Two academics published a set of experiments on brand bias in LLM product recommendations. It is the clearest account we have read of how an unknown brand actually competes inside an assistant's answer.
Xi Chu of Trine University and YuPeng Hou of Texas A&M posted it to arXiv on 16th Jun. It is a preprint and carries no peer-review statement.
The design is ten skincare products at a time: one real brand such as CeraVe, and nine fictional names. The fictional names went through generation, web-search deduplication and a recognition screen.
Three models were tested, GPT-4o-mini, Claude Sonnet and Gemini 3 Flash, in English and Chinese.
The first result is the one to sit with. All ten products carried identical ratings, prices, review counts and ingredient text.
The real brand was recommended in 100% of 670 valid trials. Not one fictional brand was ever chosen.
That dominance turns out to be brittle rather than absolute. Give the fictional brand better specifications and the models choose it about 96% of the time.
The threshold is very low. A 0.075-star rating advantage was enough to break the tie, and the paper's abstract rounds this to "less than a +0.1-star" advantage.
Decomposing the variance makes the mechanism plain. Once any distinguishing information exists at all, product parameters explain 82.4% of ranking variance.
List position explains 6.5% and brand identity 1.2%.
The authors call this a conditional monopoly. The barrier facing a new brand is not the incumbent's brand equity.
It is the absence of anything that tells the model the two products differ.
The second experiment is where marketing enters. Five kinds of persuasive language were added to the fictional brand's description, with every specification held equal.
The results split cleanly in two.
Authority language broke the incumbent's hold 73.3% of the time and social proof 50.7%, against a baseline near 4%. Anchoring, scarcity and loss aversion moved almost nothing, at 12.9%, 11.7% and 9.6%.
So the models largely ignore sales talk and treat credentials as evidence. The authority prompts included fabricated citations of the form "peer-reviewed clinical trial, n=120, p<0.01".
The authors put a price on it. Authority language bought the same advantage as a real improvement of +0.17 rating points.
That equals a 15.3% price cut, or 1.9 times the review count.
The models differ, and Claude differs most. Authority worked on Claude 55% of the time against 99% on Gemini.
Stacking authority with social proof cut Claude's rate to 21.2% while raising GPT's to 91.2%.
The third experiment is the one to read twice. It runs the same optimization across more and more of the nine challengers.
A single optimized challenger drops the incumbent's survival rate from 100% to 19.8%. With all nine doing it, the incumbent recovers to 93.8%, because the signal no longer distinguishes anybody.
Per-brand payoff falls from +0.802 for the first mover to +0.007 under universal adoption, in the authors' own payoff proxy. Across 4,745 trials, brands that did not optimize while their competitors did received zero recommendations.
The incentive and the collective result point in opposite directions. Optimizing always beats not optimizing, and everyone optimizing returns the field to roughly where it started.
One probe matters for search specifically, and the authors flag it as directional only. They added a minimal retrieval step, and brand familiarity stopped helping.
The real brand ranked 8.50 out of 10 on embedding similarity, and its survival rate fell to 0.0%.
Authority language raised retrieval similarity as well, so it helped at both the retrieval and the generation stage.
The authors also separate three tiers of this content. They are real certifications and published trials, vague phrasing such as "clinically proven", and invented studies and endorsements.
What they tested is the third, as an upper limit rather than as advice.
What the study does not measure is the important part. No search engine is called anywhere in it.
So there is no AI Overviews, no AI Mode and no ChatGPT search result in any of these numbers.
The retrieval probe is a single embedding model with no reranking and no query rewriting. The authors say they do not claim it generalises to commercial systems.
The main experiments are also one product category, one fixed user persona and three closed models. The tie-breaking result was replicated on USB cables and AA batteries.
But the language and competition results were never tested outside skincare.
For those reasons no figure above is restated as a statistic anywhere on this site.
