Entity collapse: when AI flattens a brand to one thing

From outside, entity collapse looks like a brand that AI describes accurately but only in part. Picture a multi-division group that most people know for a single consumer category. Ask an answer engine about that category and the group appears, described well. Ask instead about the industrial supplies, the financial services or the logistics arm that sits under the same parent, and something odd happens: the group is absent. A different set of companies answers the question, and the division that genuinely competes there is nowhere in the response.

This holds across ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode. The engines run on different indexes, but they read the same structural signals at the source, so a collapse on the site surfaces across all of them. The brand has not been described inaccurately. It has been described incompletely, reduced to its most famous face while its other businesses fall out of view. For the divisions that never surface, the effect is the same as not existing: in the place where buyers now ask, they are simply not part of the answer.

At the pattern level, three things combine. First, the strongest citation signal dominates. One division has been written about, linked to and referenced far more than the others, so when a system resolves the brand it settles on that best-known part and treats it as the whole. Second, the weaker divisions lack their own resolvable presence. They may have pages, but nothing gives them a distinct, machine-readable identity that an engine can hold separately from the parent's famous arm.

Third, and most fixable, nothing on the site tells a machine that the divisions are distinct parts of one organisation. The relationships that are obvious to a human reader, that these are separate businesses under a shared parent, are never stated in a form a crawler can read. Our research on what AI crawlers actually see found that structured data is routinely absent from the raw, non-JavaScript response, which is exactly the view these systems rely on. When the machine-readable relationships are missing there, collapse is the default outcome rather than the exception.

You can test for entity collapse directly. Write a short, category-level prompt for each division, the kind of question a buyer in that market would actually ask, and run it. Then read which organisation answers. If the same well-known division surfaces every time, and the others never appear in their own categories, suspect collapse. The tell is not that the brand is described badly; it is that only one part of the brand is ever described at all.

Run the prompts across engines rather than trusting a single one. A division might resolve on one engine and vanish on the others, which is itself a sign that its presence is thin and held together by a single weak signal rather than a stable, machine-readable identity.

Fixing entity collapse is structural, and it happens at the pattern level rather than through more marketing copy. Five things matter. Each division needs its own server-side structured data, present in the raw response, so that a crawler can resolve it without running any JavaScript. There should be one canonical organisation node that the divisions all connect to, so the parent is a single, stable identity rather than a set of competing fragments.

Naming should be consistent across the site and the markup, so a machine sees the same entity described the same way everywhere. And the relationships between parent and divisions should be stated explicitly on the site, so that a system does not have to infer, and usually fail to infer, which businesses belong together. Give each division its own entity home, a page that authoritatively defines it, and declare the distinct topics that division actually leads on, so the engine has somewhere to resolve each division to and a reason to surface it in its own category. Done together, these give each division a presence an engine can resolve on its own terms, rather than folding it back into the one part that already dominates.

This definition comes from entity resolution work developed through Growth.pro engagements.