Agentic Commerce for Retail Leaders: A Quick-Start Guide

Why this matters now

I spent more than 20 years building a digital commerce channel from zero into the No. 310 retailer among North America’s largest online retailers by web sales (Digital Commerce 360, 2023), in one of the hardest categories to sell online. After leaving that role in May 2026, I had the space to go deeper on a question I had been tracking closely: how AI is changing retail discovery, shopping behavior, and the path to purchase.

The conclusion became hard to ignore. Being discovered, understood, trusted, and eventually transactable by AI assistants is no longer just a marketing issue. It is a leadership challenge.

Shoppers are already handing more research, comparison, and purchase preparation to AI assistants. At the same time, the infrastructure is arriving quickly. In roughly six weeks, Google introduced Universal Cart, Mastercard debuted Agent Pay for Machines, and Salesforce shipped what it called its biggest Agentforce Commerce release yet. These are not isolated announcements. They point to the same shift. AI is becoming a new decision layer between the shopper and the retailer.

The question for retailers is no longer only whether their site ranks. It is whether an assistant can access their catalog, understand their products, verify their claims, trust their policies, and guide the shopper toward the right purchase without creating operational risk.

That creates a new kind of visibility problem. A product page may look polished to a human while the information an AI assistant needs is incomplete, inconsistent, or buried across product data, feeds, policies, reviews, variant logic, and checkout rules. A person can often work through that friction. An agent may simply move on.

The visibility problem is already here

Public Adobe Analytics data shows AI-driven traffic to U.S. retail sites grew 235% year over year across January through May 2026. During Prime Day 2026, that traffic jumped 89% year over year and converted 40% better than non-AI channels like search, email, and social. Human-approved checkout inside or adjacent to AI surfaces is real and expanding, though the models vary by platform and remain early. Fully autonomous purchasing is further along in protocol infrastructure than many retailers realize, but consumer trust in in-chat checkout is still limited.

That means the urgent work is not autonomous transaction governance. It is discovery and catalog readiness. If an AI assistant narrows the field before the customer ever lands on your site, you either appear in that consideration set or you disappear from the journey.

I have lived through enough ecommerce shifts to know the difference between a trend and an operating problem. Mobile, site speed, personalization, retail media, and platform migrations all looked like digital projects at first. In reality, they became leadership challenges because they changed how customers found products, trusted information, and completed purchases.

Agentic commerce is not one thing

A lot of confusion in this market comes from treating all AI commerce as the same thing. It is more useful to separate it into layers:

• AI-assisted discovery: The assistant helps the shopper research, compare, shortlist, and understand tradeoffs.
• AI-to-cart or AI-to-PDP handoff: The assistant sends the shopper to the exact item, variant, configuration, or cart state on a retailer-owned surface.
• AI-mediated checkout: The shopper approves a purchase inside an AI surface or embedded checkout flow, with the merchant still owning the order, payment rules, fulfillment, and service relationship.
• Autonomous agent purchasing: The shopper pre-authorizes an agent to complete a purchase later within defined limits. This is the most sensitive layer because it raises consent, fraud, liability, auditability, identity, and payments questions.

Most retailers should not build their roadmap around the last layer first. The practical work starts with the first three: discovery, handoff, and trustworthy checkout readiness.

Get the language right

There is a lot of loose vocabulary in this space, and using it loosely is a fast way to lose credibility with senior leaders. Here is the practical distinction I use:

• SEO: Search engine optimization gets a human to your site from a search engine. The traffic source is a classic results page, where a person types a query, scans links, and clicks through.
• AEO and GEO: Answer engine optimization and generative engine optimization influence how your brand and products show up when AI systems summarize, compare, and explain options. They shape whether you appear in the answer, how accurately you are represented, what sources are cited, and which claims the model has enough evidence to trust.
• ACO: Agentic Commerce Optimization is a term coined by Scot Wingo, founder of ReFiBuy, and is gaining traction as working vocabulary for this layer of the stack. The practical definition is the product-level work that makes your catalog eligible, understandable, trustworthy, and transaction-ready for AI assistants and shopping agents. I use it as an executive framework, not a settled standard, but it is the clearest shorthand available for what retailers actually need to do.

The clean executive rule: SEO wins the click. AEO and GEO influence the answer. ACO prepares the product to be selected, verified, and acted on.

Own the product card

In traditional search, the goal was to win visibility for a page. In agentic commerce, the goal is broader: own the product card. That means AI shopping engines can correctly recognize the product, map it to the right canonical item, understand its variants and offers, and see your product as accurate, complete, trustworthy, and eligible to recommend.

