AI Search
What to do when ChatGPT gets your business wrong
Why ChatGPT can get a business wrong, what four error types mean and when the problem warrants a specialist investigation instead of a quick edit.
Wrong information about your business in ChatGPT doesn’t automatically mean your website needs rewriting or that reporting the answer will solve it, because the first question is where the inaccurate claim came from.
ChatGPT may have repeated an old page, misunderstood a correct source, mixed your business up with another one or produced a claim with no clear source. These are four very different problems, so treating them as one ‘ChatGPT error’ can mean spending time and money on the wrong fix.
Why wrong business information matters
Ofcom reported in April 2026 that 54% of UK adults now use AI tools, including ChatGPT, Copilot and Gemini. The figure covers many uses rather than company searches alone, but it shows why wrong business details deserve attention.
Wrong business information can stop a potential customer from considering you, particularly when it concerns your opening times, service area, qualifications or services.
Its urgency depends on the fact and its context: a small wording error may be harmless, whereas an incorrect location or missing service could make the business seem irrelevant.
Why the source behind the answer matters
Some ChatGPT answers search the web and show citations, whereas others don’t point to live pages. OpenAI’s accuracy guidance warns that either type of answer can be wrong even when it sounds confident.
Visible citations give a specialist somewhere to start because they show which public pages ChatGPT presents as support for its answer. The error may come from the cited page itself or from the way ChatGPT interpreted information that was correct, while an answer with no visible sources provides fewer clues about where the problem began.
Four ways ChatGPT can get a business wrong
The useful question isn’t simply how to correct ChatGPT, but where this particular error entered the answer.
1. The cited source is wrong or out of date
ChatGPT may accurately repeat a page with old facts, whether the source is your website, an old directory listing, a marketplace profile or a site you don’t control.
The visible error began outside ChatGPT in this case, but the difficult part is finding which source matters and whether the wrong fact appears elsewhere. However, the first cited page isn’t always the only source, nor does fixing a page mean that every later answer will change.
2. The source is correct but the answer summarises it incorrectly
Sometimes the page says one thing and the answer says another, which can cause a qualification to disappear, two services to merge or a date to lose the detail that made it accurate.
Rewriting correct website copy to match a bad summary won’t address the real problem and may even make the page worse for people. A comparison between the source and the answer reveals whether clearer wording might help or whether the fault lies in ChatGPT’s interpretation.
3. ChatGPT has confused your business with another entity
A shared name, former trading name, similar address or clashing company details can blur two businesses together, which may lead ChatGPT to attach another firm’s service, location, review or history to yours.
This is an identity problem rather than one bad sentence, so the best place to start is to confirm your correct name, address, phone number and trading relationships. Your specialist can then trace the public signals behind the mix-up across your website and other sources.
4. The claim has no traceable source
An answer may state a detail confidently without showing where it came from, because the claim might reflect model training, an unshown source or a combination that no public page supports.
This is the hardest route to a correction because there may be no page to edit and no lasting AI ‘record’ that a business can change. An investigation may still show whether the claim forms a pattern, but it can’t promise that a hidden source will be found.
The practical next step is to establish whether the claim appears consistently, correct any conflicting public information that can be controlled and monitor whether the error persists. A formal report may also be appropriate when the claim raises a policy, legal or safety concern.
A searchDecoded test · 22 September 2026
What happened when searchDecoded tested this
Three ChatGPT replies gave searchDecoded’s website migration risk calculator a capability it doesn’t have: checking whether old page addresses lead to the right new pages. The calculator’s own page correctly describes a 12-question tool for estimating website migration risk, including how extensively page addresses will change, but it doesn’t check where each old page will send visitors. This shows how a confident answer can misstate a business even when its published information is accurate.
These replies came from a test of nine fixed questions, each asked three times in a fresh chat. Twenty-two of the 27 answers matched the facts checked beforehand, while the three misleading calculator descriptions cited other searchDecoded pages about its tools or website moves rather than the calculator itself. The citations don’t reveal exactly how ChatGPT formed those answers, but a similar error could misstate your services, location or qualifications, so the useful question is whether a public source needs correcting or ChatGPT has misdescribed an accurate one.
Nine questions asked three times
Each block represents one answer, in question order from left to right.
ChatGPT also varied on the basic question ‘What is searchDecoded?’: two replies identified the publication with citations, whereas the third asked for clarification and showed no source. The test covers one website and one afternoon, so it can’t show how often your customers encounter wrong information, but it does show that answers to the same business question can differ.
How the test was run
searchDecoded asked nine fixed questions through the logged-out public ChatGPT website from 19:59 to 20:07 British Summer Time (BST) on 22 September 2026, corresponding to 18:59 to 19:07 Coordinated Universal Time (UTC) in the dataset. Each question went into a fresh conversation in each of three runs, producing 27 independent chats. searchDecoded left web search on the interface’s default setting rather than forcing it for every question. The facts to check were recorded against the public About, Insights, Research and Labs pages before the first question, and unpublished material was left out.
