Do AI Engines Favor Big Brands Over Smaller Ones?

AI answer engines do mention big brands far more often than small ones, but the research says they are not attached to them. A University of Toronto team asked ChatGPT and Perplexity 50 unbranded questions about the best and most popular colas. Of 986 brand mentions, 62.2% went to 20 major brands and 9.0% to 20 niche ones. A controlled experiment published in June 2026 found something that looks like the opposite. Across 14,395 trials, brand identity explained 1.2% of how three LLMs ranked products, and the product details shown to the model explained 82.4%.

Both findings hold. This post covers how they fit together, what that means for a challenger brand, and a test that tells you which of two different problems you have.

Key takeaways

Four studies bear on the question. They use different designs, which is why their headline numbers seem to disagree.

Sources: Kamruzzaman, Nguyen & Kim; Chen, Wang, Chen & Koudas; Chu & Hou; Uberti-Bona Marin et al..

Only the first paper is peer-reviewed. The other three are arXiv preprints from university groups, and none is from an AI visibility vendor.

Models carry a familiar-brand prior. Kamruzzaman and colleagues asked four models to match brands from 15 countries with positive and negative attributes. In all eight model-and-direction settings, the models linked global brands with favorable traits and local brands with unfavorable ones. When the authors swapped brand names for the phrases "global brand" and "local brand," the pattern held and tended to get stronger. The models have learned that "global" means "good," whatever the specific brand.

Live engines lean toward leaders on generic prompts. In the Toronto cola test, Coca-Cola and Pepsi led on both engines, with Dr Pepper third. Perplexity was more skewed than ChatGPT, giving 67.9% of mentions to major brands and 5.8% to niche ones, against 56.3% and 12.3%. The same paper found that when the four engines it tested answered questions about niche brands, their sources agreed less (71–76% agreement against 76–81% for well-known brands). Engines converge on leaders and scatter on everyone else.

The prior shows up in other domains. A separate team testing investment advice found models consistently favored a few dominant names such as Apple and Microsoft across 567,000 generated recommendations, and the preference survived debiasing techniques.

Chu and Hou built product lists with one real brand, such as CeraVe moisturizer or Anker USB-C cables, and nine invented brands. They then varied the specs.

Kamruzzaman's paper shows the same thing from another angle. When the prompt said which country the user was in, GPT-4o picked the local brand over the global one 75% of the time. A single piece of context outweighed the "global is good" association.

The familiar-brand preference is a default. Models use it when neither the prompt nor the evidence gives them a better reason to choose. That is the incumbent tie-break, and it reconciles the numbers.

In Chu and Hou's experiments, the challenger's advantage was placed directly in the prompt, so the model could not miss it. A real challenger's advantage has to be retrieved, and retrieval favors leaders twice over:

Chu and Hou also ran a small retrieval experiment. In it, the real brand ranked poorly on embedding similarity and survived in 0% of runs. The authors call this result directional only. Still, it suggests the bias can fall at the retrieval stage too, against whoever's description fits the query worst rather than in favor of the incumbent.

So the field result and the lab result describe the same mechanism. Generic prompts supply no tie-breaking evidence, the evidence that retrieval does find is mostly about leaders, and the prior settles the rest. Where the evidence for a challenger is present and explicit, the prior gives way quickly.

Challengers should not expect the same treatment everywhere.

"Most popular" has a correct answer. Some of the Toronto prompts asked for the most popular colas, and Coca-Cola is the most popular cola. Part of that 62.2% is the engines answering accurately. The test also covers one consumer category on two engines, at one point in time.

Fictional brands are not small brands. Chu and Hou's invented names start with no reputation at all. A real challenger may have a thin or mixed record that the model has partly learned, which could help or hurt. Their main results also come from one product domain.

The marketing-language result is a warning, not a tactic. Adding fabricated clinical-trial claims to the challenger's description lifted it above the incumbent in 73.3% of trials. The authors used made-up claims on purpose, to find an upper bound. When all nine fictional competitors used the same language, the incumbent's survival rate went back up to 93.8%, and the individual payoff fell to almost nothing. It is deceptive, and it stops working once your competitors do it too.

None of the studies measures revenue. They measure mentions and picks. Those are only the top of the funnel.

A challenger brand has two possible problems that look the same on a share-of-voice dashboard. Either the engines know about you and default to the leader when nothing breaks the tie, or they are not finding evidence about you at all. The fixes are different, so separate the two before you act.

Elmo is an open-source, self-hosted AI visibility platform. It runs your prompt sets repeatedly across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, and records every brand named and every source cited on each run. Tag prompts as generic or constrained, and the generic-versus-constrained gap against your category leader is a query over your own data.

For the fundamentals, start with AI share of voice and AI competitor analysis, then how to track your brand in AI search. For the vocabulary, see the AI search glossary.