← Blog AI visibility GEO Data Trustworthiness SMEs
When AI says something incorrect about your business — and how to spot it
2026-08-27 · Martin Nymann · 8 min reading
AI can cite your website and still say something incorrect about you in the same reply: research shows that only around half of the sentences in AI replies are fully supported by the reply’s own sources. And the errors tend to go in one particular direction — when the database and the website contradict each other, the AI resolves the conflict against the company. Here are the three typical types of error, how to spot them, and why the solution is to correct the sources, never to complain about the response.
Key takeaways
- Citations do not equate to credibility: the research review (arXiv:2607.14035) finds that only ~51.5 per cent of the statements in AI responses are fully supported by the response’s own sources, and ~74.5 per cent of the citations support the claim to which they are attached.
- AI responses are constructed from isolated snippets from multiple sources — errors occur when the snippets do not agree on the same fact.
- Based on our own customers’ experiences: when the CVR register and the website contradict each other, the AI typically selects the version in the register or makes a reservation — the conflict is resolved against the company.
- The three typical categories of errors: incorrect town/place of origin, out-of-date benefits and mixed name variants — all three are resolved by ensuring the sources are consistent.
- You can’t simply dismiss an AI response — but you can correct the sources on which the next response is based. That’s the only approach that works in the long run.
AI can cite your website as a source — and yet say something incorrect about your business in the very same reply. This is not a rare occurrence: a research review of generative search engines found that only around half of the sentences in AI responses are fully supported by the sources to which the response itself refers. And in our own measurements with real customers, the error tends to occur in one specific scenario: when the official register and the website contradict each other, the AI resolves the conflict against the business. Here’s why this happens, how to spot it — and why the solution is never to write to OpenAI, but to correct your own sources.
How often does AI say something that the sources do not cover?
It has been measured. A critical research review of all GEO research from 2023 to 2026 (arXiv:2607.14035) compiles the figures: in the studies of citation fidelity, only around 51.5 per cent of the sentences in AI responses were fully supported by their sources, and around 74.5 per cent of the citations actually supported the claim to which they were attached. The rest are inaccuracies: the model tightens up a formulation, mixes two sources together or fills a gap with something that sounds right.
The conclusion of this research is worth committing to memory: a citation does not equate to credibility. The fact that AI refers to your page does not mean that everything in the response is correct — it merely means that your page was retrieved and read. That is also why we make a clear distinction between being cited and being recommended in our own analysis of 576 AI responses: these are two different outcomes, each with its own mechanism.
Why do the errors affect your business in particular?
An AI response with sources is built from isolated excerpts. The search engine retrieves a handful of passages from several pages — your website, a catalogue, the official register, an old news article — and the model must then merge them into a single response. The problem arises when the snippets do not agree.
We’ve seen this pattern among our own customers, and it’s consistent: if there are discrepancies between the CVR register and the website — name, sector, town — the AI typically chooses the register’s version or adds a disclaimer, which makes the company appear unreliable. The error feels unreasonable, but the mechanics are logical enough: the register is an authoritative source, and when two sources about the same company contradict each other, the ‘safe’ solution for a model is to rely on the register or hedge its position. Each of your pages may look correct when read on its own — it is the discrepancy that triggers the error.
The three typical error categories — and what to do about them
| The error in the AI response | The typical cause | What to do |
|---|---|---|
| Incorrect city or registered office (“the company is based in Aalborg” — you’re based in Copenhagen) | The registered address is an old address, a holding company or the accountant’s office | Correct the registered address (via Virk), or make the correct address unambiguous on the website — and don’t let the two contradict each other |
| Outdated services (AI mentions something you stopped offering two years ago) | Old subpages, directories and profiles that were never updated | Update or remove the old pages, and correct the external profiles — AI reads them all |
| Incorrect or mixed-up name (the registered name and brand name are confused) | Legal name and brand name appear randomly and interchangeably across sources | Declare both in your structured data (name + alternateName + sameAs) so that machines can resolve the variants to the same company |
The common thread is the same as in the rest of AI visibility: consistency across sources is the signal itself. AI engines cross-check, and consistency between your website, the register, the directories and your profiles is what makes you a source on which an answer can be based without reservation.
How do you spot when AI gets something wrong?
The manual method is straightforward, but needs to be done regularly: ask yourself the questions your customers ask — in ChatGPT, Perplexity and Gemini — and read the answers with one question in mind: is there a claim here that my own verified facts contradict? Remember that the same prompt yields different answers from day to day, so a single check is a snapshot, not a definitive list.
Geoa does the same thing systematically: every night, the platform compares the AI responses that mention your business with your verified registered facts — and specific contradictions become a card in your workflow, featuring the verbatim quote from the AI response and the fact it contradicts. There are two important things to be honest about regarding this mechanism: the assessment is itself carried out by an AI, so the card is an invitation to verify, never an automatic correction — and it only flags discrepancies with facts that we have actually verified. The fact that the AI omits something about you is not an error that can be measured in this way.
What should you not do?
You cannot have an AI response removed by lodging a complaint. Neither OpenAI, Google nor Perplexity correct individual responses about small businesses, and even if they did, the next response would be built from scratch using the same sources. Chasing after individual responses is a waste of time.
What works is correcting the sources from which the answers are built: your website, the register, the directories and your profiles — so that they all tell the same story. It is also the only approach that stands the test of time: a corrected answer disappears with the next question; a corrected source is read every time.
Frequently asked questions
Can I get ChatGPT to correct an incorrect answer about my business?
Not directly. AI providers do not correct individual answers about small businesses, and the next answer will be generated from scratch using the same sources anyway. The effective approach is to correct the sources — your website, the register, the directories and your profiles — so that they all tell the same story. Then the answers will correct themselves over time.
How often does AI get it wrong when citing sources?
The research review on generative search engines (arXiv:2607.14035) summarises the findings: only around 51.5 per cent of the sentences in AI responses were fully supported by the response’s own sources, and around 74.5 per cent of the citations supported the claim to which they were attached. A citation is therefore not the same as credibility.
Why does the AI choose the version from the register over the one on my website?
Because the register is an authoritative source, and models resolve conflicts by relying on the most reliable source or by expressing reservations. We have measured this among our own customers: discrepancies between the register and a website are typically resolved against the company. The solution is not to override the register, but to bring the two into alignment — correct the register via Virk, or correct the website.
What if it is the register that is incorrect?
Then correct the register. CVR details such as address and sector code can be updated via Virk, and the correction will be reflected in the sources that rely on the register. Until the correction is processed, it helps to make the correct information unambiguous on your own website — but never allow the two to contradict each other for longer than necessary.
Does Geoa detect it itself when AI says something incorrect about me?
Yes — every night, the AI responses that mention your business are compared with your verified register details, and any specific discrepancies are turned into a card in your work queue containing the exact quote. The assessment is generated by the AI itself, so the card is an invitation to verify and correct the source — never an automatic correction, and never a guarantee that everything will be caught.
Isn’t it just a matter of time before the models improve?
Perhaps the accuracy will improve — but the mechanics remain the same: answers are built from your sources, and contradictions between them will always be a problem that the model must resolve on your behalf. Consistency across websites, registers and profiles is therefore not a temporary workaround; it is the foundation, no matter how good the models become.
Get your free GEO Score — 60 seconds, no credit card required.
Get started for free