When AI says false things about your company: causes and countermeasures

Why AI claims false things at all

Language models like ChatGPT, Gemini or Perplexity know nothing in the human sense. They calculate which word is statistically most likely to follow next. When robust data about your company is missing, no honest I don't know arises, but a plausible-sounding invention. Experts call this hallucination. The model fills a gap with what would be usual for similar companies and presents the guess with the same confidence as an established fact.

On top of that comes the time factor. Many models were trained on data that is months or years old. A relocation, a new managing director, a changed range of services or new opening hours are simply not contained in the training material. The AI then answers correctly from its perspective, but wrongly from yours. A tax advisor who has merged his firm, a trades business with a new owner or a clinic with a new department experience this regularly.

The third reason is contradictory sources. When three different phone numbers, two addresses and inconsistent company names circulate on the web, the model has to guess which version is correct. It often chooses the most frequent or the most confidently phrased, not the most current. This is exactly where the most embarrassing errors arise, because to outsiders they seem entirely credible.

The most common types of false statements

Not every error is equally dangerous. Confusion with a company of the same name is particularly tricky, because the answer sounds coherent in itself. An engineering office shares its name with an online shop, and suddenly the AI blends products, reviews and locations of both into one phantom. The user notices nothing, because the answer is fluent and detailed.

Alongside this there are the classic factual errors: wrong prices, outdated certificates, invented awards or services you don't even offer. The opposite also occurs. The AI omits one of your most important services because it is nowhere clearly documented. For a service provider this is doubly harmful, because potential customers don't even consider them for the actual core competency.

The heaviest are reputation-damaging statements. An AI that derives a general quality defect from an old, long-resolved complaint, or that adopts an insolvency rumor from the wrong context, causes real damage. Such cases are rare, but they require fast, documented action.

How to detect false statements in the first place

The first step is banal and yet rarely done: ask the systems yourself. Pose in ChatGPT, Gemini, Perplexity and Copilot the same questions a customer would ask. Who are you, what does it cost, where are you located, are you certified. Note the answers verbatim. Repeat this with slightly varied phrasings, because models answer differently depending on the question.

What matters is systematics instead of chance. Set up a fixed list of questions and check them at regular intervals, say monthly. This way you recognize whether an error is new or persistent, and whether a correction has worked. Document date, model and answer. Without this measurement you fly blind and notice deteriorations only when a customer raises them.

Also pay attention to the sources that some systems provide. Perplexity and Copilot often show what they rely on. If an outdated industry directory or a foreign profile appears there, you have found the source of the error directly. That saves you a lot of guesswork in the later correction.

Tackling the causes at the root

The most effective countermeasure is unspectacular: consistent facts everywhere machines read them. Company name, address, phone number and core services must be identical on your website, in industry directories, in map services and in social media profiles. Every deviation is a gateway for confusion. This basic hygiene sounds boring but decides eighty percent of cases.

In addition, structured markup helps. With machine-readable data, for example the schema standard, you tell search engines and indirectly the AI systems unmistakably who you are. A clear about-us page, an unambiguous overview of services and a well-maintained contact page work more strongly than any advertising message, because models favor facts they can assign unambiguously.

Take targeted care of the sources that are frequently cited. Wikipedia, large industry portals, official registers and established review platforms carry a high weight. A false entry there weighs more heavily than ten correct ones on small sites. Correct the sources with the largest reach first, then work your way down.

What to do in case of acute reputation damage

When an AI claims something objectively false and harmful, document it first with a screenshot, date, model and the exact question. This preservation of evidence is the basis for every further step and is often forgotten, because answers change again at the next call. Without proof you later stand empty-handed.

The major providers have reporting channels for faulty outputs. Use them factually and with evidence instead of arguing. State what is false, what is correct and where the correct information comes from. In parallel you correct the underlying source on the web, because a report alone rarely changes the model's behavior permanently.

In case of serious, persistent reputation damage with economic loss, legal advice is sensible. In Germany personality and competition law apply even to automated statements. This is the last step, not the first, but you should know that it exists and that your documentation is what makes it possible in the first place.

