The commercial push toward autonomous shopping hinges on a fragile assumption: that AI agents can evaluate complex product catalogs rationally and execute financial transactions without going off the rails. As engineering teams wire foundation models into multi-step execution loops, delegated e-commerce seems like the next logical frontier. Yet empirical benchmarks reveal that modern architectures remain far too brittle to be trusted with corporate or consumer credit cards.
Researchers at the Wharton School at the University of Pennsylvania exposed this volatility using the ACES simulator (Agentic e-Commerce Simulator). Testing six frontier and lightweight models on a visual selection task—choosing a fitness watch from a fixed product grid via screenshots—the team evaluated how minor information retrieval shifts alter agent behavior. The setup let models inspect page visual layouts, ingest external product recommendations, and commit to a purchase. The findings are sobering: even microscopic perturbations in context flip purchase decisions entirely.
External Sources and Order Sensitivity
Baseline models displayed distinct native biases across the catalog even before ingesting third-party inputs. Introducing a single external source completely scrambled those preferences. The Wharton researchers tested three realistic external artifacts: a Reddit thread backing the Garmin Forerunner 55, a Wirecutter review endorsing the Fitbit Inspire 3, and a Strategist article favoring the WHOOP 5.0. Wirecutter exerted an overwhelming gravitational pull, boosting the probability of selecting the Fitbit Inspire 3 by up to 90 to 99 percentage points across several leading architectures compared to baseline controls.
Stacking additional reference materials failed to stabilize the models, amplifying variance instead. The sequence in which identical materials were ingested also dictated the final outcome.
Presentation order is itself a driver of product selection, causing identical content to yield conflicting purchase decisions.
When all three reviews were ingested in varying sequences, certain lightweight models exhibited wild volatility, swinging selection probabilities for the target product by over 50 percentage points solely due to input order. The packaging format introduced further instability: delivering external citations in a single batch versus streaming them sequentially produced drastic divergence in final recommendations, underscoring how deeply token delivery mechanics warp agentic reasoning.
Memory Snippets and Rational Trade-offs
To test baseline economic rationality, the Wharton team staged an asymmetrical catalog where one option dominated every metric: an Alexa-enabled smartwatch priced at $29.99 with a 5.0 rating across 430 reviews, pitted against alternatives costing upwards of $359 with sparse reviews. The researchers then seeded trivial episodic memory fragments into the prompt, such as the statement "I love hiking!"
These brief memory artifacts caused multiple models to instantly abandon the objectively dominant budget option in favor of overpriced alternatives, confusing superficial semantic alignment with genuine utility.
These findings delineate the real technical boundary of agentic commerce. Models in the ACES benchmark act as hyper-sensitive context aggregators rather than deterministic evaluators, rendering autonomous payments reckless without deterministic guardrails and strict human-in-the-loop verification. For marketers betting on Generative Engine Optimization (GEO) and Agent Engine Optimization (AEO), the takeaway is equally clear: agent decision-making is currently driven by sequence noise and context vulnerabilities, not systematic evaluation.