Why AI now decides how your business is described

What changed, exactly?

For thirty years a search engine returned a list and a person chose from it. The business kept control of its own description, because the description a searcher read was the one the business had written on the page they clicked.

That intermediate step is disappearing. A generative engine does not return a list to choose from. It returns an answer, already composed, in which a business is described in sentences nobody at that business wrote. The page is still there. It is simply no longer the first thing read, and often not read at all.

A fifth surface has appeared, and it is the only one written by someone who has never spoken to you.

Who writes the description?

Nobody, in the sense that matters. It is assembled — from the business's own pages, from directories, from a professional profile, from whatever third parties have published, and from the statistical shape of how similar businesses are usually described.

That last ingredient is the dangerous one. Where the material is thin, the engine fills the gap with the category default. A business that has not made itself legible gets described as the average of its category, which is precisely the reduction that perception work exists to prevent.

The mechanism is the same one a human market uses. It files you under the nearest available category when you have given it nothing better. The difference is speed, scale, and the fact that the filing is now written down and repeated verbatim to everyone who asks.

What does the research actually say?

There is one controlled study worth citing, from a team including Princeton, which tested nine tactics against generative engines and measured which ones changed the likelihood of being cited.

• Statistics with a clear source roughly +26 percent
• Citations from named experts roughly +28 percent
• Explicit source references roughly +25 percent

And the tactics that did not work: keyword stuffing, padding, artificial simplification of the language. The things a content agency sells by the kilo are the things that measurably do nothing.

Read the list again and notice what the three effective tactics have in common. None of them is about writing more. All three are about being verifiable. An engine assembling a description prefers material it can attribute, because attributable material is material it can defend.

Why do keyword tactics fail here?

Because the engine is not matching a query to a page. It is composing a statement it will be held to, and it selects material accordingly.

Keyword density was a signal in a system that ranked documents. In a system that writes sentences, density is noise at best. What counts is whether a passage can be lifted, attributed and stood behind — which is a question about the substance of the passage, not its vocabulary.

This is unusually good news for a business with something real to say, and unusually bad news for one whose surfaces were written to rank.

What can a business actually control?

Not the description. The material the description is made from.

Three things are within reach, and none of them is a technical trick. A clear, repeatable statement of what the business is, placed where it can be found and phrased so it can be quoted in one sentence. Consistency of that statement across every surface, so an engine assembling from three sources finds the same reading three times instead of three variants. And verifiable substance — figures with a source, positions with a name attached, claims that can be checked.

There is a fourth, mechanical and frequently overlooked: making sure the material is reachable at all. A page rendered entirely by JavaScript is invisible to a share of these crawlers, and several content delivery networks now block AI crawlers by default. A business can be doing everything right editorially and still be absent from the assembly, for reasons nobody there ever chose.

How do you check yours?

Ask. It takes five minutes and it is the single most uncomfortable audit available.

Put the name of the business to three or four generative engines with no other context, and read what comes back as if you were a prospect who had never heard of it. Then ask the harder version of the question: what kind of business is this, and who is it for.

Two failures are common and they are different problems. The engine may describe the business as the average of its category, which means the material is too thin to distinguish it. Or it may describe it accurately but blandly, which means the material is consistent and says nothing worth repeating.

The first is a conformity problem. The second is a decision problem, and no amount of publishing will fix it.

Is this a transient advantage?

Probably not, and the reason is structural rather than strategic. An engine that composes answers has to select its sources, and selection requires a criterion. Verifiability is the only criterion available to a system that cannot check anything itself: it can tell whether a claim carries an attribution, not whether the claim is true.

That means the incentive points the same way for as long as machines write the answers. Businesses that are easy to quote accurately will be quoted accurately. The rest will be described by the average of their category — a description that belongs to no one and distinguishes nothing.

Which returns the problem to where it started. The engine is new; what it rewards is not. It rewards a business that has decided what it is and made that decision checkable.