This matters because agents do not always organize the journey around your PDP. They may organize it around a product card: the recognized product, its attributes, images, reviews, sellers, prices, availability, delivery promises, policies, variants, and related offers. If your identifiers are weak, your variant structure is confusing, or your offer is attached to the wrong product record, you can lose visibility even if your own website looks polished.

Owning the product card means the agent can identify the product correctly, trust the information attached to it, and connect your offer to the shopper’s intent. That is the practical bridge from ACO strategy to catalog execution.

In my experience, catalog problems rarely show up as “catalog problems.” They show up as conversion leaks, confused customers, and costly returns. AI assistants will only make that more visible.

The mental shift that matters is this: for the better part of two decades, we optimized to win a click. Now we also have to optimize to win a recommendation we may never see happen. The funnel can start somewhere the retailer does not control.

Four moves for retailers

Move 1: Get your catalog to the agents, cleanly and completely

The principle is simple: your product data has to be available where AI assistants can actually use it, and it has to be consistent enough to trust. A polished PDP is important, but it is only one path. Feeds, APIs, commerce platform catalogs, structured data, and protocol integrations may all shape what an assistant understands, verifies, and recommends. In practice, retailers need more than one path because each source can break, lag, or contradict the others.

• Create an agent-access policy, not just an unblock list. Review robots.txt, bot-management rules, crawler permissions, rate limits, and partner access. Decide which AI crawlers, feeds, APIs, platform catalogs, and partner agents should be allowed, how spoofing is detected, and who approves changes. Opening everything is naive. Blocking everything is self-defeating.
• Audit structured data on every PDP. Price, availability, brand, GTIN, MPN, condition, ratings, review count, variants, size, color, material, dimensions, shipping, delivery options, return terms, warranty, seller, merchant of record, and business policy data should be present where relevant and consistent with the feed and checkout.
• Protect product-card ownership. Make sure core products are mapped to the right canonical item across AI shopping engines, marketplaces, feeds, and shopping surfaces. Strong identifiers, clean titles, accurate brand and model data, complete variants, consistent images, and offer-level accuracy determine whether the agent can recognize your product and prefer your offer.
• Deliver a clean feed or platform catalog in parallel. An optimized PDP and a direct data path are not redundant. Feeds and catalog integrations give agents fresher, more controlled data. PDPs provide context, authority, and corroboration. Treat feed freshness, suppression reasons, missing attributes, and policy errors like revenue-impacting metrics.
• Make inventory and price trustworthy. Product availability, variant eligibility, delivery promise, promotional price, tax and shipping estimates, and final checkout price must reconcile across feed, PDP, cart, and checkout. If the agent sees contradictions, the safer recommendation is often the competitor with cleaner data.
• Model the variant tree correctly. In high-consideration categories, one item may be a tree of sizes, colors, finishes, configurations, delivery rules, service options, warranties, and compatibility constraints. If those variants are modeled as disconnected products or hidden behind page logic, an agent can easily recommend the wrong thing.

Move 2: Make product information agent-readable and intent-rich

For years, short product copy often worked. A tight title, a few clipped specs, and lifestyle copy could get a human to click. That is no longer enough. AI assistants read for meaning and match against natural-language shopper intent. Complete, structured, accurate product data can answer the shopper’s real question.

• Fill in the product record, not just the page. Materials, dimensions, compatibility, warranty terms, certifications, constraints, service requirements, assembly, delivery limitations, and product limitations all matter. The goal is completeness and accuracy, not word count for its own sake.
• Add conversational attributes. Older product feeds rarely captured use case, occasion, compatibility, buyer intent, skill level, customer constraints, pairing logic, audience fit, or decision context. Agents need these fields to answer questions about what is best for a specific situation.
• Use direct language early. Make the product type, ideal customer, core use case, most important constraint, and key differentiator explicit in the first lines and in structured fields. Do not make the model infer important facts from lifestyle copy.
• Use customer language as missing data. Site search terms, zero-result searches, support chats, return reasons, reviews, delivery complaints, financing questions, and sales associate notes reveal the language real shoppers use. Feed that language back into attributes, FAQs, PDP copy, buying guides, and product taxonomy.
• Separate truth from persuasion. Agents need facts they can verify. Marketing language still matters, but the core product record should be literal, specific, and testable.

Move 3: Win the evidence and trust layer

An AI assistant should not take your marketing copy at face value. It corroborates. It may compare your PDP, structured data, feed, reviews, third-party sources, price, delivery promises, return terms, Q&A, and public reputation. The agent needs evidence, and that evidence lives both on and off your site.

