In August 2026, Wharton Generative AI Labs released “Agentic Shopping is Complicated and Contingent,” asking whether an AI shopper’s product choice can be predicted or reliably steered with marketing. Across roughly 26,000 screenshot-based runs in the ACES mock store, six current models picked a fitness watch from an eight-product grid. Choices were stable when the agent saw only that grid. They became unstable once the researchers added a single review screenshot, competing sources and their order, a one-line user “memory,” or a change in how retrieval was bundled. Ethan Mollick summarized the result as a flat no.
That claim matters because agentic commerce is being treated as a new ranking and advertising surface. If the buyer is a model, sellers want a GEO playbook; platforms want to know which levers still work; builders want to know whether a shopping agent will buy the same item tomorrow. Independent work using ACES, MIT’s ABxLab, and a McGill hotel-search experiment all find that shopping agents are highly sensitive to context. They do not all find that sensitivity unsteerable. The useful conclusion sits between those claims.
What we checked
- Can marketers consistently predict or, using marketing, influence what a shopping agent chooses? partly confirmed. Shopping agents are highly influenceable: reviews, source order, memories, tool-call packaging, listing text, ratings, and nudges all move choices, often by tens of percentage points. The mapping from any one intervention to a product is model-specific and path-dependent, so a seller cannot count on a single marketing playbook working across agents. [1][4][2][5]
- Do small differences in viewing order of pages and injected memories change AI shopping preferences in unpredictable ways? partly confirmed. Wharton showed that shuffling the same three review screenshots, or inserting a one-line memory such as “I love hiking!”, can swing selection by tens of points, including away from an objectively cheaper, higher-rated watch. The swings are not uniformly chaotic: Claude Haiku 4.5 stayed stable on order, and Gemini 3.5 Flash largely ignored the hiking memory. [1][4]
- How stable are page-order and memory effects across models, tasks, and real shopping environments? partly confirmed. Order effects appear in fitness-watch grids, a WebArena-style shopping site, and a 100-hotel search task, but sign, size, and whether they survive to final choice depend on model, reasoning effort, and interface. Memory-like user profiles also reweight choice in ABxLab. No opened source measured live checkout on a production retailer. [1][2][5][7]
- Do clean-page shopping benchmarks underestimate real-world agent variability? confirmed. On a bare product grid, each Wharton model had a stable favorite. Adding prior sources, order, memory, or retrieval packaging made recommendations less consistent. The authors state that benchmarks that test only the purchase interface underestimate real-world susceptibility. [1]
- Can sellers still move agent market share by changing listing content they control? partly confirmed. In the ACES seller-response experiments, one-shot description edits produced statistically significant average share gains for five of six buyer models, including +14.89 percentage points for GPT-5.1, but 67 percent of category–model pairs showed no significant change and some edits reduced share. [2]
What we found
- Wharton’s public technical report, dated 27 August 2026, used ACES, the Agentic e-Commerce Simulator. The agent acted as a personal shopper with no stated requirements, saw an eight-product fitness-watch grid as a screenshot, and chose a product. The team tested Claude Haiku 4.5, Claude Opus 4.8, GPT-5 Mini, GPT-5.5, Gemini 3.1 Flash Lite, and Gemini 3.5 Flash, with 200 runs per condition at temperature 1.0. Study 5, on tool-call packaging, dropped the Anthropic models. [1][2][3]
- With only the grid, each model settled on a different favorite: Claude Opus 4.8 usually picked the Garmin Forerunner 55, GPT-5.5 the Garmin Vivoactive 5, and Gemini 3.5 Flash the Fitbit Versa 4. A single source screenshot often overrode that default. Gemini 3.5 Flash chose Wirecutter’s Fitbit Inspire 3 in essentially every run, a shift of +99 percentage points; THE DECODER reports a +90 point Wirecutter shift for Claude Opus 4.8. Multiple sources did not average: Wirecutter dominated both Claude models, Reddit held the most sway over GPT-5.5, and the Reddit plus Strategist pairing, which excluded Wirecutter, moved no model. [1][4]
