It is the wrong question.
Three commerce protocols are now being driven in parallel by the three companies that actually shape consumer AI. None is converging. A fourth standard, older than any of them, is already running underneath every LLM answer about a product today, and most merchants have not optimized for it.
The merchants getting AI-discovered traffic in 2027 will be the ones who treated the protocol layer as four problems, not one. Here is what each one is, why none of them is winning soon, and the one substrate that already pays off.
𝐓𝐡𝐞 𝐟𝐨𝐮𝐫 𝐩𝐫𝐨𝐭𝐨𝐜𝐨𝐥𝐬, 𝐢𝐧 𝐩𝐥𝐚𝐢𝐧 𝐄𝐧𝐠𝐥𝐢𝐬𝐡
𝐔𝐂𝐏 – 𝐔𝐧𝐢𝐯𝐞𝐫𝐬𝐚𝐥 𝐂𝐨𝐦𝐦𝐞𝐫𝐜𝐞 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (Google). Announced in early 2026. Aimed at being a common language for agents, businesses, and payment providers across the full commerce journey. At launch, the first execution environment was Google’s own surfaces: Search AI Mode, the Gemini App, and Google Shopping, with Google Pay as the checkout rail. Shopify joined Google as a co-governor (confirmed on Shopify’s Q1 2026 earnings call), with Amazon, Meta, Microsoft, Salesforce, and Stripe joining the Tech Council last month.
𝐀𝐂𝐏 – 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐂𝐨𝐦𝐦𝐞𝐫𝐜𝐞 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (OpenAI). Open, cross-platform, designed to be independent of any single agent surface. Defines how AI agents discover products via merchant-provided feeds, surface accurate pricing and inventory, and initiate checkout using delegated payment tokens. The first implementation was ChatGPT Instant Checkout. OpenAI publicly walked that back in March 2026, redirecting transactions to merchants’ own storefronts (Modern Retail covered the retreat in detail). The protocol itself is still active and Checkout.com adopted it for enterprise merchants in early 2026.
𝐌𝐂𝐏 – 𝐌𝐨𝐝𝐞𝐥 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (Anthropic). Not commerce-specific. MCP is the open standard for connecting AI assistants to external tools and structured data. It is the protocol Claude, Cursor, and a growing number of MCP-compliant hosts use to access anything outside the model’s training. Commerce is one of many use cases; merchant catalog data is one of many data shapes MCP can carry. But because it is general-purpose and shipping in production today across the Anthropic ecosystem, it is the protocol least likely to be replaced.
𝐒𝐜𝐡𝐞𝐦𝐚.𝐨𝐫𝐠 𝐉𝐒𝐎𝐍-𝐋𝐃. Not a protocol in the agentic-commerce sense. A W3C-backed structured-data vocabulary that has been the foundation of search-engine product understanding since the early 2010s. It is the layer LLMs are already parsing today when they answer questions about products. Every protocol above sits on top of it, either explicitly or because the LLM’s training data depended on it. Scheme.org is the substrate.
𝐖𝐡𝐲 𝐧𝐨𝐧𝐞 𝐨𝐟 𝐭𝐡𝐞𝐬𝐞 𝐰𝐢𝐧𝐬 𝐬𝐨𝐨𝐧
The Net Revenue ran a piece in April 2026 that framed this perfectly: “Three main models pursue similar goals but are incompatible with one another. The standard that ultimately prevails will depend less on technical excellence than on who controls the platforms with the largest user bases. This is the classic pattern of any standards war: Betamax versus VHS, Amiga versus PC, cartridge versus CD-ROM. The best technology does not win; the most widely distributed one does.”
That maps cleanly onto where we are. Google has UCP and the Google distribution surfaces (Search, Gemini, Google Shopping). OpenAI has ACP and ChatGPT’s enormous installed base. Anthropic has MCP and the agent-tooling ecosystem that is hardening fastest. Each company owns surfaces the others do not. Each has reasons to keep its protocol distinct rather than concede to a competitor’s specification.
The Checkout.com piece on ACP versus UCP arrived at the same conclusion from the payments side, observing that “the priority for merchants is not choosing one over the other, but preparing to support both as agentic commerce continues to evolve.” Different surfaces, different protocols, different moments of intent. Coexistence, not consolidation.
The historical analogy worth holding in mind: ten years passed before HTTPS was universal. OAuth has multiple incompatible major versions running concurrently in production today. RSS arguably never resolved. Internet protocols do not converge fast.
𝐓𝐡𝐞 𝐦𝐞𝐫𝐜𝐡𝐚𝐧𝐭 𝐭𝐫𝐚𝐩
Here is the trap, and it is the same trap merchants fell into during the SEO era and the social era before it.
Pick the wrong protocol, and the budget you spent on agentic-commerce readiness produces zero return when a different protocol wins. Pick the right one too late, and your competitors have an 18-month head start on agent ranking. Pick all of them at once, and you have three integration projects, three feed pipelines, and three sets of attribute requirements to maintain, and that does not scale for the 99% of merchants who are not Walmart.
The dominant industry response so far has been to wait. That works until shoppers normalize buying through AI agents, at which point the merchants who waited are invisible to the dominant surface and have to play catch-up against incumbents whose feeds are already in the agent’s training rotation.
There is a better move, and it has been hiding in plain sight.
