Before a customer rings you, buys from you or signs up, there is a good chance they have already asked an AI assistant about you. What it told them is not under your control. It is, however, something you can read.
A case published in MIT Sloan Management Review shows what happens when a company finally looks. The article, “The New Rules of Brand Visibility in AI Search”, describes a regulated investment platform in New Zealand. The authors call it Stakewise, which is a pseudonym.
What the firm saw
The Stakewise team noticed that visitors were arriving from ChatGPT. By the middle of 2025 that source was sending roughly 700 visits a month, up from a handful the year before. Those visitors also looked good: lower bounce rates than other referrals, and more pages viewed per visit.
On the standard measures the site was doing well. What those measures could not show was what the AI had said to people before they clicked.
So the team asked. They put to ChatGPT the questions customers were likely to ask about fees, fund choices and comparisons with competitors. The answers, the authors write, were confident and fluent, and often wrong.
What was wrong
The article lists several errors. A claim repeated across sources said one of the firm’s products offered five funds. It offers seven.
Fee structures were described incorrectly. A protective safeguard in the product was presented as a restriction. And features of the firm’s platform were mixed up with those of a separate product.
Each of these is the kind of mistake that matters when the reader is deciding where to put retirement savings. Each arrived with the same calm tone as a correct answer. Nothing about the wording signalled doubt.
How they fixed it
A specialist agency traced the wrong answers back to the web pages they appeared to come from. It then corrected the underlying content at the source.
That is the part that matters. The AI was not the thing being edited. The pages it learned from were.
The article’s detail on exactly how the agency worked sits behind MIT’s subscription, so this post does not describe it. The principle is enough to act on: a wrong answer has a source, and a source can be corrected.
It also changes who is responsible. A firm cannot edit ChatGPT. It can edit its own website, its listings and the comparison pages that mention it, and those are the places the wrong claims were found.
The effort was not large. It took someone willing to ask the questions, read the answers against the facts, and follow each error back. Most firms have never done the first step.
What this does and does not show
This is one case, anonymised, from a research project based on 15 interviews with senior marketing and digital leaders. The 700 visits a month is one firm’s number from one point in 2025. It does not show how often AI assistants get things wrong about companies in general.
It does show that it can happen to a careful, regulated business with a good website. And it shows that the cause can be traced and sometimes fixed, which is better than assuming nothing can be done.
The researchers also found that many leaders were relying on traffic figures that cannot see any of this. Analytics count the people who arrive. They say nothing about the answers given to those who never did.
The five-minute check
You do not need a tool, a budget or a specialist. You need a list of questions and a few assistants.
Write down ten questions a customer would ask. What does it cost. What is included. Where are you based. What are your opening hours. How do you compare with the named competitor. Use the customers’ words, not your marketing language.
Ask them in more than one assistant. Try ChatGPT, Claude, Gemini and whatever else your customers use. Ask each the same ten questions in a fresh conversation.
Compare against the facts. Put the answers beside your actual prices, terms and policies. Mark each as right, wrong or missing. Pay particular attention to numbers, because that is where the Stakewise errors were.
Trace the wrong ones. Ask the assistant where it got the claim, or search for the wrong phrase yourself. Often the origin is an old page, a comparison site or a directory listing that nobody updated.
Correct the source and repeat. Fix the page you control. Ask the owner of the page you do not control to change it. Check again in a month.
Why the numbers matter most
Pay attention to which errors came up. They were figures: a fund count, a fee, a feature. An assistant that gets a tone or a summary slightly wrong does little harm. One that gets a price or a count wrong can send a customer elsewhere, or set an expectation you then have to break.
It also compounds. If one page says five funds, other pages copy it, and the assistant sees the claim repeated and treats it as settled.
Who should own it
In most firms nobody does. Marketing owns the website, sales owns the price list, and the customer who asks an AI assistant belongs to neither. The check is quick enough that it can be a quarterly job for whoever owns the customer relationship.
You may find everything is accurate. That is worth knowing too. If it is not, you have learned something about your business that no dashboard would have told you, and found it before a customer did.
Source: Shahper Richter, Alla Lvovich, Wietske Van Osch and Alex Richter, “The New Rules of Brand Visibility in AI Search”, MIT Sloan Management Review, 13 October 2026. The full article is partly paywalled; this post uses the sections readable without a subscription.