• Build product-level review depth. Volume, recency, specificity, and usefulness matter. Forty recent reviews with fit, size, use-case, delivery, assembly, durability, and return detail are often stronger evidence than four generic five-star ratings.
• Use PDP FAQs to tell your authoritative side. Genuine product-level Q&A gives the agent clean, attributable text and gives you a place to correct stale, incomplete, or wrong third-party claims.
• Earn presence in credible retrieved sources. Identify which buying guides, category authorities, review sources, marketplaces, forums, and expert roundups show up when AI systems answer category questions. This is not old link-building. It is evidence management.
• Close contradictions. If your PDP says one delivery promise, your feed says another, reviews complain about a third, and checkout adds a fourth, the agent has a trust problem. It will usually resolve uncertainty by recommending something easier to verify.
• Make policies machine-readable. Returns, warranties, delivery, financing, installation, service areas, white-glove options, and exclusions should not be buried in PDFs or vague policy pages. These are buying-decision inputs.

Retailers often underestimate how quickly small inconsistencies become trust problems. A customer might forgive one confusing delivery message. An assistant comparing five options may simply avoid the uncertainty altogether.

Move 4: Prepare for AI-to-cart, AI-surface checkout, and payment trust

This is the area most teams either overhype or ignore. Human-approved AI checkout is no longer hypothetical. Microsoft Copilot Checkout launched in January 2026 with PayPal, Shopify, and Stripe; Google’s Universal Cart started rolling out in the U.S. in May 2026. At the same time, consumer trust and adoption are still developing. The right posture is practical readiness, not science fiction.

• Treat AI-to-cart as a conversion path. If an assistant sends a shopper to your site, land them on the exact variant, preserve price and availability, keep cart state where possible, surface shipping and returns quickly, and avoid forcing the customer to restart the journey.
• Know which checkout model you are in. The purchase may happen on the merchant site, in an in-app browser, through an embedded AI-surface checkout, through a native AI checkout, or through a future pre-authorized agent payment. These are different operating models.
• Define payment and authorization governance now. Decide who owns consent records, fraud risk, chargebacks, loyalty identification, tax, returns, customer service, refunds, substitutions, backorders, and order changes when an agent initiates or completes part of the transaction.
• Map failure modes. Wrong variant, stale price, unavailable item, incompatible accessory, delivery restriction, service-area mismatch, backorder, promotion mismatch, tax or shipping change, financing change, and return-policy surprise are not edge cases in complex commerce.
• Measure the handoff. Track AI-originated sessions, feed-driven referrals, AI channel attribution, app-driven journeys, conversion rate, abandoned cart causes, post-purchase service contacts, returns, and customer satisfaction after AI-influenced purchases.

Know the protocol stack

Leaders do not need to personally implement every protocol, but they do need to understand what each layer is trying to solve. These layers can stack. They do not all compete for the same job.

• ACP, the Agentic Commerce Protocol: OpenAI’s protocol, developed with Stripe and leading merchants, is open infrastructure for AI-native commerce in ChatGPT. It supports agent-ready commerce flows, including richer product discovery and merchant-provided product data. The important merchant principle is keeping merchants involved as merchant of record, with responsibility for fulfillment, returns, support, and post-purchase communication wherever ACP-powered checkout occurs.
• UCP, the Universal Commerce Protocol: Google’s broader commerce-journey layer is designed to give AI agents, merchants, commerce platforms, and marketplaces a shared language across discovery, cart, checkout, and post-purchase actions. Agents need reliable access to product data, pricing, delivery promises, checkout options, and order status without one-off integrations for every emerging surface.
• AP2 and payment-trust protocols: Google’s Agent Payments Protocol, Visa’s Trusted Agent Protocol, Mastercard’s Agent Pay, and related FIDO work are aimed at agent identity, user authorization, cryptographically signed mandates, payment credentials, spending limits, fraud controls, auditability, and liability. The infrastructure is moving quickly, but it does not mean fully autonomous consumer retail purchasing is mainstream today.
• MCP, the Model Context Protocol: MCP is an open standard created by Anthropic and donated to the Agentic AI Foundation under the Linux Foundation. It provides a secure way for AI systems to connect with external tools and data sources. In commerce, MCP can help an agent query catalog, inventory, order, service, CRM, or ERP systems, but it does not by itself solve commerce semantics, payment authorization, fraud controls, customer consent, or merchant policy.

The practical takeaway: supporting more than one path is how you stay present across ChatGPT, Gemini, Copilot, Claude, brand-owned agents, and whatever comes next. The goal is not protocol fandom. The goal is reliable product understanding, safe action, and measurable commerce outcomes.

