Do AI shopping agents show persistent position bias in text interfaces?

The paper behind ACES (Agentic e-CommercE Simulator) reports that AI shopping agents "exhibit strong position biases—varying across providers and model versions, and persisting even in text-only 'headless' interfaces—undermining any universal notion of a 'top' rank." The bias survives two interventions that would be expected to weaken a merely superficial effect: "explicitly prompting an agent to 'ignore position' does not significantly mitigate position effects," and the bias "persist[s] even in 'headless' environments" where the agent receives only a ranked JSON list of product data with no images at all.

ACES isolates the choice step of shopping rather than full web navigation: a vision-language-model agent opens a mock storefront, screenshots a fixed 2×4 grid of eight products per category, and declares a single purchase (the paper's "Veni–Vidi–Emi" flow). Because the mock-app's layout, attributes, and badges are all programmatically controlled, the authors can run randomized trials that isolate position as a causal factor rather than a correlate of product quality. They replicated the position-bias finding on three focal models in the headless, image-free API setting and "once again found similar biases." The authors conclude that "position and other effects are not driven (only) by visual processing from parsing a webpage, and are likely to persist under standardized agent protocols" — in their words, the effects are "deep-seated characteristics of AI models" rather than "superficial artifacts of the interface used."

This sits alongside Does a single benchmark score actually predict agent readiness?: ACES shows the same split play out inside a single capability area. The paper separately tests "basic economic rationality" (does the agent pick the one product that is objectively dominant?) and finds agents have "improved over time to achieve near-perfect consistency" on that axis — while position bias, a different axis, "remain[s] fundamentally prone to strong model dependence" even as rationality improves. A model can pass a rationality check and still be a reliably exploitable chooser. The finding also gives empirical weight to Will agents compete for attention just like users do?: if position is a stable, model-specific lever on agent choice, it is exactly the kind of signal an "agent-facing attention economy" would be built to target.

What the excerpt does not establish is how long any given model's bias signature lasts or whether it generalizes beyond ACES's eight product categories and eight-item assortments; the paper itself notes that "specific empirical coefficients... capture a moment in technological time" and that a model transition (Gemini 2.5 Flash Preview to its official release) "inverted the model's position bias from the bottom row to the top row." The implication the authors draw, and the evidence supports, is not that any particular rank is permanently favored, but that some model-specific positional bias is a structural feature likely to recur after every model update — which is why they frame continuous auditing, not a one-time bias check, as the necessary response.