The AI Is Describing You—And Getting It Wrong

Buyers now ask a model what your company does before they ask you. Do you know what it says?

Try something uncomfortable this afternoon. Open a chatbot, pretend you’re a prospect who has never heard of your company, and ask it what you do, who you serve, and how much you cost. Then read the answer slowly.

If you’re like most people who run this test, one of two things happens. Either the model shrugs and gives a generic non-answer, or—worse—it answers with total confidence and gets something wrong. It invents a product you sunset two years ago. It quotes a price you never charged. It places your headquarters in a city you’ve never had an office in. It confuses you with a competitor who happens to share three words in your name.

We’ve spent a year on this newsletter talking about visibility: how to get cited, how to get mentioned, how to stop being overlooked. But there’s a quieter problem sitting underneath all of it. Getting mentioned is worthless if the mention is wrong. A confident, incorrect answer isn’t a gap in your marketing. It’s an active liability that repeats itself thousands of times a day, to exactly the people you’re trying to win, and you’re not in the room to correct it.

Why the machine misremembers

Models don’t lie on purpose. They average. When you ask a model about your company, it isn’t reading your homepage in real time and reciting it back. It’s reconstructing a picture of you from everything it absorbed during training and whatever it can pull from a scattered handful of sources at answer-time. If those sources disagree, the model doesn’t flag the contradiction—it smooths it into a single fluent sentence and hands it over like settled fact.

So the question isn’t really “why does AI hallucinate about my brand.” It’s “what does the web actually say about me, and does it agree with itself?” Usually the answer is no. Your homepage says one thing. An old press release says another. A three-year-old review site lists a pricing tier you retired. A directory you forgot you were ever in has your category wrong. A journalist once described you with a slightly-off analogy and that phrasing got quoted, then re-quoted, until it calcified into the internet’s consensus version of you.

The model reads all of it and splits the difference. The version of your company that lives in AI answers isn’t the version on your website. It’s the statistical center of gravity of every claim ever made about you, weighted toward whatever got repeated most. If the loudest, most-repeated claims are outdated, that’s the you that buyers meet first.

The compounding cost of a wrong answer

Here’s what makes this sharper than an ordinary factual error. A wrong line on a page is passive—someone has to find it, read it, and believe it. A wrong answer from a model is delivered on demand, personalized to the question, and wrapped in the authority of a system people increasingly trust more than the open web. When a buyer reads “they don’t integrate with the tool I use” or “they’re really more of an enterprise product,” they usually don’t verify it. They just quietly cross you off and move to the next option. You never see the loss. There’s no bounce in your analytics for a deal that died in a chat window you’ll never read.

And these errors are sticky. Because models learn from the web, and the web increasingly contains AI-generated summaries, a mistake can launder itself into permanence. The model gets you wrong, someone publishes a post built on that wrong answer, and now there’s a fresh source confirming the error for the next model. Left alone, a small inaccuracy doesn’t fade. It accretes.

Fixing the record

The instinct is to fight the model. You can’t. You can’t argue with it, you can’t submit a correction form, and you certainly can’t out-shout it in the moment. What you can do is change what it reads.

Start by finding out what’s actually being said. Ask several different models the same set of buyer questions and write down every wrong or outdated claim, word for word. That list is your real to-do list, and it’s usually more honest than any brand audit, because it’s showing you the version of yourself that customers meet before they meet you.

Then go make the truth boring and consistent. Say the same core facts—what you do, who it’s for, what it costs, what it integrates with—in the same plain language everywhere you control: your homepage, your about page, your pricing page, your profiles on the directories and review sites that models actually pull from. Models trust agreement. When ten sources say the same clean sentence, that sentence becomes the answer. When they say ten different things, the model picks one, and you don’t get a vote.

Finally, hunt down the loud old wrongness. The retired product page still ranking. The stale third-party listing. The interview where you described yourself in a way that no longer fits. You may not be able to delete all of it, but you can bury it under a larger, fresher, more consistent chorus of the correct version. You’re not editing a single page anymore. You’re managing a consensus.

The old job was making sure people could find you. The new one is making sure that when a machine speaks for you—and it is speaking for you, right now, in conversations you’ll never see—it gets your name right. Go ask it what you do. The answer is your next quarter’s homework, whether you like the grade or not.