AI is telling people the wrong thing about your business — and there's no one to email.

Someone forwards you a screenshot. They asked an assistant about your company and it answered with total composure: a price that expired two years ago, a service you dropped, an office you closed, a founder who left. No hedging, no source, no way to tell where it came from. And the buyer who saw it before you did has already moved on.

The first instinct is to find the complaint department. It's the right instinct for every other channel — a wrong Google listing has an edit button, a bad review has a flag, a journalist has an inbox. AI answers have none of that, and the hours spent looking for it are hours the wrong answer keeps getting served.

Why AI gets your business wrong in the first place

It helps to know that almost none of this is malice or even, strictly, error. A model is producing the most plausible continuation given what it has absorbed and whatever it retrieved a moment ago. When your own material is thin, stale, or contradicts itself, plausible and true come apart — and the model has no way to notice, because nothing in the process is checking.

Three conditions make it much more likely, and all three are yours to change:

• The truth is hard to reach. If the current fact lives only in a PDF, a slide, an image, or a page an engine can't fetch, it may as well not exist. Retrieval works on what it can read.
• The truth contradicts itself across the web. Your site says one thing, a directory says another, an old landing page says a third. Faced with three versions, an engine picks — often the one repeated most, which is usually the oldest.
• Nothing states the fact plainly. Buried in a paragraph of marketing language, a fact is inferable but not liftable. Engines quote what's stated; they guess at what's implied, and a guess is where the error enters.

Three causes that look identical in the answer

Every wrong answer reads the same from the outside — confident, specific, unattributed. But underneath there are three very different problems, and treating them the same is why corrections so often fail to stick:

Telling them apart is the whole diagnosis, and it's not something you can do from the answer alone — you have to look at what the engine was working from. That's why the useful version of this check captures more than a verdict: what was said, by which engine, and what it cited, if anything. An answer with a source you can open is cause one. An answer whose source describes a different company is cause two. An answer with nothing behind it at all is cause three.

Why there's no one to email

Now the uncomfortable part, and the reason most advice on this topic wastes your time. Contacting the AI provider does not fix the answer. Not because the companies are unhelpful, but because of how the thing is built.

What a model learned during training isn't a row in a database that support can edit. There's no record to open and correct. Feedback channels exist and are worth something in aggregate — they inform future work — but nothing about them promises that the next person who asks about your company gets a different reply, and nothing about them operates on a timeline that matters to you.

The retrieval layer is more tractable, but it still doesn't have a form. When an engine looks something up before answering, it takes what it finds. Change what's findable and you change the answer. That's a lever, and it's the only one anybody hands you.

Which is better news than it sounds, because it's a lever you control — no queue, no permission, no waiting on a company that will never write back. It's the difference between appealing to a system and simply outranking it as a source about yourself. And unlike an appeal, you can tell whether it worked.

What actually works: publish the truth, loudly

Since you can't edit the model, you compete with it. The correction that works is a publishing move, not a support ticket: put the accurate version somewhere engines read, state it in a form they can lift, and make it easier to find than the version they're currently using.

The shape of it:

1. Write the fact down plainly, where you control it. On your own site, in ordinary sentences that answer the question directly — not implied by a brochure paragraph, not living in a PDF. If the wrong answer is about pricing, there is a page that says what it costs.
2. State it in machine-readable form as well as prose. Structured data is how you say the same thing again in a format that doesn't depend on interpretation. Prose and markup agreeing is the point; markup that contradicts your own page is worse than none.
3. Fix the contradictions you're still publishing. The stale directory entry, the old landing page, the outdated profile. You don't need to erase the internet — you need to stop being the source of the confusion.
4. Repeat the fact consistently everywhere else. The same numbers, the same names, the same wording, across every profile and page that mentions you. Consistency is what makes a fact look settled rather than contested.

We don't publish the specific correction structures we generate, or which signals we weight when deciding a fact is established — those are the parts a look-alike tool would copy and apply badly, and getting them wrong makes a brand look more contradictory, not less. The principle generalises fine without them: engines don't arbitrate truth, they follow the strongest signal, and you can be that signal.

Written out as four steps that sounds like a project. In practice it's a short list of specific sentences on specific pages, and you don't have to work out which ones — that's the part we automate. Hallucination Watch asks the engines the questions your buyers ask, captures what each one said and what it cited, judges it against your verified facts rather than a guess, and hands you the correction already drafted: the wrong claim, the true version, and the markup to publish alongside it. You review it and publish. Most brands are looking at a ranked list of real problems within an afternoon of starting, not a research project.

When the problem is that AI thinks you're someone else

Cause two deserves its own treatment, because publishing more about yourself doesn't fix it — and can make it worse. If an engine has merged you with another company, every new page you publish is more material attached to a blurred identity.

The fix is disambiguation: making it unmistakable which organisation you are. That means the identifiers that travel with an entity rather than a page — consistent naming, the profiles that anchor you, the public knowledge bases engines lean on when resolving who's who, and explicit statements of what you are not.

Why you have to re-prove it, not just fix it

Here's what makes accuracy different from most GEO work: publishing the correction is not evidence the correction worked. You've changed the inputs. Whether the output changed is a separate question, and the only way to answer it is to ask again.

The lag is real and uneven. Retrieval-based answers can shift within days of the new page being crawlable. Anything leaning on what the model already absorbed moves on a much slower cycle, and may never fully let go — you're competing with the old version rather than deleting it. Engines also disagree with each other: it's routine to see the corrected fact in one and the stale one still in another, which means "is it fixed?" doesn't have a single answer.

So the loop is detect, diagnose, publish, re-ask — and the last step is the one that turns it from hope into a result you can show someone. It's also the step that catches a correction that quietly regressed, which happens more than you'd expect. You don't have to remember to run it: put it on a monthly cadence once and the re-check happens on its own, with a plain summary of what changed since last time.

You can't argue with a model. You can out-publish the thing it's reading — and then check that you did.

That whole loop is what Hallucination Watch runs: ask the engines about you, capture what each one says and what it cited, sort the wrong answers by cause, draft the correction, and re-run to prove it moved. AI Pulse puts it on a monthly cadence, which matters because accuracy decays — you change a price, an engine picks up an old page, and nobody finds out until a buyer repeats it back to you.