From Product Search to Preference Articulation: The Economics of Agentic Commerce

Generative AI is moving digital commerce from browsing toward agentic search, in which consumers delegate product discovery to AI agents. We model manual search, which accurately evaluates a limited product set, against agentic search, which screens a broad catalog using noisy representations of preferences and products. Preference complexity measures the number of satisfaction-relevant dimensions that are difficult to articulate before search but readily evaluated upon inspection. Consumers hold a finite attention budget and choose both a search regime and the intensity with which to use it: the number of products inspected under manual search and the depth of preference refinement under agentic search. We obtain three main findings. First, as preference complexity grows, manual search eventually collapses: inspection ceases, mismatch returns to its no-search level, and platform revenue falls to zero. Agentic search, by contrast, avoids a finite complexity collapse. Complexity raises the marginal value of refinement, so once refinement becomes worthwhile, it remains worthwhile at every higher level. Articulation effort and mismatch may increase, but mismatch remains below the no-search level and revenue remains positive. Second, the platform ranks search regimes by conversion revenue, whereas the consumer also bears search expenditure. When manual inspection is sufficiently inexpensive, agentic search becomes revenue-superior before it becomes privately optimal for the consumer. Over the intervening range, the consumer rationally continues manual search, creating an adoption lag. Third, the platform may provide lower AI fidelity to consumers with larger attention budgets because they can compensate through additional refinement, producing an inverted fidelity allocation. Agentic commerce thus shifts scarcity from product inspection to preference articulation, making consumers’ willingness and ability to interact central to voluntary use and platform fidelity design.

Introduction. Agentic commerce allows consumers to delegate product search, comparison, and, in some settings, transaction execution to AI agents. This shift from browsing toward delegated action is already underway. Amazon’s Alexa for Shopping supports conversational product discovery and comparison, price monitoring, and automated purchasing, including checkout on third-party brand websites through Buy for Me. Walmart’s Sparky assists consumers with product discovery, comparison, and personalized recommendations within its retail interface. Google and OpenAI have introduced conversational shopping tools that extend from product discovery and comparison toward in-interface arXiv:2608.08395v1 [econ.TH] 9 Aug 2026 2 Author: The Economics of Agentic Commerce checkout, supported by new protocols connecting AI agents with merchants. Payment networks are building the complementary transaction infrastructure: Mastercard Agent Pay and Visa Intelligent Commerce authenticate agents, encode consumer authorization, and support agent-initiated payments. Together, these offerings span the shopping journey from product discovery and comparison to checkout and payment. Throughout the paper, we use agentic commerce for this broader paradigm and agentic search for its search-and-matching stage. Our analysis centers on the latter because the value of downstream transactions depends on how accurately the agent represents what the consumer wants.

Despite these investments, available evidence points to a continuum between AI assistance and more extensive delegation. Adobe reports that 39 percent of U.S. consumers had used generative AI for online shopping by early 2025, primarily for product research, recommendations, and deal discovery (Adobe 2025). More recent surveys indicate greater caution as AI assumes more control. Worldpay finds that 44 percent of U.S. consumers would allow an AI assistant to browse for them, whereas only 6 percent would relinquish complete control over the purchase (Worldpay 2025). Contentsquare similarly reports that 30 percent of consumers would allow an AI agent to complete a purchase on their behalf (Contentsquare 2025). Although these estimates are not directly comparable, together they suggest that consumers increasingly use AI to assist parts of the shopping process, but are more cautious about delegating the entire search process and final transaction to an AI agent.

Consider a consumer buying home-office furniture online, such as an ergonomic desk chair or a modular standing desk. Some attributes are easy to specify in advance: price, dimensions, color, material, delivery time, and brand. Other satisfaction-relevant features are harder to articulate: whether the chair will feel supportive during long work sessions, whether the design fits the room, whether the product looks professional on video calls, and whether its dimensions translate well into the available space. These considerations need not be unknown to the consumer, but they can be difficult to encode ex ante as search criteria and easier to assess when she evaluates concrete alternatives. We use preference complexity to describe the extent to which a satisfactory match depends on many such dimensions. As complexity rises, a close match becomes harder to find through limited manual browsing, while communicating to an AI agent what constitutes a close match also becomes more demanding. Agentic search therefore changes rather than removes the consumer’s search burden: it relaxes the scope constraint of product-by-product inspection but shifts scarce attention toward preference articulation.

These observations lead to one governing question: how does agentic search shift the burden of preference complexity from product discovery to preference articulation? We examine two strategic consequences of this shift. When does the platform’s revenue ranking of agentic search diverge from Author: The Economics of Agentic Commerce 3 the consumer’s voluntary-use decision? And how does the platform design fidelity when consumers differ in the attention they can devote to refining their preferences?

In practice, shopping journeys are often hybrid: consumers may use AI for research, comparison, or initial screening while retaining control over the final product choice and purchase. To isolate the tradeoff created by search delegation, we abstract from these mixed modes and compare two canonical regimes—manual search and agentic search. Manual search permits direct and accurate evaluation of a limited set of products; agentic search screens a broad catalog at negligible marginal cost but relies on imperfect representations of the consumer’s preferences and product attributes.

