The Product Ontology for the Agentic Commerce

Build your data layer for UCP and ACP.

In 2026, “data engineer” is no longer a metaphor. It is the job description.

The Reasoning Web is here, and it does not speak HTML. It speaks protocols.

As commerce shifts from the Web of Documents to the Web of Agents, the goal is to be readable, verifiable, and executable across multiple agentic economies. Before we talk architecture, I have spent some time wrapping my head around two standards shaping the terrain: Google’s Universal Commerce Protocol (UCP) and OpenAI’s Agentic Commerce Protocol (ACP).

Two Standards, Two Business Models, One Mandate: User Intent

Both protocols exist to operationalize the same thing: a user mandate expressed through an agent.

The difference is not whether the agent has permission. In both cases, it does.

The difference is where execution happens and what must be proven, in real time, to complete the loop.

UCP: Ecosystem workflow (Discovery → Fulfillment)

UCP is built for Google’s surfaces (Google Search, Google Shopping, Google AI Mode and the ultimate AI gatekeeper: Gemini, now also available across the entire Apple ecosystem) where the agent orchestrates a full commerce workflow: inventory, shipping, identity, returns, loyalty, fulfillment. It enables deep handoffs to merchant systems while keeping the session coherent.

UCP is optimized for orchestration at scale. It turns your catalog into an interface agents can operate. Strategically, it also modernizes Google’s monetization engine: ads and promotions move from “drive the click” to “shape the conversion.” Sponsored offers can be embedded inside the agentic journey, and negotiation becomes part of the funnel. 

ACP: Headless checkout (Conversation → Payment)

ACP is built for conversational storefronts where the interface is the dialogue and the purchase can complete without a traditional website visit.

ACP is optimized for transactional completion. In Instant Checkout, merchants pay a fee on completed purchases, while product ranking is not supposed to be influenced by whether checkout is enabled.

The strategy

For most brands, agentic commerce is not an either or decision. You build once and you run everywhere.

Look at platforms and marketplaces like Shopify, Etsy, and Walmart. They cannot afford to pick a single rail. Their merchants need to reach across ecosystems, and the winning move is compatibility, not allegiance.

So you ship a single agentic commerce feed that supports both:

• UCP for end-to-end workflow orchestration
• ACP for conversational checkout

And yes, this assumes you are ready to deal with agents. You still sell to humans, but more and more you will negotiate and transact through software. If you are reading this, you have already accepted that shift.

The Invisible Rails: Why GS1 is the “Primary Key” of Agentic Commerce

You cannot build a graph on shifting sand.

When we look at the Unified Agentic Commerce Feed and its strict requirements for product_id, certification, and shipping_weight, it is easy to see them just as “database fields.” But for an Autonomous Agent, these are coordinates in a map.

GS1 is the physical protocol that makes UCP and ACP possible. Think of it this way:

• GS1 gives you stable nouns (product, party, location, package, batch/serial) and validated master data flows.
  
• UCP and ACP add verbs (negotiate, quote shipping, authorize payment, complete checkout) plus agent-era control signals.

The GS1 Solution (GTIN): The Global Trade Item Number (GTIN) is the only global “Primary Key” for physical goods.

• In UCP: Google strictly enforces GTINs to merge your id-web and id-store into a single canonical entity (this is the important split that we need to make before March 2026 when we are selling the same product on ecommerce sites and retail at different prices).
• In ACP: When an Agent executes a BuyAction, it relies on the GTIN to ensure it is purchasing the specific variant (Size M, Color Red) and not a similar SKU.
• Takeaway: If you mint internal IDs (e.g., SKU-123), you are invisible to the global graph. You must use GTINs (EAN/UPC) to be “interoperable” across agents.

This is the reason WordLift publishes every product entity in its Product Knowledge Graph using the GS1 Digital Link standard. GS1 is the foundational layer that bridges the digital and physical worlds, while UCP and ACP serve as the action layers. Without strict adherence to the GTIN, you are out of the game.

The 3-Layer Architecture of an Agentic Commerce Feed

Agents operate in three modes: verify, transact, negotiate. Your data stack must match.