Major platforms

The major AI platforms are moving toward the same broad destination, but they are not interchangeable. Each is developing a different mix of discovery, product data ingestion, checkout handoff, embedded checkout, agentic workflow, and payment authorization.

• ChatGPT and OpenAI are important consumer AI surfaces for product discovery and comparison. Being discoverable in ChatGPT, being accurately represented in product comparisons, and being eligible for an approved checkout or app integration are related but separate readiness questions. The broad priority remains catalog quality, feed integrity, product accuracy, and a clean path from AI discovery to merchant-owned conversion.
• Google is building a broad commerce layer across Search, AI Mode, Gemini, Shopping, Google Pay, and merchant integrations. UCP and Universal Cart point toward a more standardized path from discovery to cart, checkout, and post-purchase action while keeping the retailer as merchant of record in eligible flows.
• Microsoft Copilot is a clear example of AI-surface checkout. Copilot Checkout can let eligible shoppers move from discovery to purchase inside Copilot with participating commerce and payment partners while merchants remain merchant of record. That raises operating questions around attribution, customer data, loyalty identification, returns, service ownership, and the post-purchase relationship.
• Claude and Anthropic should not currently be treated as a major top-of-funnel consumer shopping destination. Their importance is different. Claude, MCP, and Anthropic’s agent research matter most for internal merchant workflows, tool use, catalog operations, customer-service workflows, data cleanup, and experiments in agentic decision-making.
• Shopify is turning AI-channel readiness into platform plumbing through Shopify Catalog, Agentic Storefronts, ChatGPT discovery, Copilot Checkout, Google AI Mode, and Gemini participation. Platform distribution does not remove the merchant’s responsibility for clean product records, accurate policies, trustworthy inventory, variant logic, and operational governance.
• Salesforce is positioning Commerce Cloud and Agentforce Commerce as the connection layer between AI-driven shopping and the enterprise systems that govern pricing, promotions, customer context, fulfillment, order management, service, and B2B and B2C workflows. The risk is assuming platform integration alone solves governance, attribution, product data quality, or cross-functional ownership.
• Amazon is taking a controlled, first-party approach, building experiences across its own ecosystem while defending access to its store and participating in emerging standards. The lesson is that Amazon wants agentic shopping to happen on its own terms.
• Walmart is taking a more distributed but still merchant-controlled approach, participating where AI discovery happens while keeping control of product data, cart logic, checkout, service, and governance.

Two cautions matter. First, the infrastructure is still being built before the economics are fully proven. Platform fees, ranking rules, attribution standards, customer data norms, retail media models, and ownership of the customer relationship are still evolving. Second, platform plumbing does not remove the retailer’s work. It moves the work upstream into catalog readiness, product evidence, policy clarity, payment governance, measurement, and operational trust. In an agentic commerce environment, bad data does not stay hidden. It travels faster.

Why this cannot wait until Black Friday

Last holiday season, before most of the infrastructure in this guide existed, Salesforce reported that AI and agents influenced 20% of global online retail sales, or $262 billion, through personalized recommendations and deeper customer engagement. Salesforce also reported that retailers deploying their own AI agents saw a 59% higher growth rate, averaging 6.2% year-over-year sales growth versus 3.9% for retailers without them.

The work in this guide is not a 2027 roadmap, and it is not an October project. Catalog audits, structured data cleanup, feed integration, and policy clarity take months to execute and longer to stress-test under real volume. Retailers that do the work before peak are more likely to remain visible, trusted, and actionable when AI-assisted shopping journeys accelerate. Being “ready by Black Friday” is already too late. The real deadline is Labor Day, not Thanksgiving.

If your team does not have the bandwidth to run this internally on that timeline, outside help can accelerate it, whether that means senior guidance, interim leadership, or added hands from trusted platform partners and data specialists. Compressing a multi-month catalog and structured-data overhaul into weeks is possible, but it needs clear ownership and the right capacity.

Readiness checklist

A leadership-ready audit should pressure-test the following items before claiming agentic readiness:

• Products have accurate IDs, GTINs or MPNs where appropriate, brand data, taxonomy, variant relationships, titles, descriptions, images, and availability.
• Feed, PDP, structured data, cart, and checkout agree on price, sale price, availability, fulfillment promise, return terms, and shipping options.
• Top products have product-card coverage and ownership checks: they are mapped to the correct canonical product, attached to the correct variants and offers, and represented with accurate title, image, price, availability, review, policy, and fulfillment data.
• Product variants are modeled as variants, not as disconnected products or hidden template logic.
• Policy data is readable: returns, warranties, delivery zones, white-glove service, installation, financing, exclusions, and merchant-of-record details.
• AI crawlers, commerce partners, platform feeds, and API access are governed intentionally rather than handled through accidental bot blocking or blanket allow rules.
• The organization can identify AI-originated journeys, attribute orders where the platform supports it, and measure returns, service contacts, and margin after AI-influenced purchases.
• There is an owner with authority across merchandising, ecommerce, IT, analytics, CX, payments, legal, and operations.

