Decision Architecture in E-Commerce

The preceding articles in this series have shown where the classical e-commerce funnel loses explanatory power, why stable KPIs increasingly mask opacity, where purchasing decisions form before measurable contact, where attribution breaks technically, and which new metrics hold up under these conditions. This article takes up the question that follows from these findings, namely how an e-commerce system must be designed to remain effective in shaping decisions when it can no longer fully map the decision process.

The answer proceeds from a fundamental shift in perspective, namely the transition from the system as observation instrument to the system as decision architecture.

The changed position of the system within the decision process

An e-commerce system, in most cases, no longer sits at the beginning of the purchasing decision. The growing share of selection, evaluation and pre-selection that takes place outside the system, before a user interacts with the shop at all, fundamentally alters its role. The system increasingly decides not whether a product is considered, but whether an already-formed decision is confirmed or discarded.

This shift can be demonstrated empirically. An nShift survey from early 2026 found that 58 per cent of consumers have already replaced traditional search engines with AI-powered alternatives for product discovery.¹ Shopify's own data show that AI-agent-driven orders on the platform grew 15-fold year-on-year in 2025, whilst AI referral traffic grew only eightfold over the same period.² The difference is instructive. Agents are functioning not merely as a discovery channel but increasingly as a purchase channel, in which the decision has already been structured before the shop visit.

For system design, this carries a clear consequence. Additional tracking logic does not improve decision relevance when the decisive impulses originate outside the system. What is required instead is an architecture designed to operate under uncertainty and to exert targeted influence on decisions by rigorously shaping the few points at which it becomes effective.

Observation system or decision architecture

The decisive distinction lies in whether a system is understood primarily as an observation instrument or as a decision architecture.

An observation system attempts to capture interactions as completely as possible and to derive optimisation approaches from them. Its logic rests on the assumption that complete data produce better decisions. This assumption was tenable for as long as the greater part of the decision process took place within the system.

A decision architecture, by contrast, is designed to shape the conditions under which decisions are made, regardless of whether every individual factor is measurable. It focuses less on fully reconstructing the past and more on deliberately shaping probability.

The concept is not new. Richard Thaler and Cass Sunstein coined the term "choice architecture" in their 2008 book "Nudge," describing the systematic design of decision environments to make certain outcomes more probable without restricting freedom of choice.³ What has changed is the context. In e-commerce, choice architecture has hitherto been concerned primarily with the design of product pages, basket and checkout. Today, the task extends beyond one's own system, because the upstream decision phases are increasingly shaped by external systems.

Product data as a prerequisite for visibility in the extended decision space

The quality and structure of product data have always mattered. What is new is that they no longer represent merely an internal optimisation factor but have become a prerequisite for visibility in a decision space that extends well beyond one's own shop.

A product that is clearly structured, semantically unambiguous and contextually embedded within a system has a higher probability of being correctly interpreted by upstream systems and included in decision processes. The evidence base is becoming increasingly concrete.

An AscentCore study from March 2026, which examined AI agent purchasing behaviour in simulated competitive environments across five product categories, arrived at a remarkable finding. When AI agents decided solely on the basis of publicly available product descriptions (without access to internal quality data), they selected a suboptimal product in 52.6 per cent of cases.⁴ The cause lay not in the quality of the products but in the quality of their descriptions. A single ambiguous sentence in a service description led the agent to discard the cheapest suitable option and recommend an alternative that cost 67 per cent more.

At the technical level, the data formats that matter for AI agents are becoming clear. JSON-LD, the structured data format based on Schema.org, has reached a market share of 89.4 per cent among structured data formats according to an Alhena analysis.⁵ Pages with structured data are cited 3.1 times more frequently in Google AI Overviews. Seventy-one per cent of pages cited by ChatGPT and 65 per cent of pages cited by Google AI Mode contain structured data.⁵ The inverse is equally telling. Pages without structured data are selected with significantly lower probability as information sources by AI systems.

An audit of a US Shopify store found that AI shopping assistants ignored over 40 per cent of inventory because the product feed lacked structured attributes and stable identifiers.⁶ No amount of compelling marketing copy can compensate for what is absent at the structural level.

Google itself underscored this development at NRF 2026 with concrete recommendations. These include product titles of 30 or more characters, descriptions of 500 or more characters, at least three additional product images at 1,500 by 1,500 pixels minimum resolution, complete logistics attributes (shipping, return policies, delivery timeframes), and so-called "conversational commerce" attributes such as compatible accessories and alternative products.⁷ The Shopping Graph now contains over 50 billion product listings with more than 2 billion updates per hour, and it forms the backbone of product recommendations in Google AI Mode, AI Overviews and the Gemini app.⁷

The identifier as quiet gatekeeper

One aspect that is routinely underestimated in strategic discussion concerns the completeness of product identifiers. GTINs (Global Trade Item Numbers) and MPNs (Manufacturer Part Numbers) are the keys AI agents use to match identical products across platforms. When a user asks ChatGPT where to find a particular product at the best price, the agent uses GTINs to identify identical items across multiple shops. Products without GTINs are systematically excluded from this matching process.⁵

Equally critical is data consistency. When the JSON-LD markup on one's own website shows a price of EUR 39.80 but the Merchant Center feed lists EUR 42.80, this discrepancy lowers the AI agent's trust score and leads to exclusion from recommendation sets.⁵ The consequence is unspectacular but consequential. Data quality at the attribute level increasingly determines whether a product even enters the decision space of an AI system.