𝐓𝐡𝐞 𝐬𝐮𝐛𝐬𝐭𝐫𝐚𝐭𝐞 𝐚𝐥𝐫𝐞𝐚𝐝𝐲 𝐩𝐚𝐲𝐢𝐧𝐠 𝐨𝐟𝐟
Schema.org JSON-LD is what LLMs already parse. Every meaningful AI-driven shopping answer today, whether it comes from ChatGPT, Gemini, Perplexity, or Claude, leans on structured product data from the open web. That structured data is overwhelmingly Schema.org JSON-LD.
The merchants who win the next twelve months are the ones whose catalogs are clean and AI-readable at the Schema.org layer first, and then who emit protocol-specific feeds on top of that foundation for whichever agents and surfaces matter to them. The substrate is what compounds. The protocols are what change.
This is not theoretical. Shopify’s Q1 2026 earnings disclosure included a number worth circling: traffic from catalog-powered AI searches converts at twice the rate of AI searches operating on scraped or outdated data. That single stat is the cost of being on the wrong side of catalog quality. Two times conversion, with no protocol decision required. Just clean, attribute-rich, machine-readable product data.
𝐖𝐡𝐚𝐭 “𝐜𝐚𝐭𝐚𝐥𝐨𝐠 𝐫𝐞𝐚𝐝𝐢𝐧𝐞𝐬𝐬” 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐦𝐞𝐚𝐧𝐬 𝐢𝐧 𝟐𝟎𝟐𝟔
The seven dimensions, which apply across every protocol:
𝟏. 𝐍𝐨𝐫𝐦𝐚𝐥𝐢𝐳𝐞𝐝 𝐭𝐚𝐱𝐨𝐧𝐨𝐦𝐲. Garment type, category, occasion, drawn from a controlled vocabulary an LLM can map against shopper intent.
𝟐. 𝐀𝐭𝐭𝐫𝐢𝐛𝐮𝐭𝐞 𝐝𝐞𝐩𝐭𝐡. David’s Bridal made this concrete when they audited every product for silhouette, neckline, fabric, train length, and size range. Not for human filtering. For AI agents that get asked “is this sleeveless?” and need to answer.
𝟑. 𝐂𝐨𝐥𝐨𝐫 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞. Not “blue.” Normalized shade-level values mapped to a shared color space.
𝟒. 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐞𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐬. Vector representations of image, title, and description, so visual and semantic similarity work at agent retrieval time.
𝟓. 𝐑𝐞𝐚𝐥-𝐭𝐢𝐦𝐞 𝐩𝐫𝐢𝐜𝐢𝐧𝐠. Current price, including promotions and regional adjustments. Agents that recommend a product that 404s on add-to-cart stop recommending that merchant.
𝟔. 𝐀𝐜𝐜𝐮𝐫𝐚𝐭𝐞 𝐢𝐧𝐯𝐞𝐧𝐭𝐨𝐫𝐲. Same logic.
𝟕. 𝐈𝐧𝐭𝐞𝐧𝐭 𝐬𝐢𝐠𝐧𝐚𝐥𝐬. The merchant’s own behavioral data, fed back into the optimization layer.
A catalog that ships fewer than five of these is being read by AI agents the way a sun-bleached menu gets read by a tourist. Maybe they order something. Probably they leave.
𝐓𝐡𝐞 𝐭𝐰𝐨-𝐛𝐞𝐭 𝐬𝐭𝐫𝐚𝐭𝐞𝐠𝐲 𝐭𝐡𝐚𝐭 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐜𝐨𝐦𝐩𝐨𝐮𝐧𝐝𝐬
The Net Revenue’s standards-war frame is right, and the practical implication is two bets:
Bet one: catalog quality at the Schema.org JSON-LD substrate. This pays off no matter which protocol wins. It is the discipline that compounds whether Google, OpenAI, or Anthropic eventually owns the dominant surface, because all three are training on, or directly consuming, structured product data.
Bet two: multi-protocol distribution. Generate UCP, ACP, and MCP-shaped feeds from the same optimized catalog. Build the pipeline once. Ship to every agent surface that matters. Re-run if the protocols shift. The cost is mostly engineering, and it is bounded. (Bounded meaning 12-18 months and a serious ML team – bounded does not mean trivial. The merchants buying this from a platform vendor instead of building it are making a buy-vs-build call on the same 12-18 months.)
The merchants who treat agentic commerce as a single protocol bet are signing up for a standards war they cannot afford to lose. The merchants who treat it as a substrate plus syndication problem are buying optionality against whoever wins.
This is what platforms like 𝐯𝐢𝐬𝐮𝐚𝐥𝐀𝐈 are being built for. 𝐜𝐥𝐞𝐚𝐧𝐞𝐫𝐆𝐏𝐓 cleans and enriches the catalog at the substrate layer. 𝐬𝐡𝐨𝐩𝐩𝐞𝐫𝐆𝐏𝐓 turns the on-site AI discovery surface into the intent-signal feedback loop that closes the 7th dimension. 𝐜𝐚𝐭𝐚𝐥𝐨𝐠𝐆𝐏𝐓 emits the protocol-specific feeds on top.
The broader point is not about any one platform. It is that the protocol layer was never going to converge fast, and the merchants treating that as a strategic problem rather than a wait-and-see problem are the ones building durable AI-commerce positions.
If a shopper asks an AI agent tonight about what your store sells, which protocol’s feed is your catalog speaking? And underneath that, is your catalog even readable?