AI shopping agents still show biased favorites based on list position, even with no pictures in the mix — just text.

This asks whether AI shopping agents keep favoring products by where they sit in a list even when nothing visual is shown to them, just plain text. The short answer is yes, but each model has its own pattern. The ACES study removed the visual storefront entirely and told agents outright to ignore product position. Strong position biases remained anyway, and each model had its own signature lean Do AI shopping agents show persistent position bias in text interfaces?. So the bias isn't a leftover from screen layout, the way a human's eye drifts to the top of a page. It belongs to the model itself. It also can come back after each model update. A seller can't tune against it once and move on, because the target keeps shifting.

The pattern fits a wider picture of how these agents make choices. When researchers gave 17 models and a group of humans the same binary product choices, the agents shifted their picks far more sharply in response to price, rating, and nudge changes. People made modest, probabilistic adjustments. The agents collapsed into near-rule-like behavior Do shopping agents respond more strongly to price and rating cues?. Position bias looks like one more case of the same thing: a small, irrelevant cue turned into a decisive one. Another paper suggests where these habits come from. LLM agents reproduce human-style decision biases without having the attention or memory limits that cause those biases in people. That points to patterns learned from training data and reward signals, not shared cognitive constraints Where do LLM choice biases actually come from?. In other words, the models may have learned "first is best" from us, without needing our reasons for it.

Why doesn't telling the agent to ignore position work? Research on persona prompts offers a parallel. Instructions change what a model says on the surface, but group-level bias underneath stays the same Can persona prompts actually reduce bias in language models?. Prompting works at the output level, and position preference seems to sit deeper. One proposed fix works through training instead of instructions: teach models to give the same answer whether or not a prompt contains irrelevant variations, using their own clean responses as the target Can models learn to ignore irrelevant prompt changes?. That's a plausible route for teaching agents to treat list order as noise, though the corpus doesn't show it being tested on shopping agents.

The less obvious point is that agent biases depend on the setup around them. A study of eight commercial LLMs found agents resisted marketing manipulation when every product detail was visible. They fell back on simple price-based shortcuts only when looking up attributes was costly and the user's goal was vague When do pricing cues actually mislead AI shopping agents?. That raises a fair question: is position bias also worse when agents can't cheaply compare everything? The corpus doesn't answer that yet. Humans have a version of the same problem. Eye-tracking shows AI Overviews now take the top-of-page attention that once went to the first search result Where do searchers look when AI Overviews appear?. Whether the shopper is a person or an agent, the top spot is worth a lot, and agent shoppers now give sellers a new and model-specific top spot to compete for.