From specification to semantic context

Structured data alone, however, are not sufficient. For AI agents to answer complex consumer queries, they need contextual information that goes beyond pure specifications.

A classically SEO-optimised product description ("Waterproof lightweight outdoor jacket men's size L") covers the attributes but cannot answer a query such as "I'm travelling to Europe in April, can you recommend a jacket that works in rain and fits in carry-on luggage?" For an AI agent to recommend a product for this query, descriptions such as "Handles light, sustained rain but is not designed for heavy downpours" and "Folds to 30 by 20 centimetres, fitting into the side pocket of a standard carry-on bag" are essential.⁵

SAP CX has articulated this approach conceptually, arguing that products need to be organised by "problems they solve" rather than by categories.⁵ Three decisive questions emerge. Who is this product for? In what situations is it used? For whom and which situations is it not suitable? Semantic summaries addressing these three points are the decisive factor in improving AI agent recommendation accuracy.

PIM as infrastructure for the extended decision space

When product data no longer serve merely internal catalogue management but form the basis for whether a product enters the recommendation space of AI agents, the role of the PIM system (Product Information Management) changes fundamentally. Its traditional function, namely distributing accurate product information to various sales channels, gains a new dimension, as PIM evolves into a knowledge source for AI agents.

Informatica has proposed the term "Agentic AI PIM" for this development, in which AI agents autonomously execute product information extraction, classification, enrichment and validation.⁸ Global product launches that previously took weeks could thereby be completed in hours. In March 2026, Pimly launched a Product Intelligence solution on Salesforce AgentExchange that unifies product data governance and activation across Agentforce, illustrating that PIM is becoming core infrastructure within the AI agent ecosystem.⁹

The implication for e-commerce decision-makers is concrete. Investment in PIM quality is no longer primarily a question of internal efficiency but a question of external discoverability. A shop with cleanly maintained PIM, complete identifiers and semantically enriched descriptions has a structurally higher probability of being included in AI-driven recommendation processes than a shop whose product data are optimised for marketing copy and visual presentation.

User guidance under changed conditions

When a user arrives at the shop with an already-formed preference, the task is no longer to guide them through an open decision process. The task shifts, because the system must now validate, refine or, in rare cases, deliberately unsettle the existing decision.

Research on AI nudging provides early empirical signals. A randomised experimental study published in Decision Support Systems in early 2026 examined the influence of AI-driven nudging on decision quality in online recommendation environments.¹⁰ The results show that contextual, situationally adapted interventions can improve decision quality, though only under certain conditions and with clearly bounded effect.

For the design of e-commerce systems, this means concretely that the balance tips towards less explaining and more confirming, less persuading and more reassuring. A user who arrives at a product page via an AI recommendation does not need persuasion but confirmation that their pre-selection is correct. Reviews, comparison data, availability information and transparent return policies become more important in this context than storytelling or emotional staging.

At the same time, there are moments when a deliberate interruption of the existing decision is sensible, for instance when a user has selected a product for which a more compatible or better-suited alternative exists. The art lies in designing this intervention so that it is perceived as assistance rather than distraction. Checkout optimisation research confirms this finding. Every friction point eliminated at checkout has a direct impact on revenue, because the user at this stage is no longer open to new decisions but wishes to complete the one already made.¹¹

The limits of decision architecture

This shift has clear boundaries. A decision architecture can shape the conditions under which decisions are made. It can raise the quality of information feeding into the decision process. It can reduce friction and strengthen trust. What it cannot do is control the upstream process in which the actual pre-selection takes place.

The points at which an e-commerce system becomes effective in shaping decisions today are clearly bounded. They lie in the quality and structure of product information, in the consistency of brand and trust signals, in the ability to make context rapidly comprehensible, and in the reduction of complexity at the points where the user is still open to adjustments. They do not lie in the seamless capture of every single touchpoint.

Consequences for system decisions

This also shifts the perspective on system decisions in e-commerce. The selection of a platform, the design of data models or the integration of tools can no longer be evaluated solely on how well a system can measure and optimise. It must be oriented towards how well the system can support decision processes that partly originate outside its direct sphere of influence.

