AI shopping agents are starting to do the work consumers used to do by hand, and they read product data, not product pages. This article is written for operators running complex multi-channel businesses, the ones with ERP connections, dropship platforms, marketplace APIs, headless front ends, and wholesale distribution, who need a technical plan for making their catalog, APIs, and checkout readable to those agents.
A 2026 IBM Institute for Business Value study found that 45 percent of consumers already use AI for part of the buying journey. The systems they hand the work to do not browse a storefront. They read feeds, evaluate structured data, and execute against APIs. A commerce stack built to convert humans clicking through a site is now serving a shrinking share of demand. The operators who win the next two years will be the ones who rebuild the catalog, the API surface, and the checkout flow to be readable by a machine that never sees the homepage.
Agentic commerce is an integration problem first and a standards problem second. Better copy does not move it. In early 2026, three competing protocols launched in quick succession, each backed by a platform large enough to force the question. Here is how to think about the work.
The three protocols: UCP, ACP, and AP2
Three standards now define how agents and merchants transact. Knowing what each one does, and which layer of the stack it touches, is the starting point for any sane roadmap.
The Universal Commerce Protocol, or UCP, was co-developed by Google and Shopify and announced at NRF in January 2026. It is an open-source standard built on REST and JSON-RPC that lets an AI agent connect to any merchant's catalog, start checkout, and manage orders through one unified API. UCP is a catalog-and-order interoperability layer. It standardizes how an agent discovers what a merchant sells, what it costs, whether it is in stock, and how to place and track an order.
The Agentic Commerce Protocol, or ACP, was announced by OpenAI as the first live standard for programmatic commerce flows between AI agents and businesses. Where UCP aims at broad catalog interoperability, ACP is wired into one of the largest consumer AI surfaces in the world. ACP readiness means distribution into ChatGPT-mediated shopping.
The Agent Payments Protocol, or AP2, was created by Google to handle the layer the other two leave open: payment authorization. AP2 lets agents make secure payments on a buyer's behalf, with explicit guardrails and accountability, so a buyer can set strict limits on what an agent may spend and under what conditions. Stripe's Agentic Commerce Suite and Google Shopping's new Universal Cart sit in the same payments-and-trust layer.
The three serve different jobs, so "pick one" is the wrong way to plan. UCP and ACP are catalog and checkout interoperability standards competing for the same role. AP2 is a payments standard that works with either. A defensible 2026 architecture treats protocol support as an adapter layer. Build the catalog and order systems to one clean internal contract, then expose protocol-specific adapters on top. Supporting a second or third protocol then becomes an adapter project rather than a re-platforming project. Operators who hard-wire to a single protocol will pay for it when the standards consolidate.
Agent readiness is an architecture decision
Being agent-ready means product data, APIs, checkout flows, and post-purchase workflows are all readable by an AI agent with no human in the loop. That sentence holds four separate engineering commitments, and most complex operators are weak on at least two.
Product data
Agents do not interpret merchandising copy or lifestyle imagery. They consume structured attributes. Every SKU needs a complete, normalized attribute set: price, availability, variant relationships, dimensions, materials, compatibility, and identifiers, exposed the same way across every channel. For operators whose product data lives split across an ERP, a PIM, a marketplace feed, and a headless CMS, the first job is one source of truth. An agent that reads one price from a UCP endpoint and a different price from a Google feed will distrust the whole catalog.
APIs
Agents execute against APIs in real time. A catalog endpoint that returns stale inventory, paginates inconsistently, or rate-limits hard will get pushed down the ranking. Inventory accuracy becomes a ranking input on its own. An agent that completes a purchase against an API and then receives an oversell cancellation has every reason to route the next buyer somewhere else. If inventory sync between ERP and storefront runs on a batch cycle, agentic commerce is the reason to move it toward real time.
Checkout flows
Agent-initiated checkout has to complete by API call. Multi-step flows that depend on JavaScript-rendered interstitials, CAPTCHA gates, or account-creation walls are friction an agent cannot or will not push through. The UCP order-management surface assumes a checkout that a program can drive, not one that needs simulated clicks.
