applies to that conversation only and influences neither the model nor the answer another user receives. The widespread advice to correct the model in chat rests on a misunderstanding of how these systems work.

How do I tell whether an answer comes from training or from a web search? Usually from the citations. If the system names links, it searched live, and a correction at the source can take effect comparatively quickly. If it answers without sources, the statement comes from the training state and changes at the earliest with the next training cycle.

ChatGPT names products we no longer offer. What helps? Usually the old product page is still reachable and was only removed from the navigation. For search engines and AI systems it therefore remains a valid source. What works is either redirecting it to the successor offering or marking it explicitly as discontinued. Old PDFs in the download area and directory entries should be checked as well, because both tend to preserve outdated figures.

How important is the exact company name? More important than it appears. Language models assign name mentions to the correct entity in roughly three out of four cases; varying spellings with and without the legal form make that attribution harder still. Consistent spelling across all sources is one of the few measures that costs nothing.

Is schema markup enough to become unambiguous? It helps but does not replace running text. Structured data supplies the category in machine-readable form, and it was precisely that category information which lifted accuracy in the studied attribution task from 75.4 to 81.1 percent. What gets quoted is still the text people read.

What about false claims about me personally as managing director? Here the legal position is clearer than for company data. This concerns personal data, for which Article 16 GDPR provides a right to rectification, and the general right of personality may additionally be affected, protecting professional reputation and social standing.

Does this also apply to companies in Austria and Switzerland? Yes. AI systems answer for the whole German-speaking area largely from the same body of sources, so measurement and source work apply in the same way. What differs per country are the relevant directories, trade media and registers, which should be reviewed separately. On the legal side, Section 1330 of the Austrian Civil Code and, in Switzerland, Article 28 of the Civil Code and the Unfair Competition Act take the place of Section 824 BGB.

Which agencies specialise in correcting false AI statements? The market splits into three groups: SEO agencies carrying the topic as an add-on, tool vendors that measure but do not correct, and providers that combine measurement with source work. Which group fits depends on whether you only want to know what AI systems say about you or also want the statements corrected. Ask every provider in writing for the measurement method, the number of queries per question and an explicit statement of what is not guaranteed.

Conclusion

False and vague claims in AI answers are annoying, but they are not an attack; they are a side effect of how these systems work: they attribute, compress and fill gaps. Understanding that means you stop looking for the correction form that does not exist and start cleaning up the sources the answers are built from, writing your self-description so that even the shortest possible version still holds true and still distinguishes.

The first step is manageable: ten to fifteen questions, several systems, screenshots with dates, plus the shortening test on the first three sentences of your home page. After that you know whether you have a problem and, if so, which one. Everything else follows from there.

If that first hour turns up more entries than you can work through alongside everything else, that is the point where outside help pays for itself.

This article was published on 9 August 2026, extended on 11 September 2026 with the five most common practical cases and guidance on choosing a provider, and factually revised on 14 September 2026. On 18 September 2026 the legal position in Austria and Switzerland was added. On 23 September 2026 the article on vague representation from 16 August 2026 was merged into it. It does not constitute legal advice.

Sources

• Regulation (EU) 2016/679 (GDPR), Article 16 (right to rectification)
• Section 824 German Civil Code (credit endangerment)
• Regional Court of Kiel, judgment of 29 February 2024 (case 6 O 151/23) on liability for AI-generated false information about companies
• Regional Court of Munich I, judgment of 28 May 2026 (case 26 O 869/26) on AI overviews in search results
• Austria: Section 1330 of the Civil Code (ABGB) - Austrian Legal Information System
• Switzerland: Article 28 of the Civil Code and Article 3(1)(a) of the Unfair Competition Act - Fedlex
• Gerard Pons, Besim Bilalli, Anna Queralt (Universitat Politècnica de Catalunya, May 2025): Knowledge Graphs for Enhancing Large Language Models in Entity Disambiguation, arXiv:2505.02737. Accuracy of 75.4 percent without and 81.1 percent with category hierarchy across ten datasets
• Haofei Xu, Umar Iqbal, Jacob M. Montgomery (May 2026): Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact, arXiv:2605.14021. 55,393 search queries, 98,020 individual claims, 11.0 percent unsupported by the cited page