When Agents Shop, Most Catalogs Don’t Show Up

Over the 2025 holiday season, Salesforce reported that AI agents influenced 17 percent of all retail orders across the Thanksgiving weekend, translating to approximately $13.5 billion in sales. Adobe’s measurement of Black Friday traffic told a different version of the same story: AI-driven visits to U.S. retail sites grew 805 percent year-over-year. Agentic commerce is no longer a planning scenario; it is a current channel with measurable volume, and the companies that prepared for it are already capturing a disproportionate share of what that volume produces.

What most assessments of this shift underemphasize is where the competitive differentiation actually lives. It is not in channel integrations, partnership agreements with AI platforms, or technology investments. It is in product data, specifically whether a retailer’s catalog can be understood by a system that operates entirely on explicit, structured information and has no tolerance for ambiguity or gaps.

What Agents Actually Need to Recommend a Product

AI agents do not evaluate products the way shoppers do. They do not respond to creative merchandising, editorial context, or the implicit signals that high-quality photography sends to a human browser. When an agent receives a shopping query, it searches for products whose data allows it to make a confident match. If product attributes are incomplete, if inventory status is delayed, or if pricing and promotional data are disconnected from availability, the agent cannot construct a reliable recommendation. McKinsey’s October 2025 report on agentic commerce describes the model as one in which “autonomous agents do the legwork — searching, filtering, comparing and even purchasing on behalf of the customer,” and that process requires data that is explicit, complete, and structurally accessible. Secondary reporting on McKinsey’s research suggests AI-generated product recommendations achieve conversion rates substantially higher than traditional search, a gap that widens as catalog completeness improves and narrows to nothing for products an agent cannot parse.

The catalog infrastructure most retailers built over the past two decades was designed for human shoppers navigating search results and product pages. Attributes were optimized for keyword matching, not semantic completeness. Dimensions, compatibility, material composition, and variant specifics were included where they aided SEO, omitted where they added workflow friction. A human shopper encountering missing attributes faces inconvenience. An agent encountering the same gaps has no mechanism for inference and no reason to wait.

The product data problem is not uniform across the supply chain, which matters for how different organizations should think about the work ahead. A top-ten CPG brand with dedicated catalog management teams and a mature product information management infrastructure faces a different version of this problem than a mid-market brand managing SKU data across multiple retail partners with inconsistent attribute standards. Among the largest players, the question is often one of integration and consistency across existing data systems, not the existence of the data itself. Among mid-market brands, catalog governance may need to be built before it can be improved.

Where Discovery Is Actually Happening

The traffic data makes the urgency concrete. Botify’s analysis of retail and e-commerce sites found a fourfold increase in AI bot activity between January and September 2025. According to Digital Commerce 360’s reporting on January 2026 Search Engine Land data, U.S. organic search traffic fell 2.5 percent year-over-year, while Seer Interactive’s September 2025 study across 42 organizations found that organic click-through rates dropped 61 percent on queries where Google AI Overviews appeared. Taken together, the data confirms that a measurably larger share of early-journey consumer attention is now resolving inside AI surfaces before a retailer’s own properties enter the picture.

Both Amazon and Walmart have made structural commitments to this reality. Amazon disclosed in its Q3 2025 earnings call that more than 250 million customers used its Rufus AI shopping assistant during the year, with monthly active users up 149 percent year-over-year and interactions up 210 percent. CEO Andy Jassy said Rufus is on pace to generate more than $10 billion in incremental annualized sales. Customers who engage with Rufus during a shopping trip are, per Amazon’s own data, over 60 percent more likely to complete a purchase. Walmart has taken a different architectural approach. Its CTO Suresh Kumar described in a July 2025 LinkedIn post a “super agent” framework centered on four coordinating agents, including Sparky, the customer-facing shopping agent already live in the Walmart app. According to Walmart’s corporate blog, Sparky is built on the Model Context Protocol and can handle tasks from review summarization to recurring household orders, with the system designed to be interoperable with third-party AI platforms. In October 2025, Walmart announced a partnership with OpenAI enabling customers to complete purchases inside ChatGPT, a move Walmart characterized as an AI-first retail experience.

Brands selling through both platforms face the practical consequence that the same product data must perform legibly across two distinct agent architectures with different data intake requirements, and both architectures are already active.

The Performance Gap and What It Signals

The current measurement picture is more complicated than most coverage suggests, and the complication matters for how catalog and commercial teams should prioritize their work. Professors Maximilian Kaiser of the University of Hamburg and Christian Schulze of the Frankfurt School of Finance and Management published a study in October 2025 analyzing 12 months of first-party data from 973 e-commerce websites with $20 billion in combined annual revenue, covering more than 50,000 ChatGPT-referred transactions against 164 million transactions from traditional channels. Their finding: ChatGPT referrals converted below every traditional channel except paid social, with overall LLM traffic representing less than 0.2 percent of sessions across the dataset. The authors wrote that their results “contradict widespread expectations of LLM superiority,” while also noting that conversion rates from ChatGPT referrals improved steadily throughout the 12-month observation period.

The study measures a specific and early phenomenon: organic referral traffic arriving at retailer sites from external LLM platforms. Platform-native agents like Rufus or Sparky operate differently, with AI-assisted discovery occurring within a retailer’s own ecosystem and drawing on catalog data, purchase history, and real-time inventory signals that external LLMs cannot access. Amazon’s 60 percent purchase-completion lift and the Kaiser-Schulze conversion finding are both accurate and cover different surfaces, different data conditions, and different points in the consumer journey.