Related work. Our work connects three research streams: consumer search, AI-mediated search and interaction, and the economics of AI agents. Consumer-search models provide the direct-inspection benchmark; research on AI-mediated search and human–AI interaction identifies the roles of representation error, preference articulation, and the cognitive cost of using AI; and the agentic-economy literature situates delegation within broader questions of market design and alignment. We connect these streams by making refinement depth a consumer choice subject to interaction friction and an attention budget. This consumer-side decision shapes search performance and voluntary adoption; conditional on agentic participation, it also shapes the platform’s fidelity allocation across observable attention segments.

2.1. Consumer Search The foundational search literature models consumers sampling from product distributions to locate acceptable matches (e.g., Stigler 1961, McCall 1970). Weitzman (1979) derives optimal stopping rules under heterogeneous option values, and Anderson and Renault (1999) links search intensity to spatial competition and preference heterogeneity. In these models, inspecting an item reveals its relevant attributes, so search effort primarily expands the set of products evaluated. Related work allows the information environment itself to evolve with search (Urgun and Yariv 2025, Malladi 2023). Our model instead holds the consumer’s latent preferences fixed and lets her choose how much effort to devote to refining the intermediary’s representation of them. Search effort therefore operates through preference articulation rather than additional product inspection or preference discovery.

2.2. AI-Mediated Search and Conversational Commerce Although Liang (2026) studies matching among people rather than product search, it provides our closest spatial benchmark. Liang compares direct encounters with search over an arbitrarily large pool of noisy AI representations and shows that representation error can overturn the scope advantage of AI-mediated search, especially in high-dimensional spaces. We retain this noisy high-dimensional geometry but allow the consumer to refine the agent’s initially noisy representation of her preferences through interaction. The consumer therefore trades the cost of additional interaction and the attention it consumes against a more accurate representation. This demand-side choice allows us to analyze how refinement shapes match performance as preference complexity grows, when consumers voluntarily use agentic search, and how the platform allocates fidelity across consumer segments.

Method. We study a product-search platform that helps consumers locate suitable products from a large catalog. The platform may be interpreted as an online marketplace, a travel-booking platform, or a 1 Related research examines GenAI’s economic effects in labor markets (Ide and Talamas 2025, Eloundou et al. 2024), marketing (Goli and Singh 2024, Wang et al. 2026, 2025), healthcare (Adida and Dai 2025, Dai and Singh 2025), pricing delegation (Keppo et al. 2026), and human–AI collaboration (Bastani and Cachon 2025, Dai and Taylor 2025, Siderius et al. 2026).

Author: The Economics of Agentic Commerce 7 retail intermediary. Its economic role is to reduce search frictions between consumers and products.

Consumers arrive with latent, imperfectly articulated preferences, choose how to search, and purchase if the resulting match is sufficiently good. The platform earns revenue from successful transactions.

The central technological distinction is how search is mediated. A consumer can search manually by inspecting a finite number of listings herself, or she can delegate the search process to an AI agent. These regimes differ along a scope–fidelity tradeoff. Manual search is accurate but narrow: the consumer evaluates each inspected product with perfect fidelity but can inspect only a few products. Agentic search is broad but imperfect: the AI can screen the full catalog but must rely on noisy representations of the consumer’s preferences and product characteristics, as formalized in Section 3.3. Throughout, scope denotes the breadth of the catalog evaluated, whereas fidelity denotes the accuracy with which consumer–product fit is evaluated. We formalize this tradeoff using a Lancasterian framework in which utility derives from product characteristics (Lancaster 1966), building on the spatial model of Liang (2026) for AI-mediated matching.

3.1. Products, Preferences, and Complexity Consider a platform with a large catalog of differentiated products. Each category is characterized by preference complexity k, defined as the number of satisfaction-relevant dimensions that are difficult to articulate or query ex ante but readily evaluated through rich listing inspection. We treat easily measured and communicated characteristics—such as price, brand, size, and battery capacity—as resolved before search. The model focuses on latent features that consumers can assess from photographs, specifications, maps, and review text but cannot readily encode in an initial query. Low-k categories involve few such dimensions; high-k categories involve many that matter jointly. Two examples illustrate the range:

• Low-k. Choosing a phone case. Price, brand, and device compatibility are easy to specify in advance; only a few dimensions—the exact shade, how the case looks on the device, how bulky it is in hand as conveyed by photographs and reviewer comments—are difficult to articulate beforehand, yet are readily evaluated once the consumer inspects a listing.

• High-k. Booking a hotel. Many hard-to-articulate features matter jointly—room feel, neighborhood character, noise, view, layout, and comfort as reported by reviewers. These are difficult to encode in a short query but easier to evaluate once the consumer inspects photographs, maps, and review text.

Products and consumer preferences are represented in a k-dimensional Euclidean characteristic space Rk. A product i is described by its attribute vector Xi ∈Rk. A consumer has a latent ideal point x0 ∈Rk, which represents the product that would perfectly satisfy her preferences along these hard-toarticulate dimensions. This does not assume that the consumer can name, measure, or directly locate 8 Author: The Economics of Agentic Commerce x0. Rather, x0 is a modeling device that summarizes her latent tastes. The consumer’s mismatch from product i is the Euclidean distance between the product she receives and this latent ideal:

Mismatch(Xi) = ∥Xi −x0∥.

Without loss of generality, we normalize the consumer’s ideal point to the origin, x0 = 0. This is only a change of coordinates: for each consumer, we center the local product space at her latent ideal point.