An AI agent can search your catalog, compare offers and finish checkout without a hiccup — and still hand your customer the wrong item. Not because the transaction failed, but because the agent never had a reliable way to tell your product apart from a near-identical one: a different pack size, a seasonal reskin, a promotional bundle. Standards bodies now have a name for this: the “last meter” problem, and it was the center of discussion at the first neutral, multi-vendor workshop on agentic commerce.

What happened in Zurich

The W3C and GS1 Workshop on E-Commerce for Humans and AI Agents ran September 8–9, 2026, hosted at Google’s Zurich offices, bringing together Google, OpenAI, Shopify, Mars and the standards community around one question: agents are already good at executing a purchase — are they any good at knowing, with certainty, which physical item they’re buying? Per Forkast’s post-workshop report, the honest answer is not yet: “the industry is racing to build agentic commerce protocols, [but] the foundational data layer remains fragmented.”

GTIN is becoming the shared key, but resolution is still missing

The workshop’s clearest point of convergence: both Google’s Universal Commerce Protocol (UCP) and OpenAI’s Agentic Commerce Protocol (ACP) rely on the GTIN (Global Trade Item Number) as the primary identifier for physical goods, according to Forkast. Shopify, which has UCP and WebMCP deployed across millions of merchants, used its session to flag “the necessity of protocol layering and the current gaps in attestation” — i.e., agreeing on GTIN as a field is not the same as having a trustworthy way to resolve it.

That resolution layer is where GS1 Digital Link enters: an open GS1 standard that expresses a product’s GTIN as a web address rather than just a barcode number, giving a physical item a canonical, brand-controlled URL that any agent can dereference. Digital Link CEO Paula Rivero, presenting the session “Oops! My Agent Bought It Again” (Digital Link sells GS1 Digital Link infrastructure, so treat the framing as vendor-interested), put the underlying problem plainly: “an AI agent can execute a purchase flawlessly and still buy the wrong product if it cannot pin down the precise item behind a name or description.”

The workshop had a concrete demonstration, not just a proposal: per Forkast, Mars and K10X ran a pilot using GS1 Digital Link and the GS1 Web Vocabulary specifically to solve Consumer Pack Variant (CPV) disambiguation — telling apart, say, a limited-edition wrapper from the standard SKU it sits next to on the same shelf.

The gap that’s still open

Two problems surfaced without a fix attached. First, a vocabulary interoperability gap: Paola Di Maio of the W3C AI Knowledge Representation Community Group noted friction between schema.org, GoodRelations and the GS1 Web Vocabulary — three overlapping ways to describe the same product, none of which agents can be assumed to prefer. Second, liability: Claudia Caluori of Studio Polimeni pointed out that when an agent buys the wrong thing, the legal exposure sits with the business that deployed the agent or accepted its order — not with whichever standard it happened to be using. Per Forkast, that’s “a strong financial incentive for brands to control their own product data” rather than wait for the vocabulary fight to resolve.

Notably, the workshop closed without a formal outcomes report — this is a standards-convergence conversation still in progress, not a shipped mandate.

What this means for your store today

You don’t need to wait on GS1 Digital Link rollout to act on the part that’s already actionable: expose a gtin (or sku/mpn at minimum) on every Product in your JSON-LD, and make sure each variant in a ProductGroup carries its own distinct identifier rather than sharing one across sizes or colors. That’s precisely the check this audit already scores under “Product identifier (SKU / GTIN)” — see the field-level breakdown for exactly what schema.org requires. The Zurich workshop is evidence the entire industry considers this unsolved at scale; it’s also confirmation that merchants who fix it now aren’t solving a hypothetical.

FAQ

What is the “last meter” problem in agentic commerce?

It’s the failure point where an AI agent has successfully navigated a store and can technically complete checkout, but cannot definitively identify which exact physical product variant — pack size, edition, bundle — a shopper actually meant, risking a technically-successful purchase of the wrong item.

Do I need GS1 Digital Link to be agent-ready today?

Not yet as a hard requirement — no working group or mandate emerged from the September 2026 W3C/GS1 workshop, and the event closed without a formal outcomes report. What you can act on immediately is the underlying identifier problem: publishing a real gtin or sku/mpn on every product and variant in your structured data, which both UCP and ACP already rely on.

What’s the difference between a GTIN and a SKU for an AI agent?

A SKU is a merchant-internal identifier meaningful only inside your own systems. A GTIN is a market-wide identifier that lets an agent match your listing against the same physical product wherever else it appears — which is why Google’s UCP and OpenAI’s ACP both use it as their reference identifier for physical goods, per the September 2026 workshop.

Sources

Whether your own catalog carries a real GTIN or SKU on every variant — the exact field the industry just admitted it hasn’t solved for agents at scale — is one of the checks an AgentReady scan runs today.