Post-purchase workflows
Order status, shipment tracking, returns, and cancellations all need API exposure. The agent that sold the order is also the agent the buyer asks about delivery. If that agent cannot get a structured answer, the experience breaks on the channel the merchant just won.
Structured data is now the cost of being found
Before an agent transacts, it has to find the merchant, and discovery now runs on structured data. The evidence is clear. Google's AI Overviews now appear on 14 percent of shopping queries, a 5.6x rise in roughly four months. SE Ranking found that 65 percent of pages cited by Google's AI Mode and 71 percent of pages cited by ChatGPT include structured data. Content with proper schema markup has a 2.5x higher chance of appearing in AI-generated answers.
JSON-LD is the implementation standard. Google, Bing, Perplexity, and ChatGPT all rely on it to pull structured signals from a page. For product pages, the properties that carry the most weight are price, availability, GTIN, and shipping data. GTIN deserves direct attention. It is the identifier Google uses to match a product against the same product from other sources, which is what lets an agent comparison-shop accurately rather than guess. A product page with correct price, availability, GTIN, and shipping schema is far easier for an AI system to evaluate, and to trust, than one without.
For complex operators, the schema problem grows with scale. A headless architecture helps here. Because the front end is decoupled, the team can render JSON-LD on the server from the one source of truth and keep it matched to what the APIs return. The failure mode to hunt down is divergence: schema on the page that disagrees with the UCP endpoint that disagrees with the marketplace feed. Every mismatch is a reason for an agent to mark the catalog down.
Where multi-channel operators are most exposed
A wholesale-plus-DTC hybrid hits a structural seam. The DTC catalog and the wholesale catalog often carry different pricing logic, different availability rules, and different identifiers. Agents shopping the consumer side need the DTC truth. B2B procurement agents, arriving close behind consumer agents, will need the wholesale truth. Operators who never cleanly separated those data models carry the exposure twice.
Dropship and marketplace dependencies create a second exposure. If a real share of the assortment ships from third parties, agent-facing inventory accuracy is only as good as the worst supplier feed. Agents punish oversells. Operators who treat supplier data quality as a contract obligation, with service-level terms on feed freshness and accuracy, will out-rank competitors who treat it as someone else's problem.
ERP integration is the third. The ERP is usually the system of record for cost, inventory, and order status, but it is rarely built to serve low-latency external API traffic. The pattern that works is a commerce data layer between the ERP and the agent-facing APIs that caches, normalizes, and rate-shapes ERP data, so agents get fast, consistent responses without overloading a system that was never designed for that load. Any middleware evaluation in 2026 should put agent readiness on the requirements list at the start.
Priority actions
Build one source of truth for product data. Before touching any protocol, reconcile price, availability, variants, and identifiers across ERP, PIM, feeds, and CMS. An agent that sees conflicting data distrusts everything.
Publish complete JSON-LD product schema and make GTIN non-negotiable. Render schema on the server from the source of truth so it provably matches the APIs. Put price, availability, GTIN, and shipping first.
Move inventory sync toward real time. Batch cycles produce oversells, and oversells get a catalog pushed down by agents. Treat inventory accuracy as a ranking input.
Build protocol support as an adapter layer. Define one clean internal catalog and order contract, then expose UCP and ACP adapters against it, and plan AP2 into the payments layer. Do not hard-wire to one protocol.
Audit checkout for programmatic completion. Remove CAPTCHA gates, JavaScript-only interstitials, and forced account creation from the agent path. If an agent cannot finish by API call, the sale is lost.
Put service-level terms on supplier and marketplace feed quality. Agent-facing accuracy is only as strong as the weakest dropship feed, so make freshness and accuracy contractual.
Merchants who file agentic commerce under 2027 are misreading the clock. The protocols are live, consumer behavior is already at 45 percent, and AI shopping surfaces grow every month. The advantage goes to operators who rebuild for machine readability before their catalog becomes the one agents quietly route around.
