AI hallucination: how wrong AI answers hurt your brand and what to do about it

What is AI hallucination?

An AI hallucination is output from a language model that is phrased convincingly but is not backed by correct information. The model sounds sure of itself while it is wrong.

The most cited review of the topic, the survey of hallucination in natural language generation by Ji et al. (2023) in ACM Computing Surveys, distinguishes two basic types. An intrinsic hallucination contradicts the source material the model has in front of it. An extrinsic hallucination can be neither confirmed nor refuted from that material: the model adds something that is not written anywhere.

Both types matter for businesses. If an AI assistant reads your services page and then quotes a different price, that is intrinsic. If it credits you with a certification that appears on none of your pages, that is extrinsic.

The problem will not disappear completely. Language models calculate the most likely next word, and the most likely word is not always the correct one. Better models and a web search running in the background reduce errors, but as long as answers are built on probabilities, some risk remains. That is also the research consensus Ji et al. (2023) reflect: hallucination cannot be fully eliminated from the output of large language models.

Which kinds of AI hallucination affect businesses?

AI hallucinations about businesses follow recurring patterns, and most of them trace back to missing or conflicting information on the web. The table below lists the most common error types.

The common thread: hallucinations appear where the model finds no clear, well-supported answer. Closing those gaps lowers the chance of a wrong description.

Why is AI hallucination becoming riskier for brands?

The brand risk from AI hallucination is rising because more people read answers directly in an AI interface and click through to sources less often. An error reaches more people, and fewer of them notice it.

Two data sources point the same way. The comparative analysis by Iyappan (2026) in the GOYBO International Journal of Marketing Intelligence compares traditional search with AI-driven systems across a corpus of 162 units of analysis. According to its Table 5, source verification falls from 44% to 27%, while the share of conversational queries rises from 29% to 91%. Almost every interaction then becomes a question followed by an AI-synthesised answer, rather than a keyword query followed by a list of links.

A Pew Research Center analysis based on real browsing data confirms the direction. In the March 2025 browsing data of 900 US adults, users who saw an AI summary on a Google search clicked a traditional result in only 8% of visits, against 15% when no summary appeared. They clicked a link inside the summary in 1% of visits.

For your brand this means: if an AI answer misrepresents what you offer, the reader is unlikely to visit your website and correct the error themselves. On Iyappan's figures, in 73% of AI-driven search cases the user does not verify the answer afterwards. His data presents this as the norm already, and it owes less to careless users than to an interface that delivers answers with an expert's confidence. Fogg et al. (2002), in their research on how people judge web credibility, found that perceived source credibility strongly shapes whether people adopt information. In AI search, the AI system itself carries that credibility signal: users attribute accuracy to the interface instead of judging individual sources, so a hallucinated statement about your business inherits the trust users place in that interface. More on this behaviour in our analysis of zero-click search.

Do source citations protect against AI hallucination?

Citations in AI answers offer limited protection against hallucination, because a cited source does not prove the model actually used it. A link under an answer looks like evidence, but it is not evidence by default.

Wallat, Heuss, de Rijke and Anand examined this in their paper "Correctness is not Faithfulness in RAG Attributions" (ICTIR 2025). Their finding: up to 57% of the citations in their experiments were not faithful. The model had built the answer from other patterns and attached a matching source afterwards. The authors call this post-rationalisation.

Two consequences follow. Your website can appear as a source while the statement next to it does not come from you. And being mentioned somewhere is not enough: the facts on your page have to be so clear and easy to lift that the model really uses them.

Why is factual accuracy a visibility signal?

Factual accuracy works as a visibility signal in AI search because generative systems favour content whose statements match other sources and are clearly supported. In classic rankings, accuracy was mostly a matter of decency rather than a ranking factor: a heavily linked but wrong page could easily beat a correct, less linked one.

In his correlation table, Iyappan (2026) rates the relationship between factual accuracy and AI trust signal rating as positive and very strong, the highest level on his four-point scale. He names the break with the old paradigm explicitly: "Generative AI systems incorporate factual consistency checking and source credibility evaluation into retrieval scoring, creating structural incentives for epistemically rigorous content production." In GEO, accuracy therefore becomes a ranking factor. The logic is easy to follow: a statement that appears identically in several sources is easier for a system to adopt than one that turns up once with no support.

In practice, precision helps twice. Concrete data, supported figures and named people close the gaps in which hallucinations form, and they make your content easier to cite.

How do entity signals reduce hallucination risk?

Entity signals reduce the risk of AI hallucination because they give a model clear, machine-readable and consistent facts about your business. The less a system has to guess, the less it invents.

