AI Just Told a Customer Something False About Your Company

Assistants misstate business facts every day, and there is no correction desk to call. Each type of error points to its own repair, and all of them happen upstream.

Sooner or later, a customer asks an AI assistant about your company and gets back an answer you never wrote. When the answer is wrong, the first instinct is to argue with the chatbot, and the second is to look for someone at the AI company to complain to. Neither works, and knowing what does has become part of my job.

At ESBO Ltd, the digital PR and link-building agency I run, we have checked what assistants say about client brands every month since early 2025. The errors are rarely creative. They fall into five types, each pointing to its own repair.

The scale is not a niche worry. When BBC journalists reviewed answers from ChatGPT, Copilot, Gemini and Perplexity, 51% of responses about the news carried significant issues, 19% of the answers citing BBC material introduced factual errors, and the published examples included keeping a former prime minister in office months after he had left. 

A system that misplaces a head of government will not treat your pricing page with more care.

Assistants get companies wrong in five ways

The first is stale facts served as current: retired products, old prices and a founder who left in 2022 still listed as the boss. The second is mistaken identity, where a similarly named business donates its lawsuits or its reviews to your answer. The third is category drift, a consultancy described as a software vendor or a broker described as a bank.

The fourth is the old complaint promoted to headline, one bad thread from years ago framing the entire reply. The fifth is absence, when the assistant names three competitors and stops. In our checks, the fifth is the most common and the least noticed, because nobody takes screenshots of an answer they never saw.

You cannot correct the model, only its sources

There is no correction desk. The AI companies run feedback buttons rather than editorial departments, and a mid-sized business has no number to call when a machine invents its shutdown. Arguing inside the chat changes that one conversation and nothing else, and the slow route, a legal claim over a hallucination, has barely been tested anywhere and no customer conversation waits for a court.

The assistant said what it said because of the pages it read. So the repair always has the same shape: change the reading material and wait. Answers built on live retrieval can move within weeks of the sources changing, while facts absorbed during training move more slowly. Either way, the job is editing a library.

Each error points to its own repair

Stale facts respond to dated statements. Put the current prices, products and leadership on your site in plain sentences with dates attached, then push the new facts into fresh third-party coverage, because the machines favor recent pages. When Ahrefs studied the lists ChatGPT cites, 79.1% had been updated within the year, so a fresh page beats a stale one even when the stale one ranks.

Mistaken identity responds to consistency. Pick one name, spell it one way everywhere, including the directories nobody remembers filling in, and publish a page that calmly separates you from your twin. Category drift responds to the company your name keeps in print: the words you want attached to your business have to sit beside it in sentences other people publish, which is the case for brand mentions.

The old complaint cannot be deleted, only outweighed. Answer it wherever a reply box exists, since replies are text the machines read too, then stack newer coverage on top until the old thread loses the recency contest. Absence has no shortcut at all. For buying questions, the assistants mostly relay ranked lists and press, so if those pages skip you, the answer does too, and the repair is the ordinary work of becoming citable.

The check belongs on the monthly calendar

The routine we run takes under an hour. Write the five questions a buyer would ask about you: who is this company, is it legitimate, who are the best providers in its category, how does it compare with its closest rival, and is it still operating? Run them through two different assistants, because they read different libraries and fail differently. Log the answers and, more usefully, the sources they cite, since the citations are your repair list. (Write everything down. Memory flatters last month’s results.)

Most months, the check comes back clean, which makes it feel like paying for nothing right up until the month it does not. That is fine. The point of a check is the month it catches something.

The assistant who wronged your company believed what it read, which is both the problem and the repair. Nobody can argue it out of a false belief in the chat window, but anyone can change what it reads next. Give the machines better material, give them a few weeks, and ask again.

Assistants misstate business facts every day, and there is no correction desk to call. Each type of error points to its own repair, and all of them happen upstream.

Sooner or later, a customer asks an AI assistant about your company and gets back an answer you never wrote. When the answer is wrong, the first instinct is to argue with the chatbot, and the second is to look for someone at the AI company to complain to. Neither works, and knowing what does has become part of my job.

At ESBO Ltd, the digital PR and link-building agency I run, we have checked what assistants say about client brands every month since early 2025. The errors are rarely creative. They fall into five types, each pointing to its own repair.