How do shopping agents change consumer choice?

Recommendation, delegation and the measurement of consumer benefit

This retrospective review considers source versions available by 2 February 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: agentic commerce; consumer choice; recommendation; delegation; market concentration

1. Research question and scope

An automated shopping system may retrieve a product, recommend it, place it in a basket or complete its purchase. Each operation represents a different degree of delegation. The research question is how algorithmic mediation changes the choice process and what evidence can support claims about the resulting consumer benefit. A recommendation that receives attention, a product selected by a simulated agent and a purchase that satisfies a shopper are distinct outcomes.

The older literature on decision aids provides a useful starting point because automated assistance in commerce predates conversational models. Häubl and Trifts separated screening a large assortment from comparing the surviving options. Their study examined recommendation and comparison tools designed for those different tasks [1]. Contemporary agents can combine them in one interface, but combining operations does not remove their separate objectives. A filter can exclude a suitable product before a comparison ever begins.

This review examines product selection, market-level concentration and consumer reliance. It excludes forecasts of the size of agent-mediated commerce and descriptions of particular checkout releases. Those change quickly and do not establish how reliably a system chooses. The relevant evidence includes controlled consumer studies, analytical work on recommender systems and a recent simulation of model-based shopping agents. Their results answer different questions and are compared at that level.

2. Review method

Targeted searches on 21 October 2026 used the terms shopping agents, recommendation agents, consumer choice, sales diversity, algorithm aversion and trust in automation. Primary publisher pages, author manuscripts and university records were consulted. Twelve references were selected to connect the choice operation with its measurement. Selection was purposive; no comprehensive database review, independent duplicate screening or quantitative meta-analysis was performed.

The extraction distinguishes the actor being observed, the choice set, the intervention and the outcome. A human participant choosing with an aid differs from an autonomous model selecting from an artificial catalogue. Results are not combined into a single estimate of agent effectiveness. Several historical studies were inspected through their abstracts and publication records, so the discussion stays within those summaries. Access categories appear after the conclusion.

Causal identification is considered separately from realism. Randomising an attribute in a sandbox can isolate its influence on that agent under those conditions. It does not establish the prevalence of the effect in a live market. Conversely, an observed association between recommendation exposure and sales can reflect the platform's prior selection of users or products. A reliable interpretation needs both the comparison and the setting.

3. What is delegated?

3.1 Preference specification

A shopper can express a hard constraint, such as a maximum price, or a negotiable preference. An agent that treats these as interchangeable may return an attractive option that fails the assignment. Evaluation therefore begins with whether the selected product belongs to the feasible set. Preference satisfaction within that set is a second question. Conflating them makes a system appear successful when it has traded away a requirement the shopper did not authorise it to relax.

Bettman, Luce and Payne's account of constructive choice emphasises how decision processes respond to the task and its information environment [3]. This provides a conceptual reason to inspect the elicitation interface. A short conversational request cannot be assumed to reveal a complete, stable utility function. If the shopper revises a preference after seeing alternatives, that revision may reflect learning about the category rather than inconsistent behaviour.

3.2 Recommendation and execution

Adomavicius and Tuzhilin distinguish content-based, collaborative and hybrid recommendation methods [2]. Their taxonomy concerns how recommendations are produced. It does not itself describe an agent's authority to spend money or accept substitutes. A system can use sophisticated ranking while having no purchase permission; a purchasing agent may use a relatively simple ranking rule. Capability, objective and authority need separate description.

The difference affects error accounting. An irrelevant suggestion can be ignored before purchase. An executed order creates a downstream correction problem, potentially involving returns, delivery and payment. Measuring only whether the browser reached a confirmation page captures task completion while leaving preference compliance unresolved. A useful assessment records both the completed action and whether it matched the instruction that justified the action.

4. How choice architecture changes the evidence

4.1 Screening versus comparison

A recommendation system determines which products receive further consideration. A comparison interface determines how visible attributes are weighed among those products. Häubl and Trifts explicitly examined tools for these separate stages [1]. For an agent, retrieval coverage must therefore be distinguished from ranking accuracy. A ranking can be excellent within a restricted catalogue while omitting the shopper's best available option.

This distinction changes the denominator of a success rate. If evaluation includes only products already returned by the system, it cannot identify failures of discovery. If it includes every product in an external catalogue, the assessor must establish which were genuinely available and eligible at the time. Availability, geography and delivery constraints are properties of the tested choice set, not interchangeable background details.

