2025 was supposed to be the year AI shopping agents transformed retail. The protocols launched. The major retailers integrated. The projections reached into the trillions. By December, we have enough data to separate the signal from the hype.

The verdict: agentic commerce is real, the infrastructure challenges are harder than expected, and the merchants who invested in data quality captured disproportionate advantage. Here's what actually happened.

Agentic commerce protocols in 2025

Three foundational protocols emerged in 2025, each solving different layers of the agentic commerce stack.

‍OpenAI and Stripe launched the Agentic Commerce Protocol (ACP) in early 2025, creating a standardized way for AI agents to execute purchases. ACP handles the transaction layer: checkout session initiation, payment tokenization through Stripe, inventory verification, and order execution. When a ChatGPT user says "buy those running shoes," ACP is the protocol that makes the purchase happen without the user leaving the chat interface.

‍Anthropic released the Model Context Protocol (MCP), solving a different problem: how agents access and coordinate across enterprise systems. MCP prevents "bot sprawl" by allowing a single agent brain to query multiple backend databases—inventory, customer service, supplier data—through a unified interface. When Walmart's Sparky agent handles a request like "plan a unicorn-themed birthday party for 8 kids," MCP is what allows the agent to pull from party supplies, bakery, and invitations simultaneously.

‍Google introduced Agent2Agent (A2A), addressing agent-to-agent coordination. While ACP handles human-to-agent transactions and MCP connects agents to enterprise systems, A2A enables agents to communicate with each other. A customer's personal shopping agent can delegate tasks to a retailer's inventory agent, which can coordinate with a logistics agent for delivery scheduling. A2A provides the communication layer for these multi-agent workflows, enabling complex orchestration that no single agent could handle alone.

The protocol layer stabilized faster than most expected. By mid-2025, merchants had clear technical specifications to build against. The challenge shifted from "how do we connect" to "how do we make this work at scale."

How Walmart, Instacart, and Amazon built AI shopping agents

Four distinct approaches emerged among retailers who moved first. Each reveals something about where value actually accrues in agentic commerce.

Walmart built unified intelligence

Walmart's approach was the most architecturally ambitious. Rather than deploying separate chatbots for separate functions, they built a "Super Agent" architecture powered by MCP. Three agent personas—Sparky (customer-facing), Associate (employee-facing), and Marty (supplier-facing)—share a single data plane.

The insight: when Walmart updates their inventory database, they don't retrain three separate bots. The protocol ensures all agents read from the same source automatically. This matters because data consistency across surfaces is what makes agent commerce work. A customer asking Sparky about a product sees the same information an employee would see on the floor.

The infrastructure investment was substantial, but Walmart is betting that unified data infrastructure creates compounding advantage as agent usage scales.

Instacart became invisible infrastructure

Instacart made the most radical strategic bet: accept that the traditional storefront is declining and become fulfillment infrastructure for other platforms' AI agents.

Their ChatGPT integration isn't a marketing channel—it's a business model pivot. When a user asks ChatGPT to plan a week of high-protein vegan meals and order the ingredients, Instacart handles fulfillment without the user ever visiting Instacart.com. The meal plan gets generated, the cart gets built, the transaction completes via ACP—all inside ChatGPT.

The infrastructure requirement: real-time inventory sync across thousands of store locations. If the agent promises same-day delivery but the items aren't available at the nearest store, the experience breaks. Instacart's existing logistics infrastructure made this possible; merchants without that foundation would need years to replicate it.

Amazon defended its ecosystem

Amazon took a different approach with Rufus: use agents to keep customers inside the Amazon ecosystem rather than enabling external agent access.

The "Buy for Me" feature lets Rufus handle checkout steps autonomously—address selection, payment method, shipping preferences—once the user gives permission. Amazon Lens integration enables visual search: snap a photo of a broken screw, and the agent identifies and orders the replacement.

The strategy: make the Amazon app the easiest place to get product answers, reducing the need for customers to search Google or ChatGPT. Amazon isn't opening its catalog to external agents; it's using agents to defend market share.

The lesson for other retailers: Amazon can afford a closed ecosystem because of its scale. Most merchants don't have that luxury—they need to be discoverable by external agents or risk invisibility.

Shopify democratized agent access

For the millions of merchants without enterprise engineering teams, Shopify's Agentic Storefronts provided the fastest path to agent visibility. Launched in 2025, the platform handles ACP integration automatically—merchants get agent-compatible checkout endpoints without building custom infrastructure.

