LLM Agents Exhibit Strong Bias Toward Specific Content Sources

Testing 12 agent models reveals that source preference often outweighs item quality; agents select an item from a preferred source even if it satisfies fewer user requirements. This bias can be mitigated by hiding source identifiers or relabeling items, which significantly alters selection behavior.