4.2 Position and presentation

The ACES preprint uses a controlled marketplace to investigate model-based product selection. Allouah and colleagues report choice concentration, model-dependent preferences, position effects and changes following model updates. Sponsored tags were penalised while platform endorsements were rewarded in their tested agents [5]. These findings are specific to the simulation and model versions. They do not establish a durable recipe for influencing every live shopping assistant.

A position effect is analytically different from a preference for a product attribute. When product order is randomised, the same item can be evaluated at different positions. Without that variation, an apparent preference for a prominent item could arise because the platform placed stronger products there. The distinction is important when interpreting natural shopping logs, where ranking and product quality usually move together.

The manuscript's instruction tests include budget, colour and named-brand requirements. Separate rationality tests hold product attributes constant except for price or rating, so a preferred option can be defined from the manipulated attribute [5]. These controlled comparisons have a clearer answer criterion than an open request for a product that feels suitable. Performance on the simpler task should retain that distinction when reported.

4.3 Selection and welfare

A seller's increased selection share does not reveal whether the consumer obtained a better match. It may reflect improved information, a persuasive description or sensitivity to an irrelevant display feature. These explanations require different measurements. Consumer benefit needs a stated criterion such as compliance with requirements, realised satisfaction or a defensible preference-based comparison. Selection share alone supplies none of them.

Shmueli's distinction between predictive and explanatory modelling is relevant here [12]. Predicting which product an agent will select and explaining why it selects it are separate analytical aims. A model with strong predictive accuracy can exploit correlations that do not survive an interface change. An intervention that changes selection establishes an effect within its comparison, but does not automatically reveal every intermediate reasoning step.

5. Individual discovery and market concentration

Fleder and Hosanagar's analytical and simulation work shows how some recommendation designs can increase variety within an individual's purchases while reducing diversity across the market. Several consumers can discover something new and still converge on the same products [4]. This result concerns the structure of recommendation and aggregate demand; it is not an empirical estimate of current language-model shopping behaviour.

The two diversity measures have different units. Individual variety describes one person's set of selections. Aggregate concentration describes how purchases are distributed across products or sellers. A market can improve on one and worsen on the other. Reporting a single diversity score without its unit conceals this distinction and makes apparently contradictory results difficult to interpret.

Historical purchase data can also be part of both the input and the outcome. When exposure follows prior popularity, subsequent sales reflect an interaction between existing demand and the platform's allocation of attention. A catalogue with sparse data on new products poses a different measurement problem from a mature catalogue with repeated observations. Neither the absence of prior sales nor the abundance of them is a direct measure of consumer-product fit.

Salganik, Dodds and Watts used an artificial music market to study social influence, inequality and unpredictability [6]. The study concerns human responses to information about others' choices. It provides evidence that market outcomes can depend on the information environment, but does not identify the behaviour of contemporary purchasing agents. The comparison is useful because it keeps the feedback process visible without treating every form of automated mediation as the same mechanism.

Muchnik, Aral and Taylor's randomised experiment found that prior ratings influenced subsequent rating behaviour, with asymmetric positive and negative effects [7]. Ratings can therefore be outcomes shaped by earlier social signals as well as inputs to later decisions. An agent's use of a rating does not make that rating an independent measure of quality. Its origin, exposure history and relation to the purchase task remain relevant.

6. Reliance, familiarity and permission

6.1 Adoption is not calibrated reliance

Komiak and Benbasat examined perceived personalisation and familiarity in relation to cognitive and emotional trust and the intention to adopt a recommendation agent [8]. Their framework separates using an agent as an aid from accepting greater delegation. The relevant outcome is adoption intention within that study, not a measured improvement in every later purchase. A familiar interface may feel easier to use without supplying better product evidence.

Lee and See's review treats appropriate reliance as the objective of trust design [9]. That is a more specific criterion than maximising reported trust. A shopper who rejects a good suggestion and one who accepts an unsuitable purchase both create mismatches between reliance and capability. A system intended to increase delegation must be evaluated for both errors, with particular attention to the consequences of executed actions.

6.2 Aversion and appreciation

Dietvorst and colleagues documented avoidance of algorithms after people observed errors, even when an algorithm could remain useful relative to the alternative [10]. Logg and colleagues found algorithm appreciation in other judgement settings [11]. The findings discourage a universal assumption that people either distrust or prefer automated advice. Their tasks and comparisons differ, and neither alone predicts how consumers will delegate shopping decisions.

