At some point every founder has the same experience. Someone tells you "I asked ChatGPT about your product," and what it said was wrong. Old pricing. The wrong audience. A feature you never built, or one you killed two years ago. Sometimes it recommends a competitor while describing your product.
The first instinct is to treat it as a one-off glitch. It is not. An AI answer that is wrong about you is a symptom with a specific cause, and the cause is almost always findable and fixable. Here is the playbook I use.
Step one: document the wrong claim precisely. Screenshot it, but also write down the exact question that produced it, which assistant produced it, and the date. "ChatGPT said our starter plan costs X" is actionable. "AI says bad things about us" is not. Ask the same question to the other major assistants too. If three get it right and one gets it wrong, you are dealing with a single-model problem. If all four get it wrong, the source they share is broken, and that is far more common.
Step two: find the source. Wrong AI claims about brands are rarely invented from nothing. The model read it somewhere, or learned it from somewhere. Ask the assistant directly: "where did you get that?" When browsing is on, many will show you the page. When it is not, search for the exact wrong claim in quotes. Old pricing usually traces to a directory listing or a review roundup. Wrong audiences trace to a comparison page or an old press mention. Invented features often trace to a competitor's comparison table or a forum comment repeating a rumour. In my experience, eight times out of ten the trail ends at a page you can name within fifteen minutes.
Step three: classify the source, because the fix depends on it.
If it is your own page, this is the best case. Update it, make the correct fact prominent and plain, and make sure the page is crawlable. The correction will propagate as models re-crawl and as browsing-based answers read the fresh page. This is also the signal to ask why your own site was not the source of record in the first place: usually the correct fact was missing, buried, or behind a script.
If it is a third-party page you control, like a directory profile or a marketplace listing, update it there. This is why keeping a list of everywhere your product is listed pays off: the fix is a login and an edit, not a hunt.
If it is a third-party page you do not control, like a review site or a blog post, you have three honest options. Ask for a correction, politely, with the right facts and a link to where they are stated. Give the site a reason to update, like new information worth adding. Or publish the correction prominently on your own pages so that future reads outweigh the stale one. What does not work is demanding, threatening, or spamming corrections across the web; it burns goodwill and rarely changes the answer.
Keep a simple log as you go: question, assistant, wrong claim, source page, fix applied, date. Over a few months that log becomes the most useful AI-search document you own, because it shows patterns no single check reveals. Step four: verify the fix, on a delay. AI answers do not update the minute you edit a page. Browsing-based answers can shift within days once the source page changes. Training-data-based beliefs take longer, sometimes months, because they are baked in until the model's data is refreshed. Re-ask the original question every week or two and note what changed. The trend matters more than any single answer.
A few things worth knowing about what does not work.
You cannot report an error to the model and expect a correction to stick. Some assistants accept feedback, and it may help at the margins, but the next conversation starts from the same sources. The fix has to happen at the source, not in the chat.
You also cannot shout louder than the wrong page. Publishing ten new blog posts does not outrank one stale page the model already trusts. The unit of work is the source the model reads, not the volume you publish.
And do not pick a fight with reality. If the assistant says users complain about your support response times, and users do complain about your support response times, the fix is not reputation management. It is better support, then visible evidence of better support in the places the model reads.
Two more levers are worth pulling where they exist. Some platforms now offer official claim or profile routes for businesses, and claiming yours gives the model a first-party record to prefer over third-party guesses; the details differ per platform and change often, so check each one's current process. And where an assistant keeps repeating a claim with no findable source, publish the correct fact in the exact question's shape: a plainly titled page or FAQ entry that answers the question word for word gives the model something better to repeat.
One mindset note: expect to do this more than once. New wrong claims appear as your product, pricing and category change, and as new pages about you go live. A quarterly check of the questions buyers actually ask is usually enough to catch drift early, before a wrong fact hardens into the answer everyone sees.
The deeper point is that AI answers about your brand are downstream of your public record. Every wrong answer is a map to a page that needs work, and the work is ordinary: accurate pages, current listings, plain language, nothing important hidden from crawlers. Brands that treat each wrong answer as a free audit find the same handful of root causes over and over, and each one they fix lifts every assistant at once.
The brands that struggle are the ones arguing with the chat window instead of editing the record it reads.
Disclosure: I am the founder of Vouch (https://vouchgeo.com), a tool that tracks what AI assistants say about your brand and shows the sources behind their answers, so corrections land where they matter.

