When an AI recommends a product, is it judging quality — or just rewarding a brand name it already trusts?

This explores whether LLMs recommending products get swayed by *who* is offering an item more than by *how well* the item fits. The short answer from the corpus is yes, and by a wide margin. Across 12 models and three domains, LLM agents picked an item from a source they favored over a better item from a less-favored source about two-thirds of the time, even when the favored item met fewer of the user's stated requirements Do language models favor sources regardless of item quality?. The same study gives two useful clues about how it works. Hiding the source labels weakens the bias, so the model is reacting to the name itself. And training data that pairs a source with good outcomes is enough to create the preference. The model learns 'this brand means good' as a shortcut, then keeps using it when the evidence in front of it points the other way.

Viewed alongside related findings, source bias is one example of a wider pattern: LLM recommenders carry preferences from pretraining that have nothing to do with the user or the catalog. LLM recommenders show position bias, popularity bias, and fairness bias that come from the language model's training corpus, not from interaction data Where do recommendation biases come from in language models?. The clearest example is GPT-4 recommending *The Shawshank Redemption* across movie datasets with very different popularity patterns, because it is famous on the internet, not because the users in that dataset liked it Where does LLM recommendation bias actually come from?. Source identity works the same way: reputation absorbed during training stands in for judging the item.

The pattern also shows up outside shopping. Six LLMs asked for strategic business advice recommended the same side of every tradeoff across 15,000 simulations. Industry context moved the answers only 11%, while simply reordering the options moved them 19% Do LLMs consistently favor the same strategic choices regardless of context?. When surface cues like ordering or familiar names outweigh the actual details, the model is matching patterns from training rather than weighing quality. A related finding: LLMs fill in 35–49% of their claims about users from priors instead of evidence, and the models that rate themselves as least prone to this do it most Do large language models fabricate user attributes beyond available evidence?. You can't simply ask a model whether it is being swayed.

Humans have a parallel weakness. People prefer search answers with more citations even when the citations are irrelevant, so citation count works as a trust signal that has come loose from what the sources say Do users trust citations more when there are simply more of them?. That makes a worrying loop possible: models favor familiar sources, and users trust answers that look well sourced, so neither side is checking item quality.

The corpus hints at fixes through design rather than debiasing. One approach keeps LLMs out of the ranking step altogether: use them to enrich item descriptions, then let a traditional recommender rank, which beats asking the LLM to recommend directly Does LLM input augmentation beat direct LLM recommendation?. The other approach follows from the label-hiding result: strip or neutralize source identity before the model compares items. The gap is that none of these notes measures how much quality difference it takes to overcome a source preference. That threshold is the open question if you want to go further.