When an AI Model is Wrong About Your Brand

There's no correction form, no editor, and nobody to appeal to. So what do you actually do?

In late 2024, a solar installer in Isanti, Minnesota started losing contracts it had already won.

The sales team at Wolf River Electric kept hearing the same reason. Customers had searched the company name, and Google's AI Overview told them Wolf River was being sued by the Minnesota Attorney General for deceptive sales practices. High pressure tactics, hidden fees, misleading claims about savings.

The Attorney General had sued four companies over solar lending. Wolf River wasn't one of them. Google's AI Overview did cite sources. According to Wolf River's complaint, you can read every one of them and find no mention of any lawsuit against the company.

Customers kept walking anyway. Wolf River eventually sued Google, and its complaint itemises what the false claim cost. A $39,680 contract terminated in March 2025. A $150,000 one two days later. Then $174,044 of nonprofit projects the week after that. The company has put its total damages somewhere between $110 million and $210 million.

But hold the lawsuit for a second, because the interesting moment came earlier. Before the lawyers, before any of it, someone at Wolf River saw that summary and asked the question every marketing team eventually asks.

Who do we call?

There's nobody on the other end

Every correction remedy a brand team knows assumes there's a publisher somewhere. A journalist gets something wrong, you email the desk. A review site is out of date, you claim the listing. Google shows a stale sitelink, you've got Search Console, structured data, a documented path with a queue at the end of it.

None of that exists here. There's no console, no ticket, no correction form, no editor. The output isn't sitting in a database you can edit. It gets written fresh every time somebody asks, never quite the same way, out of sources you had no say in.

People call legal instead. That road is less clear than it looks.

Wolf River's case is one of the earliest real tests of it, and it's still tangled up in procedural fights about which court even hears it. Around the same time, OpenAI beat a defamation claim in Georgia brought by a radio host over a fabricated ChatGPT output. These cases are slow, expensive and unsettled, and while they run, the claim is still out there doing damage.

So the urge to make it stop has nowhere useful to go. The question that does have an answer is a different one.

Where did this come from, and what do I change so the next answer comes out different?

Stage one: work out which of the four you're holding

Every misdescription we see is one of four things. In the output they look identical. Underneath they have almost nothing in common, so guessing wrong here costs you weeks.

Stale source. The page was right when you published it. Then the price changed, or a feature got deprecated, or you rebranded, and nobody ever took the old page down. It's still live, still indexed, and it's still the clearest answer to the question somebody just asked. So the model uses it. From where the model is sitting nothing has gone wrong: it's describing you accurately, just an older version of you. This is the most common type by a distance and the easiest to fix, because the page belongs to you.

Your own contradiction. The docs say one thing, the pricing page says another, the sales deck says a third. The model picks whichever is best structured, which is very often not the current one. Second most common, and the most awkward, because the source is you.

Competitor framing. Somebody in your category has built a page called "you versus them" and written both halves of it. Nothing on it has to be false for this to hurt you. They chose which facts went in, they wrote your column, and they decided what counted as a fair comparison in the first place. Same goes for the listicles where the running order tracks affiliate payouts more closely than it tracks product quality. A model has no way to tell any of this apart from a neutral review. It's shaped like a comparison, so it gets read as one, and your competitor ends up narrating you.

Fabrication. You go looking for the source and there isn't one. This is rarer than people think. Most of what gets called hallucination turns out on inspection to be one of the first three, sitting in plain sight once you actually check. But it does happen, and Wolf River is what it looks like. A real enforcement action against real solar companies, real reporting about it, all recombined into a sentence about a lawsuit that never existed. Every input was true. The output wasn't. This is also the one type you can't fix by editing a page, because there is no page.

So the question that sorts them is simple. Can you find the source? A stale page, your own contradiction, a competitor's comparison: each of those leaves a trail that ends at a specific URL, and that URL is what tells you which fix to reach for. Work this out before you commit to anything, because a wrong diagnosis here costs you a quarter.

Stage two: trace it

Find the actual source, not a general sense of where it came from.