The caveat: Shopify solves connectivity, not data quality. A Shopify store with incomplete product data, stale inventory, or thin descriptions will be just as invisible to agents as any other poorly-maintained catalog. The platform removes the technical barrier but does not fix the underlying infrastructure gap.

Early adoption data showed Shopify merchants who had already invested in product data quality (primarily those selling through Google Shopping or Amazon) saw faster agent traction. Those with minimal product data investment remained invisible despite having the technical integration enabled.

Michael Kors optimized for conversion

The luxury and fashion play looked different. Michael Kors deployed Mastercard's Shopping Muse (built by Dynamic Yield) as a digital stylist rather than a transactional agent.

The capability: translate aesthetic language ("minimalist beach wedding outfit," "cottagecore vibes") into specific product recommendations. Early tests showed 15-20% conversion rate improvements compared to traditional keyword search.

For luxury brands, the protocol layer matters less than semantic understanding—translating the language of style into the language of inventory. The data infrastructure requirement shifts from real-time inventory sync to rich product attribution and cross-category relationship mapping.

What the data revealed about ChatGPT shopping performance

By October 2025, researchers had enough transaction data to measure what was actually happening.

‍Kaiser and Schulze's study analyzed 973 e-commerce websites with $20 billion in combined annual revenue, comparing 50,000+ ChatGPT-referred transactions against 164 million transactions from traditional channels. The findings challenged the hype narrative:

• ChatGPT referrals account for less than 0.2% of all e-commerce sessions
• Affiliate links convert 86% better than ChatGPT referrals
• Organic search outperforms ChatGPT by 13%
• ChatGPT dominates LLM traffic (90%+), but LLM traffic overall remains small

The researchers found that when ChatGPT traffic does arrive, it converts poorly. The gap isn't consumer demand—it's merchant infrastructure. Products with incomplete data don't appear in agent recommendations. Products with stale inventory create failed transactions. Products with inconsistent pricing across channels trigger checkout failures.

The silver lining: Adobe reported that AI traffic to US retail sites increased 805% year-over-year on Black Friday 2025, with Cyber Monday showing 670% growth. The trajectory is steep even if the absolute numbers remain small. ChatGPT now drives more than 20% of referral traffic to Walmart, nearly 15% to Target, and 10% to eBay.

Consumer adoption is accelerating. Salesforce reports that 39% of consumers—and over half of Gen Z—already use AI for product discovery. Morgan Stanley estimates 23% of Americans bought something via AI in the past month. Bloomreach found that more than 60% of consumers have used conversational AI for shopping at least once. The demand exists; merchant readiness doesn't.

The specific data problems merchants face are well-documented across forums and community discussions:

• Missing GTINs: Without valid Global Trade Item Numbers, agents can't reliably match products. One analysis found that a single missing GTIN is enough for agents to skip a product entirely.
• Incomplete descriptions: Thin or generic product copy forces agents to skip rather than guess. Agents need contextual information to understand products.
• Stale inventory: Products showing as available in agent recommendations but out of stock at checkout. Feed updates lag behind actual inventory changes.
• Variant confusion: Size and color options not properly represented in feeds. When a merchant sells "blue running shoes in sizes 7-13," each variant needs accurate availability.
• Price mismatches: Different prices across channels trigger transaction failures and damage reliability scores.

The infrastructure gap isn't theoretical. It shows up in every failed transaction, every invisible product, every customer who converts somewhere else because the agent couldn't find what the merchant actually sells.

Agentic commerce strategies that worked in 2025

Patterns emerged among merchants who captured value from agentic commerce this year.

Data quality as competitive moat

The merchants who performed best had invested in product data quality before agentic commerce arrived. Complete GTINs, accurate inventory sync, consistent pricing across channels, rich product attributes—these fundamentals became the difference between visibility and invisibility.

One analysis found that merchants with 95%+ data fill rates on core attributes saw dramatically better agent discovery. Below 80%, products were routinely skipped. The bar is high because agents have no tolerance for ambiguity—they move to competitors with cleaner data.

The infrastructure lesson: data quality isn't a backend operational issue anymore. It's a direct revenue driver with measurable impact on agent visibility. Research from Mirakl estimates businesses lose an average of $15 million annually due to poor data quality—costs that multiply in agentic commerce where agents ruthlessly filter incomplete listings.