Layer 1: Trust, Identity, Compliance (Physics)

If the agent cannot verify the product, it cannot recommend it. This layer anchors claims to constraints.

• return_rate (ACP)
  A risk signal. It implies an inferred uncertainty, which in turn suggests a strong, and non-obvious, dialogue between the ecommerce team and both logistics and support. This level of cross-functional communication is a data point most ecommerce operations lack. In KG terms it means internal operational efficiency.  

• shipping_weight (UCP)
  Needed for accurate landed cost, shipping promises, and return calculations. Remember Google wants to negotiate the sales and speak about the sales conditions with certainty. 

• certification (UCP)
  Compliance as IDs, not adjectives. Registry-backed references beat “eco-friendly.” Use GS1 identifiers and Digital Link to anchor the product, then attach external registry IDs as first-class graph entities, not strings.

Layer 2: Commercial Logic and Gatekeepers (Transaction)

This layer controls the buy action. It separates “talk about it” from “execute it.”

• product_id splitting (UCP)
  Distinct identifiers when price/availability differ by channel (web vs local). Avoid mismatches that break eligibility.

• enable_checkout (ACP)
  A hard gate. Allow discovery and Q&A, block purchase when needed (B2B, regulated, high‑touch).

• auto_pricing_min_price (UCP)
  Margin protection. A floor prevents automation from negotiating you into bad unit economics.

Layer 3: Agent Policies and Context (Negotiation)

This is where you stop shipping “data” and start shipping “behavior.”

• negotiation_policy (UCP)
  Guardrails for offers: max discount, bundles, price-match rules, eligibility constraints.

• structured_answers (UCP) and q_and_a (ACP)
  Preemptive Q&A that kills uncertainty for your flagship products. Replace guessing with explicit answers. Again it seems obvious but it is not. It’s telling agents how well the system works behind the scenes. How effective are we in deriving relevant questions and answering them from our clients, support teams, help desk and physical stores. Once again: it’s organizational and accurate (when was the last time you have updated the FAQs? Is there a data to derive the context of that information?)
• Proactive Q&A: Eliminating Product Uncertainty

Replace guesswork with explicit, preemptive answers to address customer uncertainty about your flagship products.

While seemingly obvious, this is a critical, often overlooked practice: explicitly inform your agents how the system operates behind the scenes. This involves effectively deriving and answering relevant questions sourced from all channels, including clients, support teams, the help desk, and physical stores.

• Crucially, this is an organizational and accuracy-focused initiative. When were your FAQs last updated? Who is in charge? Do you have the necessary data to provide context for that information? Most merchants don’t. 

• relationship_type (ACP)
  Semantic linking for substitution/cross-sell. Turn out-of-stock into rerouting.

Structure Is Strategy

The difference between a feed and a graph is no longer academic. It boils down to economics.

If agents cannot validate your entities, they will not surface them. If they cannot execute safely, they will not transact. If they lack policy and context, they will not sell.

In 2026, the pipeline is the product. Build it once. Make it agent-readable everywhere.

Unified Field Matrix (UCP + ACP)

This is the unambiguous and evolving checklist. One unified schema, with protocol-specific requirements.

Priority legend

• Critical: missing or wrong data can reduce eligibility, block checkout, or trigger enforcement.
• Compliance: policy or regulatory driven fields.
• Growth: improves conversion or unlocks new surfaces.
• Pilot / New: emerging fields that may change.

Layer 1: Identity, Trust and Compliance

These attributes authenticate the product and the seller. If these are wrong, you lose trust and can lose visibility.