Our own research into the AI visibility of SMEs across 310 websites shows how large the gaps are in practice: 52% had no Organization or LocalBusiness schema, 89% no machine-readable address and 41% no machine-readable date. On 20% of the sites (60 of 296), GPTBot had no or insufficient access. Those are exactly the gaps a model fills with guesses when in doubt.

The main building blocks:

• Organization schema on the homepage. Name, description, address, contact details, logo and sameAs links to your profiles give every system a fixed anchor. In its guidance on AI features in Search, Google stresses that structured data must match the visible text and that AI Overviews need no special markup. Schema does not replace clear content; it confirms it.
• Consistent NAP data. Name, address and phone number should be identical on your website, your Google Business Profile and in industry directories. Conflicting data forces the model to choose.
• A maintained Google Business Profile. Keeping opening hours, services and category current there stops outdated details from circulating in Google's interfaces.
• Editorial mentions. Accurate articles in trade media are independent confirmations of your facts. The more sources agree, the more stable the picture.
• Wikidata where it fits. A correct item provides a source that is separate from your own website, provided your business meets the notability criteria.

How to build that kind of consistent brand identity step by step is covered in our article on brand entity as a GEO signal.

What should you do when you find an AI hallucination about your business?

When you find an AI hallucination, you correct the sources the system draws from, not the answer itself. Most AI assistants have no reporting process comparable to the one for Google search results.

1. Record the error. Ask ChatGPT, Perplexity, Gemini and Copilot the same questions about your business, your services and your region. Note the date, platform, question and answer. Tools such as Otterly.ai or Peec AI can automate this. Why one platform is not enough is explained in our piece on AI search monitoring.
2. Find the source. Perplexity shows its sources openly. Check whether the wrong information comes from an old page, a directory or someone else's article.
3. Fix your own pages. Update outdated content, redirect discontinued services cleanly and phrase key facts so each fits in one sentence and makes sense without context.
4. Clean up external listings. Correct directory entries, your Google Business Profile and, if you have one, your Wikidata item.
5. Validate your schema. Wrongly declared structured data, such as an old address or the wrong business type, reinforces errors. Check your markup regularly with the Rich Results Test and the Schema Markup Validator.
6. Secure independent confirmation. A recent, factually correct article in a relevant trade publication gives systems another source that can, over time, outweigh older and incorrect data.

Do not expect instant results. How fast a correction takes effect depends on whether a system searches the live web or falls back on older training knowledge.

Frequently asked questions

What is AI hallucination in plain language?

AI hallucination is a false statement that an AI system phrases as if it were true. It happens because language models produce likely sequences of words rather than looking things up in a database of facts. Ji et al. distinguish statements that contradict the source material from statements that cannot be derived from it at all.

How often do AI systems hallucinate about businesses?

There is no general error rate for business descriptions. How often it happens depends heavily on how much clear and consistent information about a business exists online. Small businesses with few mentions and no structured data are more vulnerable than well-known brands, because a model finds less evidence about them.

Can I control what ChatGPT says about my business?

You cannot control it directly, but you can influence it. ChatGPT and other assistants rely on content from the web. Clear facts on your website, correct schema, consistent NAP data and independent mentions change the material the answers are built from.

Are AI hallucinations becoming rarer?

That cannot be said in general: error rates vary widely by model and task. What is clear is that the business risk does not shrink by itself. More people read AI answers and fewer check the sources (Iyappan 2026 reports source verification falling from 44% to 27%), so even a rare error can reach many people.

Does schema markup prevent AI hallucinations?

Schema markup does not prevent hallucinations, but it narrows the room for them. It supplies machine-readable facts such as name, address and services that a system no longer has to guess. The condition is that the markup is correct and matches the visible text; wrong schema can even reinforce errors.

How long does it take for a correction to show up in AI answers?

It varies by system and cannot be guaranteed. Assistants with live web search, such as Perplexity, can use an updated page after the next crawl. Answers that rely on training knowledge often change only with a new model version.

How does factual accuracy affect visibility in AI search?

Positively and clearly. In Iyappan's (2026) correlation analysis, factual accuracy has a very strong positive correlation with AI trust signal rating, the highest level in the study. Accurate, supported statements protect against hallucination and make your content easier to cite at the same time.

Is a citation in an AI answer proof that the answer is right?

No. A citation only shows that a system links a page, not that the statement comes from it. Wallat et al. found in their experiments that up to 57% of citations were not faithful to what the model had actually relied on. For important decisions, open the linked page yourself.

Want to know what AI systems are saying about your business right now? Request a free analysis.