The more durable performance signal for catalog and commercial teams is not conversion rate on external referrals but visibility: whether a product appears in agent recommendations at all. Products with incomplete attributes, inconsistent inventory signals, or thin descriptive content are effectively absent from agent-mediated consideration across all surfaces, and an absent product generates no session data, no referral, no evidence of the opportunity cost.

Retail media buyers and brand managers face a compounding version of the same problem. When an agent mediates discovery, the consideration phase, comparison, and filtering of options all occur inside the AI interface before a shopper ever reaches a product page. Attribution models built on observable touchpoints across a traditional funnel lose the behavioral signal that most optimization depends on. According to eMarketer, 55 percent of U.S. advertisers already report inconsistent targeting and attribution from retail media networks, and agent-mediated commerce narrows the observable portion of the customer journey further.

What Agent-Ready Catalog Infrastructure Actually Requires

The source article from Balaji Balasubramanian at SAP CX identifies three data priorities that consistently surface in the agent-readiness discussion: making product attributes explicit and machine-readable, providing semantic summaries that describe who a product is for and what problem it solves, and organizing products by intended use case alongside category. These priorities carry different practical weight depending on where a brand or retailer sits in the supply chain.

Category managers at large CPG companies often find the machine-readability requirement most immediately achievable because structured attribute data frequently exists within established product information management systems and needs to be surfaced in the right format. The harder work is semantic enrichment. Two products may share specifications, but an agent needs to understand that one serves a use case the other does not, and that requires descriptive content written with inference in mind, not keyword density. Digital Commerce 360 reported in December 2025 that companies which “performed best in 2025 invested in consistent product attributes, naming conventions, compliance documentation and governance across marketplaces,” quoting Jorrit Steinz, CEO of ChannelEngine, who added that companies entering 2025 with siloed or outdated content are entering 2026 “on the back foot.”

Brands scaling into wholesale through Amazon or Walmart for the first time frequently discover that catalog content optimized for their own site does not meet the attribute completeness standards those platforms’ agents require. Rufus draws on product detail pages, customer Q&As, and reviews to construct its recommendations. Mars United’s December 2025 analysis of more than 1,000 Amazon products found that Rufus predominantly recommends items with a minimum four-star rating and an average of approximately 9,000 reviews, with products carrying limited review volume effectively filtered out regardless of attribute quality. A brand entering a new retail channel faces a data readiness problem and a social proof accumulation problem simultaneously, and only one of those responds to a catalog investment.

Google’s VP of Ads and Commerce Vidhya Srinivasan described AI-assisted search conversations as two to three times longer and more contextual than traditional keyword searches, characterizing the shift as the difference between typing “blue shirt” and asking for a top suited to a formal bridal shower in a specific city. Organizing product content around the problems it solves, not only the categories it occupies, is what gives an agent the contextual structure to answer that second kind of prompt.

The Infrastructure Decision in Front of Operations Teams

McKinsey’s October 2025 agentic commerce report estimated that AI agents could mediate $3 trillion to $5 trillion in global consumer commerce by 2030, with up to $1 trillion of that in U.S. retail alone. Morgan Stanley estimated the U.S. share at $190 billion to $385 billion by 2030, representing 10 to 20 percent of online retail. McKinsey noted in January 2026 that the transition may unfold faster than prior digital shifts because AI systems operate along the same digital pathways as human users, without waiting for new infrastructure to be built. Retailers that built agent-compatible data foundations in 2025 are already being indexed across discovery surfaces, and that position compounds as agent traffic scales.

The catalog data requirement extends beyond product attributes into the real-time signals agents use to evaluate whether a recommendation is actually actionable. Inventory status, pricing accuracy, and fulfillment availability all feed directly into whether an agent presents a product as a viable option or filters it out of the recommendation set entirely. Walmart’s super-agent architecture addressed this directly. MetaRouter’s analysis of Walmart’s approach described the MCP-based design as ensuring that when Walmart updates its inventory database, all agents reading from that system receive consistent information automatically, keeping every catalog signal current across every surface where agents operate.

Commercial leaders at mid-market CPG brands working across multiple retail partners face the sequencing question most acutely. Full agent readiness across every retail channel simultaneously is not a realistic near-term target. The more productive question is which retail partners’ agent architectures are already driving meaningful discovery volume, and whether their catalog data meets those systems’ specific requirements. On Amazon, that means Rufus’s data intake criteria; on Walmart, MCP-compliant product data structures; on retailers building toward agent compatibility through platform partnerships, compliance with the data standards those partnerships impose. None of those requirements converge on a single standard, which is precisely why catalog governance, the organizational capability to maintain current, complete, and consistent product data across surfaces that do not share a common protocol, is the underlying investment all of them share.

Conversations On Retail

Conversations On Retail is a gathering place and resource center for retail and CPG executives, built to make it easier to stay current, discover the technologies and solutions shaping the industry, and connect with the people driving it forward.

We publish news, views, and reviews from staff editors, contributing experts, and trusted partners. Some articles are developed internally, while others are submitted by industry contributors or adapted from interviews and recorded conversations with industry leaders.