Visible but misunderstood: why being named by AI is not always a win

Most of the conversation about AI answers is about being in them. Does ChatGPT name you when a buyer asks? Does Perplexity? How often, and ahead of whom? That is the right first question, and it is the one a visibility number answers.

It is not the only question. There is a second one that visibility cannot see: when an engine names you, what does it say you are?

Named is not the same as understood

Picture a company that spent years as one kind of business and then changed direction. New buyers, new product, new site. Ask an AI engine about the new category and the company is sometimes named. A visibility tracker counts that as a mention and moves on.

Now read the answers. Many of them name the company and describe the old business: the product it retired, the market it left, the thing it was known for three years ago. The buyer reading that answer is being told this is a company for something it no longer sells. The buyer who wants what it sells now does not hear its name in the right context at all.

Both of those answers count as a mention. One of them is a problem.

Why visibility cannot see it

A visibility score asks whether the business is in the answer, and how prominently. It does not ask whether the answer is about the right business. So a company that repositioned can watch its visibility rise and not know that part of the rise is the old identity getting louder.

That is why understanding needs its own number. We call it entity accuracy: of the answers that name you, the share that describe what you do now. It sits next to visibility, not inside it, because the two can move in opposite directions. More mentions in the old category raise visibility and lower entity accuracy. That is not progress, and a single blended score would hide it.

How it is measured, briefly

The categories come from the business, not from us. At the start, the business confirms the ways a company like it can be described, and marks which are current and which are legacy. Every answer that names the business is then placed in one of those categories from its own words only. What we already know about the business is not used to decide what the engine meant, so a vague answer does not get the benefit of the doubt.

Claims get the same treatment. A claim that was never true is incorrect. A claim that was true of the old company and is not true now is outdated. They are listed apart because the fix is different.

Where the old identity comes from

An engine is not remembering the old company out of habit. It is reading pages. When it describes the old business, there is usually a page behind it that still does:

• A retired product or solutions page that still resolves, or that now returns an error but is still cited.
• A profile on a directory, a marketplace or a review site that was written before the change and never updated.
• Press coverage and partner pages from before the change, which are often better linked than anything written since.
• The business's own old copy, left in a footer, an about page or a press kit.

Engines cite what they can read, and older pages have had longer to be linked and cited. A new site that describes the new business clearly is necessary. It is often not enough on its own, because the old pages are still there saying something else.

What to do about it

Start with the pages the engines cite when they describe the old business. Each legacy answer comes with its sources, and the sources are the list.

• Old pages on your own site: redirect each one to the page that replaced it, so the old address stops answering for the old business. A page that now returns an error is not the same as a redirect; engines can keep citing the error for a long time.
• Profiles you control elsewhere: update the description, the category and the products, in the words a buyer would use for what you do now.
• Pages you do not control: a newer source that tells the current story, including what changed and when, is believed more readily than a flat denial. An engine that can see the old pages reads "this was never true" as false. It reads "this was true until last year, and here is what it is now" as an update.

Then measure again. Entity accuracy is slow to move, slower than a simple wrong fact, because the old pages keep existing. The trend over a few months is the signal. A single month is not.

The short version

Being named by AI is the first half. Being named as what you are is the second, and after a change of direction it is often the harder half. Count them separately. A mention in the wrong category is not a win; it is a buyer sent to the wrong place, delivered in a confident sentence.