This is the step most teams skip, and it's why so many AI visibility projects turn into a content calendar that changes nothing. Wolf River's whole case rests on it. The false claim arrived with citations attached, which is precisely what made it convincing to the customers who read it, and the only way anyone established it was invented was by opening each cited link and finding nothing there. Citations are not verification. Somebody still has to read them.

There are two things to establish before you spend an hour on fixes.

Is the model looking this up, or remembering it? When it searches and cites as it answers, the thing that's wrong is sitting in a document that actually exists, and a document can be found and usually changed. When it answers from memory, the claim went in during training, months before anyone asked the question. There's no document to correct. Nothing you publish this quarter reaches it.

Is it one model or all of them? One model getting it wrong is usually a source problem. Every model getting it wrong is usually a you problem, because the true answer isn't published anywhere retrievable and they've all independently reached for the same third party filler.

That distinction drives everything after it. It's also why you can't do this by hand. You're not checking a prompt, you're checking whether a claim keeps recurring across models and sub queries and time. Wolf River found out through cancelled contracts, which is the most expensive detection method available.

Stage three: fix the source, not the model

You can't correct the model. You correct what it reads. In order of leverage:

If you own the source, kill it or update it. Old pricing pages, stale comparison content, dormant subdomains, a PDF from 2022 sitting in a resources folder. Taking a page down is faster and more certain than trying to drown it out with something new, and it's the one move here that needs nobody's approval but your own. Do this before anything else.

If the true answer isn't published anywhere, publish it, in the shape of the question. Not a page that gestures at pricing, a page that states it. Models pull out specifics and skip over vagueness, and vagueness reads as absence rather than mystery.

If a third party owns it, what works is a factual correction with evidence attached, not a complaint about tone. Review sites and listicle publishers update far more often than people expect, because being wrong is bad for them too. The ones that won't update are usually being paid not to, and you route around those by outpublishing rather than negotiating.

If the model is remembering it rather than looking it up, settle in. You cannot reach inside a trained model and delete something. All you can do is make the true answer so much easier to find that live search overrides the memory, and that takes two or three quarters of steady publishing rather than one campaign. Your exec team would much rather hear that now than in November.

Stage four: check it actually landed

A fix in one place doesn't land everywhere at once. A model that searches live can pick up the change within days. A model answering from memory can keep repeating the old claim for months after you've corrected the source it came from. So checking once and calling it done is how you end up explaining the same problem twice to the same person.

There's a footnote to the Wolf River story that's worth sitting with. After the suit became public, a law professor ran the same searches and got no AI Overview at all. The likely explanation is that Google simply switched the feature off for those queries. That's the closest thing to a remedy anyone got. Not a correction and not an acknowledgement. The claim appears to have simply stopped showing up, and getting even that far took a lawsuit.

So keep checking, model by model, on a set schedule. Track the exact claim you're trying to kill rather than your overall visibility score, because the score can climb while the one sentence that's costing you deals sits exactly where it was. And when it clears in two models but not the third, that isn't a failed fix. That's just what fixing this looks like.

What to do before any of this happens

Three things worth having in place while nothing is on fire.

A monitored claim set. The ten or fifteen factual statements about you that would lose you a deal if a model got them wrong. Price, integrations, security posture, who you're built for, whether you're currently being sued.

One source of truth for each. A single page, current, structured, answering that claim directly. If there are two pages, you've already got the second failure mode waiting to happen.

An owner. Not a working group. One person whose job it is to look at this monthly and escalate when it's ugly.

Wolf River's real exposure was never that a model got it wrong about them. Models get it wrong about everybody. Their exposure was the gap between the claim going live and anyone noticing, and then the second gap, between noticing and being able to say where it came from.

The screenshot is coming either way. What decides whether it's a bad Thursday or a bad quarter is whether you can answer "where did this come from" in an hour or a fortnight.

Quadrant watches what AI models say about your brand, across models and over time, so you hear about it from a dashboard rather than from a cancelled contract.