Which Layer I Would Fix Before a Merchant Protocol

I would not buy a merchant protocol and call the catalog problem solved. I would first name which layer is wrong: the product facts the company controls, or the open-web sources a shopping agent actually retrieves when it recommends something. A protocol can govern the first. It does not, by itself, become the second.

The comparison is L’Oréal and Unilever Fixed the Wrong AI Visibility Layer. That page is the deeper resource. This essay is how I would use it as a founder: where I would spend, what I would leave alone, and what evidence would make me reverse the call.

What the launch actually changed

On March 12, 2026, Azoma launched the Agentic Merchant Protocol as merchant-side infrastructure for how a catalog is defined, distributed, and governed across AI commerce surfaces (Azoma; Business Wire release). Azoma says AMP sits above OpenAI's ACP and Google's UCP rather than replacing them (Business Quarter). The early adopters named with the launch are L'Oréal, Unilever, Mars, Beiersdorf, and Reckitt (VentureBeat).

That list is a signal about urgency inside large CPG companies. It is not a measurement of what a shopping agent cites. Max Sinclair, Azoma's CEO, put the operating change in the launch release and a VentureBeat interview: marketplaces had controlled product detail pages, rankings, and distribution, and in an agentic setting those fixed pages are no longer the place the recommendation is assembled. The same interview records his point that companies which spent decades on brand equity are unwilling to hand representation of their products to an ungoverned model.

I would read that as a governance claim. The company wants one place to put catalog data, brand guidelines, compliance rules, regulatory requirements, competitive positioning, and persona-level targeting, and then have that package distributed. Azoma's release lists four AMP components: canonical machine-native product catalogues, programmatic open-web distribution, contextual prioritization of content, and an agent-agnostic interface. VentureBeat adds RegGuard, a compliance check, and citation tracking from the wider platform. Those are merchant-side controls. Agent-side protocols (ACP, UCP, AP2, MCP) describe how an agent connects, pays, and reads context. They do not decide whether the fact the agent repeats is the fact the brand would stand behind.

The market numbers are about behavior, not about your SKU

McKinsey's agentic commerce work projects that by 2030 as much as $1 trillion in U.S. B2C retail revenue could be orchestrated through agentic commerce, and that the global opportunity reaches $3 trillion to $5 trillion. Morgan Stanley Research estimates that agentic shoppers could reach $190 billion to $385 billion in U.S. e-commerce spending by 2030, a likely 10% share of online retail and up to 20%. A Riskified survey of more than 5,000 consumers, published October 2025 by a fraud-prevention vendor, found 73% using AI somewhere in their shopping journey: 45% for product ideas, 37% for summarizing reviews, 32% for price comparison. Only 13% reported completing a purchase after an AI referral, while 70% were at least somewhat comfortable with an agent buying on their behalf.

I would not brief a board with the trillion-dollar figure and a protocol logo in the same slide. The large numbers describe a shift in how shopping can be mediated. They do not say that L'Oréal's adoption moved a citation rate, and they do not say that a smaller catalog will be recommended because it joined the same standard. The purchase-completion figure is the one I would keep on the wall: most of the reported AI shopping behavior is still research, comparison, and review summary. The agent is assembling a description from distributed pages. If those pages are stale affiliate copy, the protocol's canonical record has to reach them or the agent never sees it.

Azoma also reports two client outcomes. VentureBeat repeats them from the company; neither the release nor other launch coverage I found audits them. Ruroc's founder says the brand is consistently the #1 most recommended ski and snowboard helmet brand in ChatGPT across multiple geographies and that ChatGPT-sourced site traffic rose 14x, which Azoma reports as a client result. Perfect Ted, a matcha brand, says Azoma's AI-search visibility work contributed to +532% year-over-year revenue growth across all channels. I would treat both as vendor-reported case results. I would not import either percentage into a forecast for a different category.

The source layer is still a wide market

Product visibility and brand visibility fail in the same place when the machine is the reader. The reader changed. The pool of sources did not collapse into the brand's own product page. Measured on the Machine Relations Index, release mri_score_v2.1+2026-11-04+94f24cb38da0, the Consumer Products category (1,825 observed runs across six question shapes) has no dominant source. The most-cited, YouTube, appears in 13.7% of runs (250 of 1,825), Reddit in 13.5%, and Forbes in 10.6%, and each question shape's leaderboard holds between 355 and 681 distinct cited domains. Those three show up at the top of that pool as examples, not as the whole pool.

That release is why I would refuse the sentence "we implemented AMP, so the agents have us." A canonical record that never leaves the vendor's pipe is a governed document. A recommendation is a retrieval event across a lot of domains. If a category's citations are spread across hundreds of domains and no source appears in more than about 14% of runs, one integration is the start of distribution, and the distribution still has to land on pages an agent will fetch. I use Machine Relations for that stack, earned authority through measurement, and the glossary when I need the term for a layer kept stable.

Reported results are a reason to ask for the denominator. Traffic from ChatGPT at 14x needs a baseline. A 532% revenue figure needs the period, the channel, and what else changed in the business. The launch reporting does not supply that audit. I would ask for it in a contract before I let either number justify a category bet.