Realistic expectations for the correction

Be honest with yourself: you cannot edit AI outputs directly like a database entry. You influence them indirectly, via the information landscape on the web. That means patience. Between a correction at the source and the visible change in the models, weeks to months can pass, depending on when and how the system refreshes its data.

Some systems with live access to the web react quickly, because they read current pages in real time. Pure training models without web search take longer, often until the next training cycle. That is why it is worth observing both types separately. A piece of good news is at the same time a warning: what you publish today co-shapes tomorrow's AI answers.

Don't expect one hundred percent control. The goal is not perfect steering, but a factual landscape so clear and consistent that the most probable answer also becomes the correct one. Whoever achieves this reduces false statements significantly, without ever being able to rule them out completely.

A practical roadmap

Don't start with tools, but with taking stock. Collect your correct core facts in one place, then check the most important AI systems and the most frequent sources on the web. Only once you know where the truth deviates from the public portrayal can you act in a targeted way instead of optimizing blindly.

After that you work in waves. First make your own website and the large directories consistent, then correct the high-reach external sources, lastly the smaller ones. Measure again after each wave, so that you see what works. This loop of measuring, correcting and measuring again is the actual core of the work, not a one-time project.

Anchor the check as a routine. A fixed appointment once a month, at which you ask the same questions and compare the answers with the previous month, is enough for most small and mid-sized companies. This way you stay able to act before a small error becomes a problem that costs you customers.

• Record core facts centrally and in writing
• Test four to five AI systems with customer questions
• Make website and large directories consistent
• Correct high-reach external sources first
• Measure monthly and compare with the previous month

Industry differences: where false statements hurt most

Not every industry carries the same risk. In the health, finance and legal fields, false AI statements can cause immediate damage: when a language model attributes to your practice a service you don't even offer, or ascribes to your firm a specialty you never had, liability questions and lost trust arise. Here close-meshed monitoring and fast correction are worthwhile, because individual false statements become expensive.

In retail and gastronomy, by contrast, operational details are more in the foreground: wrong opening hours, outdated prices, a closed location still counted as open. Such errors cost you customers directly, but are usually easier to fix, because they hang on clearly structured data sources like industry directories. B2B service providers in turn suffer above all from wrongly assigned references or invented case studies. So consider which kind of false statement weighs most in your industry, and gear your monitoring exactly to that.

A worked example: what a false statement really costs

Take a mid-sized trades business that is wrongly portrayed by AI assistants and search systems as only regionally active, although it delivers nationwide. Assume that per month 400 potential customers make an AI-supported inquiry about this service area. If the business loses only 5 percent of them through the false statement, that is 20 fewer inquiries. At a realistic close rate of 10 percent and an average deal value of 2,500 euros, around 5,000 euros of revenue are missing every month.

Calculated over the year, that is 60,000 euros lost gradually, without anyone noticing the connection. Against this stands the effort of cleaning up the data sources and initiating the correction: often a few working days plus ongoing monitoring. This calculation shows you two things. First: even small false statements add up, because they work permanently. Second: the investment in clean data usually pays for itself quickly. Plug in the figures of your own business, then you see how high your priority for the topic should be.

Frequent questions and misconceptions

A widespread error goes: if I simply send the AI the correct information once, the mistake is fixed. That is not how it works. Language models don't learn in real time from individual messages, and a chat history changes nothing about the underlying training data or live sources. Sustainable correction only arises when you change the sources the systems draw from: your website, directories, Wikipedia-adjacent databases, press articles.

A second misconception: the louder I object, the faster the false statement disappears. In fact neither public outrage nor mass reporting helps. What counts is consistency across many credible sources, so that the correct version becomes the clear majority. And finally, many believe the problem concerns only big brands. The opposite is often true: precisely with smaller companies that have a thin data landscape, the AI fills gaps with guesses, and exactly from this the most frequent errors arise. Whoever publishes little about themselves gives the systems more room to guess.