- Shuffling the same three sources changed outcomes for some models and not others. Both Claude models recommended the Fitbit Inspire 3 regardless of order, and GPT-5.5 stayed loyal to the Garmin Vivoactive 5. Gemini 3.1 Flash Lite hinged on which source it saw first: when Wirecutter led, it almost never switched to Wirecutter’s pick; when Reddit led, it switched in more than half of runs. THE DECODER reports that model’s Fitbit Inspire 3 lift swinging between 2 and 56 percentage points above control, while Claude Haiku 4.5 stayed at 41 to 42 points. [1][4]
- In the memory study, one listing was objectively best: a Smart Watch with Alexa at $29.99 with a 5.0 rating, while every other product cost at least $359 with worse reviews. Injecting “I love hiking!” sent Claude Opus 4.8 to the hiking-friendly Garmin Vivoactive 5 in three-quarters of runs; THE DECODER reports +75, +37, and +36 point Garmin shifts for Opus 4.8, GPT-5.5, and Gemini 3.1 Flash Lite. Only Gemini 3.5 Flash kept picking the cheap, high-rated watch (Wharton reports p = .784; THE DECODER reports 86 to 92 percent). Delivery format also mattered: GPT-5.5 shifted +53 points toward Wirecutter’s pick when the three sources arrived bundled, versus +6 points when they arrived one by one. [1][4]
- The ACES paper that supplied Wharton’s simulator, later revised through 15 December 2025, shows that even on a clean grid agents are not neutral shoppers. Position effects are large and model-specific, including opposite column preferences between GPT-4.1 and GPT-5.1; sponsored tags reduce selection while an “Overall Pick” badge raises it; and modest seller-side description edits moved average share by +3.66 to +14.89 percentage points for five of six buyer models, with no significant average move for Gemini 3.0 Pro Preview. Those listing edits worked in only 33 percent of category–model pairs and sometimes reduced share. Model upgrades also reshuffled market shares, for example sending Fitbit Inspire from 45 percent under Claude Sonnet 4 to 77 percent under Claude Opus 4.5, and from 25 percent under GPT-4.1 to 6 percent under GPT-5.1. [2]
- Independent shopping experiments qualify the “unpredictable” headline. MIT’s ABxLab, across more than 80,000 trials and 17 models in a WebArena-style store, found agents far more sensitive than a 30-person human baseline: human order effects were 4 percentage points, while GPT-4.1 Nano showed an +88.8 point first-viewed bias and Claude 3.5 Haiku a −35.4 point penalty. A “Wirecutter’s top pick” nudge was among the strongest cues, and explicit user profiles acted like categorical switches rather than mild weights. A McGill hotel-search study of 7,000 sessions found weaker, U-shaped rank effects on inspection than in human field data, with displayed attributes dominating final choice and higher reasoning effort reducing the middle-rank penalty for every model that had one. [5][6][7]
What it means
Read the Wharton “Nope” as a claim about search-path marketing the seller does not control, not as proof that agents cannot be influenced. Reviews, order, memory, badges, and listing copy all move demand; what fails is a stable, cross-model recipe. Builders should evaluate shopping agents in messy information environments: prior pages, competing sources, injected memory, and the actual tool-call packaging used in production. Pin model version, temperature, and reasoning effort, because those settings change both source-order and position effects. Treat stored memory as a preference channel and an attack surface; require a user confirm when memory would override price, rating, or budget dominance. Do not ship GEO from a single-model bake-off. Keep a human confirm step on spend until the agent’s choice profile is measured on your catalog, your retrieval stack, and the models you actually serve.
Still open
- Do order and memory effects survive live checkout, real reviews, ads, and multi-page navigation on production retailers, or only in screenshot and mock-store labs?
- Wharton ran temperature 1.0; ACES often used low reasoning effort. How much of the instability is sampling and decode settings rather than the model prior?
- Does the higher reasoning effort that reduced hotel position bias also damp Wharton’s source-order and memory effects?
- How fast do modal products and position maps drift after provider model updates, given ACES already saw large share reshuffles between generations?