Shopify has advanced furthest in this regard. With the rollout of Agentic Storefronts, approximately 5.6 million shops are now connected by default to ChatGPT, Gemini and Microsoft Copilot.² The Shopify Catalogue API executes 40 million LLM inferences daily (roughly 16 billion tokens), automating categorisation, attribute extraction and variant consolidation.² BigCommerce pursues a similar approach through its API-First architecture but requires manual registration via the OpenAI Merchant Program. Even Shopify's "Agentic Plan" allows merchants on Magento, Salesforce Commerce Cloud or custom stacks to register their product data in the Shopify Catalogue and thereby enter AI agents' recommendation sets without migrating their e-commerce infrastructure.²

A system is thus no longer measured by its ability to explain the past but by its ability to make future decisions more probable. This is a different category of requirement, and it marks the transition from an e-commerce that defines itself through tracking to one that organises itself through decision architecture.

What this means for the next 12 to 24 months

Gartner forecasts that by 2030, roughly 20 per cent of all transactions will execute through AI platforms.² Morgan Stanley expects nearly half of all online shoppers to use AI shopping agents by then.¹² Shopify recommends a 90-day sprint to establish a baseline of agent readiness. That begins with a product data audit in the first 30 days, followed by structured data implementation (Product, Offer, AggregateRating schemas, GTINs, semantic summaries) in days 31 to 60, and real-time synchronisation plus Share of Model monitoring in days 61 to 90.²

The strategic question is not whether this shift will occur but how quickly. The early part of the curve, where competitive advantages are built, has already begun. The closing article in this series, Strategic Consequences for Commerce Leaders, translates these findings into practical measures for the next twelve to twenty-four months.

This article is part of the Decision Journey Analytics series, examining how e-commerce decisions evolve beyond the classical funnel.

Glossary

Decision Architecture (Choice Architecture) The systematic design of decision environments with the aim of making certain outcomes more probable without restricting freedom of choice. In the e-commerce context, the term refers to the design of product pages, data structures and user guidance such that already-formed purchasing decisions can be confirmed and completed.

Agent-Ready Product Data Product data prepared so that AI agents can read, interpret and include them in recommendation processes without human mediation. Requires structured formats (JSON-LD), complete identifiers (GTINs) and semantic contextual information.

JSON-LD (JavaScript Object Notation for Linked Data) A standard format for embedding structured data in web pages based on Schema.org vocabulary. Preferred by AI agents because it can be parsed without HTML traversal.

Shopping Graph Google's product knowledge graph, comprising over 50 billion listings and forming the foundation for product recommendations in Google AI Mode, AI Overviews and the Gemini app.

PIM (Product Information Management) A system for centrally managing and distributing product information across sales channels. Increasingly evolving from an internal data management tool into a knowledge source for external AI agents.

Nudging A behavioural economics concept in which the design of the decision environment (not the options themselves) is used to make certain decisions more probable. Relevant in e-commerce for the design of product pages and checkout processes.

GTIN (Global Trade Item Number) A globally unique product identifier used by AI agents for cross-platform product matching. Products without GTINs are systematically excluded from comparison processes.

Sources

¹ nShift Consumer Survey, early 2026: "58% of consumers have replaced traditional search with AI for product discovery." Cited in Digital Applied: "E-Commerce AI Agent Readiness Assessment 2026."

² Shopify Enterprise Blog: "Agentic-Ready Product Data: How to Get It & the Cost of Inaction," 2026. Stellagent: "Agent-Ready Product Data," April 2026.

³ Richard H. Thaler, Cass R. Sunstein: "Nudge: Improving Decisions About Health, Wealth, and Happiness," 2008.

⁴ AscentCore: "The Shift to Business-to-Agent (B2A) Commerce: Why Your Product Descriptions Are Now Your Most Critical Sales Asset," Research Report, March 2026.

⁵ Alhena: "Schema Markup and AI Search for E-Commerce," 2026. Stellagent: "Agent-Ready Product Data," April 2026.

⁶ Toolient: "Product Feed Optimization for AI Agents," March 2026.

⁷ Google: NRF 2026 Merchant Center Guidelines. Dataiads: "Google Merchant Center Product Highlights: 2026 Complete Guide."

⁸ Informatica: "Agentic AI PIM: The Next Evolution in Digital Commerce," 2026.

⁹ Pimly / Salesforce AgentExchange: "Product Intelligence Launch," March 2026.

¹⁰ Yuxiao Luo, Nanda Kumar, Adel Yazdanmehr: "AI nudging and decision quality: Evidence from randomized experiments in online recommendation setting," Decision Support Systems, Volume 200, January 2026.

¹¹ Digital Applied: "eCommerce Checkout Optimization: UX Guide 2026."

¹² Morgan Stanley: Forecast on AI shopping agent adoption, cited in Adfinite: "Optimize Your Shopify Product Data for AI Agent Discovery," 2026.

For an analysis of how visibility is reorganising in AI-driven search systems, see our article on Generative Engine Optimization. And for a look at how Shopify is already building a concrete data infrastructure for AI-based discovery, see our article on the Shopify Agentic Plan.