Priority	Field name (unified)	Required by	Technical spec / example	Business logic and value
Critical	is_incentivized_review	Both	Boolean (true / false)	Declares whether reviews are incentivized. Used to enforce review-policy compliance and to let agents discount biased review signals.
Critical	shipping_weight	Google UCP	Value + unit (e.g., 1.5 kg)	Enables accurate shipping and returns calculations where weight is required. Missing weight forces estimates and reduces price certainty.
Critical	return_rate	OpenAI ACP	Decimal fraction (e.g., 0.05 = 5%)	Merchant quality and risk signal. Helps an agent decide how safe it is to recommend and transact with the seller.
Compliance	certification	Google UCP	Object (e.g., { scheme: “EPREL”, id: “…” })	Registry-backed compliance claims. Replace marketing labels with verifiable program IDs for regulated categories.
Compliance	structured_title	Google UCP	String with declared provenance	Title clarity and traceability. If you generate titles, disclose provenance so downstream systems can apply policy and quality checks.
Compliance	seller_privacy_policy	OpenAI ACP	URL	Canonical privacy policy link. Needed for user questions during assisted checkout and for trust verification.
Compliance	seller_tos	OpenAI ACP	URL	Canonical terms of service link. Required to support disputes, refunds, and user consent during checkout flows.
New	popularity_score	OpenAI ACP	Integer (1 to 5)	Explicit social proof signal. Derive from sales velocity, repeat purchase rate, or other internal metrics, then freeze the mapping.
New	sustainability_incentives	Google UCP	Complex object	Declares eligibility for rebates or incentives. Include jurisdiction, program ID, and conditions so the agent can validate eligibility before recommending.

Layer 2: Commercial Logic and Gatekeepers

Transaction controls: price, availability, and whether an agent is allowed to complete a purchase.

Priority	Field name (unified)	Required by	Technical spec / example	Business logic and value
Critical	product_id_web and product_id_store	Google UCP	String IDs	Split identifiers when channel-level price or availability differs. Prevents agents from comparing mismatched offers and reduces policy risk from price inconsistencies.
Critical	enable_checkout	OpenAI ACP	Boolean (true / false)	Hard purchase gate. false means the product can be discussed and shown, but the agent must not initiate a transaction.
Critical	installment	Google UCP	Complex object	Installment and BNPL terms for high-ticket items. Provide down payment, periodic amount, count, and any required disclosures to avoid misleading totals.
Growth	subscription_cost	Google UCP	Object (period + amount)	Enables recurring purchase flows (subscribe and save). Use an explicit period and a fixed amount.
Growth	enable_search	OpenAI ACP	Boolean (true / false)	Discovery control. false hides the item from agent search and recommendations, while still allowing direct access by ID or URL if you choose.
Pilot	auto_pricing_min_price	Google UCP	Currency (e.g., 15.00 USD)	Margin floor for automated repricing or negotiation. Prevents agents and tools from pushing price below your acceptable threshold.
Growth	product_highlight	Google UCP	List of strings (2 to 4 bullets)	Fast scanning and conversion. Keep each bullet short, factual, and testable.

Layer 3: Agent Layer (Negotiation and Context)

Policy and context that let an agent sell without guessing.

Priority	Field name (unified)	Required by	Technical spec / example	Business logic and value
Agent	negotiation_policy	Google UCP	JSON object	Guardrails for offers and discounts. Define max discount, eligibility conditions, and exclusions so the agent can negotiate safely.
New	structured_answers	Google UCP	Key value list (Q → A)	Pre-answers common questions to reduce hallucinations. Treat as a controlled knowledge base with only verifiable claims.
New	q_and_a	OpenAI ACP	Plain text	FAQ content.
New	relationship_type	OpenAI ACP	Enum (e.g., substitute, accessory)	Defines how products relate so the agent can reroute on out-of-stock and run cross-sell paths without improvising.
New	compatible_with	Google UCP	List of canonical IDs	Strict compatibility linking (A fits B). Enables bundles and prevents wrong recommendations in constrained domains.
Visual	media_video_link	OpenAI ACP	URL	Video is the fastest way to explain complex products in-chat. Use a stable, publicly accessible asset URL.
Visual	virtual_model_link	Google UCP	URL to 3D asset (e.g., .gltf)	Enables AR and spatial previews where supported. Provide a stable URL and correct MIME handling.

This is the new moat: not AI content, not better prompts.

Data Structure. Context in other words. 

Agents buy what they can parse, verify, and execute. Everything else becomes noise.