Ask an AI engine about your own business and you may not like the answer. A price list from two years ago. A service you discontinued. The wrong city. A confident description of a company that is half you and half a business with a similar name three states away.
The unsettling part is the delivery. There is no hedging, no "this may be outdated." The engine states the wrong thing the way it states everything, plainly, as fact, to a buyer who has no reason to doubt it.
You cannot email support about it. You can fix it. The process is unglamorous and it works.
Why engines get businesses wrong
Three failure modes cover almost every case. Training data is a snapshot, so a model can carry a frozen version of your business from whenever its data was collected. Retrieval reads the live web, so if stale directories, unclaimed profiles, and forgotten pages outnumber the correct record, the wrong version wins the vote. And when the entity record is thin, the model fills gaps by inference, which is where the strange blended answers come from.
We covered the mechanics of those two clocks in our retrieval versus training data post. The practical takeaway here: you fix retrieval by fixing sources, and you wait out training.
Step one: find out what they actually say
Before fixing anything, document the damage. Ask ChatGPT, Perplexity, Gemini, and Google's AI Overviews the questions your buyers ask: who you are, what you do, what you cost, who else they should consider. Log the answers, note what is wrong in each, and note which sources get cited when sources are shown.
The cited sources are the map. They tell you exactly which pages are feeding the error.
Step two: fix the surfaces you own
Your website, your About page, your schema, your Google Business Profile. Make the correct facts unambiguous, current, and easy to lift: what the business does, where it operates, what it costs, what it no longer does. If a service is dead, remove the page or state plainly that it has been discontinued, because a live page describing a dead service is you outvoting yourself.
This is also where structured data earns its keep. Organization schema with a sameAs list tying your profiles together tells machines which scattered records describe the same entity, which is half the battle in any name collision.
Step three: fix the copies
Then the third-party record: aggregators, directories, old listings, profiles created by platforms you never signed up for. Claim what can be claimed, correct what can be corrected, and prioritize whatever the engines actually cited in step one. Consistency across those surfaces is the whole game, and we wrote about why in our entity consistency post. The web is a rumor mill, and machines average the rumors.
Step four: wait on two clocks
Once the sources agree, retrieval-based answers typically come around within weeks as pages get recrawled. Training-based answers lag until the next model cycle. Recheck monthly with the same question set you logged in step one, so you are measuring drift instead of guessing.
What does not work: arguing with the chatbot, spamming feedback buttons, or publishing a blog post titled "actually we are not closed." Machines do not read rebuttals. They read records.
The record is the reputation now
A wrong answer engine is not an annoyance, it is a sales conversation you lost without knowing it happened. Auditing what the engines say, fixing the owned surfaces, and cleaning the third-party record are core work areas in our SEO / AEO / GEO Foundation, from $7,500. For a business with an active misinformation problem, it is the most valuable work we do, because every other marketing dollar is spent on top of what the machines believe.