Where this goes next: context capture

Once the catalog is clean and complete, the next edge is feeding agents the one thing they still lack: how real shoppers actually talk about your products. The richest source of context is already inside the business, often unused.

• The search and prompt language people use to find, and fail to find, products.
• Zero-result searches that reveal exactly what the catalog does not know how to answer.
• Reasons behind cart abandonment, returns, service contacts, warranty issues, and delivery complaints.
• Projects, occasions, constraints, objections, and use cases buried in reviews and sales conversations.
• The language sales associates use when they translate a messy customer need into the right product configuration.

Capturing even a fraction of this and feeding it back into product data is how you move from keeping up to getting ahead. Most merchants collect some of it. Very few operationalize it into catalog structure, PDP content, feeds, FAQs, buying guides, and AI-surface answers.

Who owns this

This is where good ideas usually stall inside retailers. Not because the goal is wrong, but because the work sits between teams. In most retail organizations, agentic readiness cuts across functions that rarely move in sync.

Merchandising owns attributes and catalog depth. IT or engineering owns feeds, APIs, crawler access, and checkout integration. Brand, CX, and service own reviews, Q&A, policy clarity, and reputation. Payments and risk own authorization, fraud, chargebacks, and compliance. Analytics owns attribution and intent data. Legal has a voice in consent, customer data, platform terms, and liability.

Each function can do its part well and the brand can still lose, because no single team owns the outcome. A senior leader must own agentic readiness end to end, with enough understanding to pull teams together and connect the work to the P&L. This is a leadership mandate before it is a project plan.

The bottom line

AI is becoming another decision layer in product discovery and consideration. It does not replace every channel, and it is not the customer. But when it narrows the field before the shopper reaches the cart, it can shape which products become visible, trusted, and actionable.

If an assistant cannot reach your catalog, understand your products, verify your reputation, reconcile price and availability, or trust the path to purchase, it is more likely to favor a competitor that is easier to understand and safer to recommend.

That matters most in the hardest categories, especially spec-heavy and return-sensitive goods where mismatches are expensive and customer confidence is hard-earned. Discovery is already shifting. Checkout is evolving across platforms, protocols, and payment models. The retailers that win will be the ones that make the journey easier for both the customer and the assistant acting on the customer’s behalf.

Underneath all of this, the job has not really changed. I have always been focused on one question: How do we make it easier for the customer to do business with us? Every replatform, page-speed improvement, search decision, merchandising decision, and operational tradeoff comes back to removing friction between a person and the thing they want to buy.

Agentic commerce is the newest version of that work. I am not approaching it as a trend to comment on. I am looking at it as an ecommerce leader who has spent more than 20 years building ecommerce systems where product data, customer trust, conversion, delivery, service, and platform execution all have to work together.

This is the problem I think is most worth solving in retail right now.

Sources reviewed

• Adobe Analytics: 2026 Prime Day Insights and retail AI-traffic reporting.
• Anthropic: Model Context Protocol documentation and Project Deal research.
• Digital Commerce 360: OpenAI agentic-commerce updates.
• EMARKETER: Research on agentic commerce, GEO, AEO, and AI-shopping forecasts.
• FIDO Alliance: Trusted AI-agent interaction standards.
• Forrester: Research on Instant Checkout, agentic commerce, and agentic payments.
• Google: Universal Commerce Protocol, Universal Cart, and Agent Payments Protocol materials.
• Mastercard: Agent Pay for Machines announcement.
• Microsoft: Copilot Checkout announcement.
• OpenAI: Product discovery, merchant product feeds, and Agentic Commerce Protocol documentation.
• Salesforce: Agentforce Commerce announcements and holiday commerce reporting.
• Scot Wingo and Retailgentic: Agentic Commerce Optimization research.
• Shopify: Agentic Commerce, Agentic Storefronts, Catalog, and Agentic plan materials.
• Visa: Trusted Agent Protocol and Intelligent Commerce materials.

About the author

Jeff Douglas

I bring 20+ years of ecommerce leadership experience, including revenue and budgeting responsibilities, customer experience and digital marketing leadership, Salesforce Commerce Cloud modernization and high-consideration retail. I advise businesses on customer experience, growth priorities, technology and AI-driven discovery.

Experience