Generative search amplifies big brands, arXiv study finds

Dominant brands in the top quintile of market share get cited 1.4 to 2.8 times more often than the average brand across a sample of 1,176 top market players. That is the headline number from a new arXiv paper auditing brand bias in generative search, and it should disabuse anyone still clinging to the idea that AI search engines are a leveler for smaller players. They are not. They are a popularity contest with better production values.

The study's more interesting finding is not the popularity bias itself, which most marketers have already intuited from watching ChatGPT answer "best enterprise cloud provider" with the same three names every time. It is where that bias comes from. The researchers tested whether grounding a model's answers in live web retrieval, the mechanism vendors pitch as their defense against stale or skewed training data, actually corrects for it. It does not. Popularity bias is "already present in parametric memory," the paper notes, and external retrieval "fails to mitigate" it. In plain terms: giving the model access to today's internet does not stop it from reaching for the brand it already knows.

Familiarity as a moat, not a feature

The paper's second experiment is the sharper one for brand strategists. Holding product attributes constant in controlled, in-context tests, the researchers isolated brand familiarity, meaning the ranking advantage a brand gets simply from being lodged in an LLM's weights, from brand popularity, meaning its actual market share. These are different things, and conflating them is where most corporate AI-visibility strategies go wrong. A mid-sized but well-documented company (heavy Wikipedia presence, frequent trade press coverage, a long Securities and Exchange Commission filing history) can rank ahead of a larger but quieter competitor, because the model has simply encountered its name more often during training. Popularity is about what the market does. Familiarity is about what the model has read. Generative search rewards the second whether or not it tracks the first.

This has a direct and uncomfortable implication for the sectors that read TCE. Financial services firms, multilateral institutions, and large industrial groups tend to assume that AI visibility is downstream of market position: be the biggest, and the citations follow. The arXiv findings suggest something closer to the reverse causality for anyone below the top quintile. A regional bank with $40 billion in assets and a sparse public footprint will lose a generative-search citation battle to a smaller fintech that has spent five years generating earned media, policy commentary, and third-party analysis. The model is not weighing balance sheets. It is weighing training-corpus density.

The quintile problem for everyone else

This matters most for the long tail, which in B2B terms means most of TCE's audience. A handful of hyperscalers, money-center banks, and legacy industrial conglomerates occupy the top quintile and absorb a disproportionate share of citations almost by default. Everyone else, including well-capitalized mid-market players, regional development banks, and sector-specific standards bodies, is competing not against the market leader's product but against the market leader's documentation history. That is a solvable problem, but not the one most marketing teams are resourced to solve. Chasing market share to win AI citations is a decade-long project. Building the kind of citable, parseable, third-party-corroborated content footprint that increases parametric familiarity is a matter of months, if a firm knows to look for the right things: attributable data, named spokespeople, consistent technical vocabulary that matches how journalists and analysts describe the category.