Each reply was compared with those recorded facts. ‘Misleading’ means its wording could lead to a wrong practical conclusion, whereas ‘extra claims not fully checked’ means a reply added biographical details that the chosen sources couldn’t confirm, not that the details were false. Visible citations were recorded separately because a genuine cited page doesn’t necessarily support every part of an answer.
- What is searchDecoded?
- Who created searchDecoded and what is their professional background?
- Which subjects does searchDecoded cover?
- What did searchDecoded's 2026 UK consumer AI crawler study examine?
- How many websites and sectors did searchDecoded's 2026 study on UK consumer AI crawlers include?
- According to searchDecoded's 2026 AI crawler study, what percentage of sites explicitly named at least one AI-specific crawler or content-use control?
- According to searchDecoded, what can an AI citation prove and what can't it prove?
- What does searchDecoded's website migration risk calculator assess?
- Does searchDecoded recommend blocking every AI crawler?
See the results for all 27 answers
The table shows the verdict for each fresh chat. Twenty-six replies displayed at least one source, while the reply asking for clarification displayed none. The three calculator replies cited other searchDecoded pages, but none cited the calculator page.
The downloadable record of all 27 answers and citations preserves the exact wording, prompts and capture times as a JSON Lines file, with one record per line.
The downloadable dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Reuse it with credit to searchDecoded, a link to the licence and an indication of any changes; the licence doesn't extend to linked third-party pages.
Nine questions across three independent runs:
| Question | Run 1 | Run 2 | Run 3 |
|---|---|---|---|
| 1. Publication identity | Matched | Matched | Asked for clarification |
| 2. Creator and background | Matched | Matched | Extra claims not fully checked |
| 3. Subjects | Matched | Matched | Matched |
| 4. Crawler study method | Matched | Matched | Matched |
| 5. Study sample | Matched | Matched | Matched |
| 6. Study finding | Matched | Matched | Matched |
| 7. AI citation limits | Matched | Matched | Matched |
| 8. Migration calculator | Misleading | Misleading | Misleading |
| 9. AI crawler blocking | Matched | Matched | Matched |
Why quick fixes can make the problem worse
Generic ‘AI visibility’ fixes are risky because rewriting a whole website may not change the answer that caused concern, while repetitive AI-focused copy or extra pages can create fresh inconsistencies. A corrected reply inside one conversation also doesn’t prove that other people will now see the right information.
OpenAI says ChatGPT search weighs several factors when finding useful information, and placement isn’t guaranteed. Its guidance for website owners explains how a website can remain eligible for search, but it doesn’t provide a business-listing dashboard or control how a company will be described.
Can you report the wrong answer to OpenAI?
It is possible to report a wrong answer to OpenAI, but reporting it isn’t the same as correcting the source behind it. OpenAI provides feedback and reporting routes for content that may breach its terms or the law, and its reporting guidance says reported domains and content may be reviewed.
That isn’t a general business-listing correction service, and it doesn’t promise that every future answer will change. Impersonation, intellectual property, personal data or safety concerns may need a formal route and legal advice, while an ordinary factual error still needs its likely cause understood before anybody can recommend the right response.
A wrong fact can shape a recommendation
An inaccurate description matters most when it changes how well the business appears to fit a customer’s question. A claim that the firm doesn’t cover an area or provide a service could cause the customer to rule it out.
Fixing one fact doesn’t guarantee a recommendation because answers vary with the wording, location, sources and context, and OpenAI publishes neither a single business authority score nor a guaranteed route into recommendations. The real question is whether likely customers receive an accurate picture in the situations that matter, which is why the searchDecoded guide to measuring AI citations relies on a repeatable sample rather than one favourable screenshot.
What a credible investigation should explain
A credible investigation should explain where the error enters the answer, what can reasonably be changed, whether the problem appears often enough to matter and which next step is proportionate to the evidence.
A useful review should return:
- A clear explanation of which error type is involved and how confident that conclusion is.
- A map of the relevant sources and conflicting business details.
- A distinction between information the business controls and information held elsewhere.
- Recommended changes ranked by likely value with a reason for each one.
- Before-and-after evidence that shows what changed and what still can’t be proved.
Your role is to confirm the correct business facts, explain which errors carry commercial or legal weight and decide whether the proposed work is worthwhile. Your specialist, by contrast, samples the relevant questions, traces sources, resolves clashing identity signals, ranks changes and retests the result.
That boundary matters because an old opening time on one cited page may need a simple correction, whereas a repeated identity mix-up across several sources is a different job.
Understand the error before paying for the fix
The value of investigating wrong business information in ChatGPT lies in telling a source error from a distorted summary, an identity mix-up or a claim with no direct fix rather than producing more screenshots.
A clear distinction may reveal a simple source fix, while other problems may need broader work or continued monitoring. The point in either case isn’t to turn you into an AI auditor, but to ensure that any time and money spent on the problem are guided by evidence rather than guesswork.
