A customer asks ChatGPT about your business, and the answer comes back wrong: it says you've closed, mixes you up with a firm that shares your name, quotes services you don't offer, or sends people to an address you left years ago. Unlike a bad review, there's no reply button and no one to call. But wrong answers have causes, the causes are usually findable, and most of them are fixable. This page explains which are which, honestly.
The five ways AI gets a business wrong
- Mistaken identity. You share a name with another firm, and the machine merges you: their reviews, their prices, their controversy, presented as yours.
- Stale facts. Closed, moved, renamed, rebranded: the machine is reading a page from your past, or repeating something it learned before the change.
- Coverage errors. Recommended for a city you don't serve, or written off in one you do. Visibility is patchy by place, and so are the mistakes.
- Borrowed history. An old incident, a previous owner's record, or a defunct branch's reviews attached to today's business.
- Confident invention. When the retrievable evidence is thin, the machine fills gaps fluently. The less you publish, the more it guesses.
Where wrong answers come from: the two layers
Every AI answer is built from two layers, and the fix depends on which one is wrong.
- The retrieval layer: what the machine reads at answer time. Live web pages, directories, registers, review pages. If the wrong claim is coming from here, it can be traced to specific sources, and corrected evidence can change the answer in days to weeks.
- The model's memory: what it learned in training. Some claims are baked in and resurface even with no source on screen. No one can edit this layer directly, us included, and anyone who tells you otherwise is selling something. What works is containment: making the live evidence strong, consistent and machine-readable enough that retrieval overrides memory, and knowing that full correction may wait for the next model release.
What fixing it actually involves
- Reproduce and date it. The same question, asked repeatedly, because answers vary between identical runs. A one-off wobble and a repeating error are different problems, and knowing which you have is step one.
- Trace the sources. Which pages the machine retrieves when it builds the wrong answer, and which of them carries the wrong claim.
- Correct the evidence layer. Your own pages first, then the third-party pages the machines in your trade actually read. Machines don't take corrections; they take evidence.
- Re-measure until it holds. One good answer proves nothing, for the same volatility reason. The job is finished when the corrected answer repeats, and you get the dated before-and-after record showing it.
The honest caveats
Two things this service can't do. If the wrong answer lives in the model's memory rather than its reading, containment is real but instant correction is not, and we'll say so after reproducing it rather than after invoicing you. And if the answer is unflattering but true, that's a reputation problem, not a machine error; no evidence-layer work should be sold for it, and we won't.
If this is happening to you now
Email felix@aeovapartners.com with the question you asked and what the machine said, a screenshot is perfect. We reproduce it the same day, tell you which layer it's coming from, whether it's fixable and how fast, and quote a fixed fee before any work begins. If it's a one-off wobble that doesn't repeat, we tell you that too, and it costs you nothing.
Common questions
ChatGPT says my business is closed. How do I fix it?
First reproduce it: ask the same question several times, because AI answers vary between identical runs and a one-off error needs different handling from a repeating one. If it repeats, the claim is coming from somewhere, usually a stale page, an old directory listing, or a machine-readable record you've forgotten exists. The fix is finding that source and correcting the evidence the machine reads, then re-measuring until the corrected answer holds. That trace-and-correct work is exactly what our correction service does.
Can I contact OpenAI or Google to correct a wrong answer about my business?
There is no correction channel a business can rely on today. The chat interfaces have feedback buttons, and Google has processes for its business profiles, but none of them guarantees a changed answer. The route that actually works is indirect: fix the sources the machine retrieves when it builds the answer. Machines don't take corrections; they take evidence.
How long does fixing a wrong AI answer take?
It depends on which layer the error lives in. If the machine is retrieving a wrong or stale page at answer time, corrections can land within days to weeks of the source being fixed. If the error is baked into the model's trained memory, no one can edit it directly: the practical fix is making the live evidence strong enough to override it, and full correction may wait for the next model release. We tell you which case yours is after reproducing it, before any work is agreed.
Is this the same as AEO?
It's the urgent end of it. Answer engine optimisation is usually about being recommended more; this page is about being described correctly, which comes first. The method is the same, measure, fix the evidence layer, re-measure and prove it, applied to a specific wrong answer rather than to overall visibility.
Updated August 2026