For commerce evaluation, willingness to delegate is a behavioural outcome requiring its own measurement. Stated willingness in a survey, choosing to use an assistant and allowing an assistant to purchase are progressively different observations. An experimental preference for algorithmic advice cannot be converted into a forecast of transaction share. The amount, category and reversibility of the purchase can change the meaning of the delegation.

7. What a defensible evaluation must measure

7.1 An explicit reference task

An evaluation needs the shopper's instruction as a reference object. If it is altered after observing the agent's output, the test can reward choices that were never part of the original assignment. Hard constraints should remain identifiable in the record. When several options satisfy them, evaluation must specify how competing preferences are compared or report that there is no single demonstrably best option.

A task with one dominant product supports a relatively clear compliance test. A task involving taste, uncertainty and several acceptable products does not. Calling both tasks shopping accuracy hides the difference. For the latter, the assessor may need a set of acceptable choices, a preference elicitation procedure or later consumer feedback. Each method introduces assumptions that belong in the interpretation.

7.2 Robustness across the choice environment

The same instruction can be tested against changed product order, changed catalogue composition and a refreshed model version. These are distinct robustness checks. Holding all but one factor constant helps identify sensitivity; changing them together can assess a realistic deployment but leaves the source of a performance change less clear. A single successful purchase is evidence of an instance, not an estimate of reliability.

Repeated calls to an unchanged agent also need careful interpretation. Variation among runs can reveal instability, but repeated runs are not necessarily independent observations about consumer preferences. They may reuse the same model, prompt and catalogue. The number of runs must remain distinct from the number of consumers or commercial contexts represented in the test.

7.3 Outcomes after checkout

Order completion is an intermediate event when the research question concerns consumer benefit. Delivery, product suitability, returns and customer satisfaction occur later. A transaction can satisfy the initial textual constraints while exposing an omitted requirement, such as compatibility. Conversely, a cancelled purchase can represent useful consumer protection if the catalogue contains no feasible option. A completion-only metric would classify that protection as failure.

Commercial analysis has another boundary: substitution. A product bought through an agent may replace a purchase that would have happened through another route. Attributed revenue from the new interface is therefore different from incremental revenue. The research reviewed here does not estimate that difference for a live agentic-commerce deployment. Any economic claim must identify its own counterfactual rather than inherit one from a task-completion benchmark.

8. Implications for interpretation of marketing claims

A claim that structured product information improves machine readability concerns information availability. A claim that it increases selection concerns the agent's response. A claim that it benefits consumers or grows demand adds further outcomes. These claims can all be investigated, but evidence for the first cannot be silently promoted into evidence for the last. Their denominators and comparisons differ.

Product information can also be correct in isolation and incomplete for the assignment. A price may omit delivery; a feature description may omit compatibility; stock may refer to another location. Evaluating factual accuracy at field level is therefore insufficient for evaluating task suitability. The relevant question is whether the information needed for the shopper's stated constraints was available, current and correctly used.

The evidence supports continuous attention to model and interface dependence. It does not establish that all existing brand preferences disappear when an agent enters the process, or that they necessarily become more important. A shopper can specify a brand as a constraint, leave it open or revise it during comparison. The reviewed experiments do not supply a general distribution over those behaviours in every category.

8.1 Interpreting a null result

A comparison that detects no difference between two agents does not establish that their choices are equivalent. The tested catalogue may contain little meaningful variation, or the measurement may be too coarse to reveal a difference. Equivalence requires a defined tolerance and an appropriate design. Reporting the task distribution and uncertainty is more informative than describing both systems as equally capable on the basis of one non-significant comparison.

9. Limitations and conclusion

This is a selective narrative review combining historical consumer-choice research with a recent model-agent simulation. Much of the foundational literature was consulted at abstract or record level. The ACES manuscript is a preprint, and its sandbox does not reproduce every feature of a real marketplace. No common effect size, risk score or market-adoption forecast is estimated.

The strongest general conclusion is a measurement requirement: delegation changes the location of the choice, but does not remove the need to specify whose objective is being served. Product retrieval, ranking, execution and realised consumer benefit need separate assessment. Evidence also shows that individual discovery can coexist with aggregate concentration and that reactions to automated advice depend on the setting. Claims about the future of commerce require direct observations of the relevant agents, consumers and market outcomes.

Source access and appraisal register

Sources were consulted on 21 October 2026. Access categories describe the material inspected for this narrative review. An abstract supports only its reported summary; it does not establish a check of the complete analysis or supplementary material. No independent replication or formal risk-of-bias scoring was undertaken.

References