Server-side collection as foundation

Merchants relying on client-side tracking (pixels, cookies, JavaScript tags) discovered a fundamental problem: AI agents don't trigger client-side events. They make HTTP requests directly to APIs and feeds.

The merchants who could capture and measure agent-mediated transactions had server-side data collection infrastructure. They could see agent traffic in their analytics, measure conversion by channel, and optimize their feeds based on actual performance data.

The infrastructure lesson: the "first mile" of data collection—server-side signal capture at the point of origin—determines whether you can measure and improve agent commerce performance.

Starting small and expanding

Merchants who tried to enable their entire catalog for agent commerce on day one struggled. Those who succeeded started with constrained product segments: bestsellers with low return rates, clean product data, and simple variants.

One merchant resource recommended: "Choose a contained product segment. Don't try to onboard your entire catalog on day one." The approach works because it surfaces data quality issues on products that matter most, allows iterative improvement, and doesn't overwhelm operations teams.

The infrastructure lesson: agent commerce is an infrastructure initiative, not a marketing channel launch. Treat it accordingly.

How agentic commerce reshaped retail media networks in 2025

The retail media implications became clearer this year. Analysts warned that retailers could lose retail media revenue if consumers bypass retailer websites to shop directly through AI platforms.

The shift: RMNs evolved from selling ad placements to licensing data access. When agents make purchase decisions, they don't see banner ads or sponsored product placements. They parse structured product data, compare pricing, evaluate loyalty benefits, and make algorithmic decisions.

Kantar's analysis suggests that retail media will shift from "impressions" to "influence signals." Instead of CPM and CPC, brands will pay for competitive positioning in agent decision-making—cost-per-agent-conversion models, revenue sharing, or algorithmic bidding for recommendation placement.

The retailers building for this future are creating central data layers that expose product feeds and shopper signals in standardized formats agents can consume directly. The measurement model shifts from "did the customer see the ad" to "did our product data influence the agent's recommendation."

The identity infrastructure gap compounds this challenge. According to eMarketer, 55% of US advertisers already report inconsistent targeting and attribution from RMNs—a problem that worsens dramatically when agents bypass traditional tracking entirely.

The state of agentic commerce heading into 2026

Agentic commerce arrived, but the transformation is still early. The protocols work. The consumer demand exists. The early adopters demonstrated what's possible.

The gap is infrastructure. Most merchants aren't ready—not because the technology is immature, but because their data systems weren't built for this. Looking back at 2025, the merchants who captured value shared common infrastructure characteristics that separated them from competitors.

Cross-surface data consistency. The same product data flows to agents, Google Shopping, marketplaces, and their own sites. No channel-specific cleanup processes. No conflicting information across surfaces.

Real-time inventory and pricing sync. Feed updates in minutes, not hours. Agent transactions that encounter out-of-stock items or price mismatches create failed experiences and damage reliability scores with agents.

Server-side data collection. The ability to capture signals from agent traffic that doesn't trigger client-side tracking. Measurement infrastructure that works regardless of how the customer (or their agent) arrives.

Identity resolution across surfaces. Agent transactions connected to customer profiles, enabling personalization, attribution, and loyalty across web, app, and agent-mediated purchases.

Schema-compliant product markup. Schema.org Product markup on product pages, allowing agents to parse structured data directly. Products without proper markup force agents to guess—and agents don't guess in your favor.

These aren't advanced capabilities. They're foundational data infrastructure that many enterprises haven't prioritized because the ROI wasn't clear until now.

The projections remain bullish. McKinsey projects $5 trillion in global agentic commerce volume by 2030. Morgan Stanley estimates $190-385 billion in US e-commerce spending by 2030—10-20% of online retail. Bain forecasts 15-25% of e-commerce.

The 2025-2026 window is when early movers establish the infrastructure advantages that late entrants can't replicate. The merchants who invested this year will compound that advantage as agent traffic scales. Those who waited will find the gap harder to close.

The question for 2026 isn't whether agentic commerce matters. It's whether your infrastructure is ready to capture it.

The retailers who will win next year are those who learned from 2025: data quality isn't optional, server-side collection is foundational, and the merchants who treat agent commerce as an infrastructure initiative—not a marketing experiment—will capture the value that laggards leave on the table.

‍

MetaRouter provides the first-mile data infrastructure enterprise retailers use to capture, normalize, and route data across all commerce surfaces—including the agent-mediated transactions that traditional tracking can't see. Learn how server-side collection powers agent commerce visibility.

‍