Your AI Shopping Agent Is Already Playing Favorites—And Your Brand Loses

LLM agents are quietly picking winners and losers by brand name before they even read the fine print. Twelve models, three domains, one ugly finding: where your product comes from matters more than what it does. This is the agentic-era version of Google PageRank, except nobody published the algorithm.

What happened

Researchers tested 12 LLM agent models across three product domains, comparing items that satisfied the same user requirements but came from different sources—same position in the results, same specs on paper. Every single model showed source attribution bias, and the models largely agreed on which sources to favor. The preference is strong enough to override quality: when a worse item (missing one requirement) came from a preferred source, agents selected it about two-thirds of the time—but when the better item came from a dispreferred source, it almost never won. The team tested two mechanisms: first, training that rewards high-quality items can accidentally bake a source in as a shortcut signal; second, hallucination-adjacent behavior kicks in when information is missing and agents fill the gap with preconceptions about the source. Two mitigations showed real effect: supplying the missing information the agent was guessing at, and adding a prompt that explicitly counters source preconceptions—pointing toward prompt-level ranking as a practical lever. Hiding source-identifying information weakened but did not eliminate the bias, and relabeling an item with a preferred source’s name raised its selection rate on its own.

Cold read

This is a controlled lab setup—items curated to satisfy “the same requirements at the same position”—which is not how real e-commerce or hotel search works, where source, price, completeness, and position are all confounded. The paper shows correlation between source labels and selection rates; it does not prove the training-as-shortcut mechanism causally, that’s a proposed route, not a demonstrated one. “Largely agreeing” across 12 models is striking, but 12 models is a small and probably skewed sample—if they’re all fine-tuned on similar RLHF data, shared bias is expected, not surprising. The mitigations (supply missing info, add a counter-prompt) reduce preference but the paper does not quantify how much of the gap closes or whether it survives adversarial real-world prompts. And “source preference” in a benchmark is very different from measurable revenue impact in a live agentic workflow—the translation from lab to P&L is the gap nobody has closed yet.

What it means for you

• Signal maturity: 3/5 — Effect is real and large, but causal mechanism and real-world magnitude are unproven
• Who gets hurt: Challenger brands and D2C operators whose products live on less prominent or newer platforms; if your inventory is on a “dispreferred” source, you lose two-thirds of agent-driven purchases regardless of product quality
• What breaks if this is true: The entire premise of GEO (Generative Engine Optimization) and LLMO—optimizing your content is secondary if the agent discounts your domain before reading it
• Why it might not land: Enterprise buyers and B2B procurement agents typically supply structured, complete data to LLMs, which the paper itself identifies as the main bias reducer; the problem may be mostly consumer-facing
• Watch for: Any major e-commerce platform or OTA (Booking, Expedia, Amazon) publishing agent-interaction logs or third-party audits showing differential selection rates by source domain—that’s the real-world confirmation event

Forecast as of 2026-10-05

By Q3 2027, at least one enterprise AI procurement or comparison-shopping vendor will publish internal data or a case study showing measurable, source-correlated selection asymmetry in live agent transactions—forcing a public response from at least one frontier model provider about bias auditing in agentic settings. If that doesn’t happen, the lab finding stays a lab finding.

Source: Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It — Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song, Yohan Jo. https://arxiv.org/abs/2610.03195v1