How I would brief the two numbers that get misused

The trillion-dollar McKinsey range and the 14x Ruroc figure do different jobs, and I would not let them share a slide. McKinsey's projection is a ceiling on how much U.S. B2C retail could be orchestrated by agents by 2030, up to $1 trillion, with a global range of $3 trillion to $5 trillion. Morgan Stanley's companion estimate is about the share of online retail inside that shift: $190 billion to $385 billion of U.S. e-commerce spending by 2030, a likely 10% share and up to 20%. Neither figure names a protocol, a domain, or a SKU. I would use them to justify paying attention. I would not use them to justify a vendor.

The October 2025 Riskified shopping-journey split is the figure I would actually operate. Seventy-three percent of consumers in that survey use AI somewhere in the journey, and the uses are still mostly upstream of a purchase: 45% for ideas, 37% for review summaries, 32% for price comparison. Thirteen percent complete a purchase after an AI referral. Seventy percent are at least somewhat comfortable letting an agent buy. Comfort and completion are far apart. The agent is assembling a description, often from review and comparison pages, and only sometimes closing. A protocol that governs the closing payload and never reaches the review page is aimed at the smaller behavior. I would fund the description the 45% and the 37% are reading, and I would treat checkout governance as a separate operations project.

Ruroc's 14x ChatGPT traffic and the matcha client's 532% year-over-year revenue are Azoma's client figures. Trade coverage of the March 12 launch reports the adopter list, L'Oréal, Unilever, Mars, Beiersdorf, and Reckitt, and the design decision to sit above ACP and UCP. It does not audit those two lifts. In a memo I would put the lifts in a column labeled vendor-reported, with the baseline missing, and I would put the adopter list in a column labeled observed. Mixing the columns is how a case study becomes a forecast. None of that is a reason to import 532% into a category that does not sell matcha.

The Index release I would attach to the same memo is mri_score_v2.1+2026-11-04+94f24cb38da0. In Consumer Products, no source appears in more than 13.7% of runs, and each question shape draws on 355 to 681 distinct domains. That spread is the argument against a single feed as a visibility strategy. If I cannot name, after thirty days, which of those domains received our facts, the protocol has governed a record and has not entered the retrieval pool. Google's Universal Commerce Protocol announcement on January 11, 2026, and the Commercetools note on agentic-commerce trends from January 8, are context for why every commerce team heard this pitch in the same quarter. Context is not a measurement of our SKU.

What I would do

I would split the work into two budgets.

The catalog budget pays for a single merchant-side record: attributes an agent can extract, regulatory constraints, and a rule for what may not be said. For a regulated product, that record is the point of a protocol. Inaccurate agent-generated claims are a legal exposure in regulated categories such as supplements; VentureBeat describes RegGuard as auditing content against rules like FDA/DSHEA standards. I would buy that governance if the company cannot keep one product fact consistent across the surfaces it already publishes.

The visibility budget pays for the sources in the retrieval pool. I would take the buyer questions that already lead to a recommendation, record which domains the answer cites, and put structured, checkable product facts on the domains that recur. I would not accept a syndication promise that cannot name those domains after a month. OpenAI's ACP and Google's UCP, announced in this sequence with Google's Universal Commerce Protocol post on January 11, 2026, are connections. Connections move a payload. They do not choose the sentence the shopper hears.

Scope: this is a call for companies whose buyers ask an agent what to buy, and whose product facts are currently split across a site, a marketplace feed, and uncontrolled affiliate pages. It is not a call to abandon the product page. The page remains the record a human can open. The agent, on the evidence in this launch, often does not.

What would change my mind: a published measurement, with a window and a denominator, showing that recommendations in my category resolve from the merchant protocol's record more often than from the open-web domains in that Index release. If agent answers started citing the governed feed directly, and the 355-to-681 domain spread per question shape narrowed to a feed I control, I would move the visibility budget into the protocol. Until that measurement exists, I keep the protocol in operations and the citation work in the source layer.

FAQ

Does a merchant protocol replace product pages?

No. ACP, UCP, AP2, and MCP connect agents to data, payments, and context. AMP, as Azoma describes it, is a merchant-side layer above those protocols. The product page is still the human-readable record. The protocol is a claim about governing the facts an agent may repeat, and about pushing those facts to third-party sources.

Are the L'Oréal and Unilever adoptions proof that agents now recommend from the protocol?

They are proof that five large CPG companies, L'Oréal, Unilever, Mars, Beiersdorf, and Reckitt, adopted the protocol quickly. The 14x ChatGPT traffic figure and the 532% revenue figure are Azoma's own client reports. Trade coverage of the March 12, 2026 launch reports the adopter list and the design, not an independent audit of those outcomes.

Should a smaller catalog adopt the same protocol?

I would adopt the governance piece if inconsistent product facts are already causing bad answers, and if the company cannot staff a separate integration for every agent surface. I would not adopt it as a substitute for the sources agents retrieve. The Index release above puts the most-cited Consumer Products source in 13.7% of runs, with 355 to 681 distinct domains per question shape. One feed does not cover that spread by existing.

What would I ask a vendor before signing?

I would ask which domains received the catalog in the last thirty days, which buyer questions cite those domains, and which of the reported lifts were measured by someone other than the vendor. I would also ask what the compliance check blocks. A protocol that distributes a claim the legal team would not sign is a faster way to be wrong.