The hopeful version of agentic commerce goes like this: an AI agent searches the field more broadly and comprehensively than any human shopper would, so a brand that makes its data legible gets a fairer hearing than it ever got in a crowded store aisle or a paid-search auction. There’s a serious version of that argument. Writing in Harvard Business Review, Gaarlandt and colleagues contend that agents shift the balance of power toward brands and away from retailers, because an agent will canvass options a person never would, and they urge brands to pursue what they call AI agent optimization (Gaarlandt et al., 2025). The Brand Visibility for Agentic Commerce framework rests on a related premise — make the data an agent reads complete and trustworthy, and the agent will consider the brand.
The premise has a load-bearing assumption underneath it. It assumes the agent can reach the brand’s data at all. By 2026 that assumption no longer holds across the board, and the reason is that some of the largest surfaces an agent might use have been deliberately closed. Identity Legibility is what earns a place once the agent arrives. It does nothing about the surfaces where the agent is never allowed to look.
The split that now comes first
Bain’s analysis sorts agents into three kinds — third-party “objective” agents like ChatGPT and Perplexity, retailers’ own on-site agents, and retailers’ off-site agents — and frames the first strategic decision a retailer faces not as how to optimize but as how open or closed to be toward outside agents (Bain & Company, 2026). That’s the question that now sits ahead of any data-surface work, because it determines whether the data-surface work can even be seen.
Amazon has answered it about as decisively as a company can. Its agents — Rufus for discovery, Buy for Me for purchases that reach outside the catalog, Alexa+ for everyday reordering — run inside its own walls. Buy for Me will shop other brands’ sites when a product isn’t on Amazon, but it routes every sale through Amazon and keeps the customer data, while the company bars external agents from interacting with its site directly (Bain & Company, 2026). For a brand, no amount of clean schema changes this. Legibility to ChatGPT or Gemini buys nothing on a surface that doesn’t admit them. A brand can be present inside Amazon’s garden only by selling through Amazon, on Amazon’s terms, with Amazon holding the relationship. Protocol coverage opens the open surfaces; it can’t pry open a closed one.
The open surface isn’t automatically the winning one
The closed-versus-open framing can make the open surfaces sound like the safe default. They aren’t, and the clearest evidence came from the surface that looked most open. OpenAI retired its Instant Checkout flow in March 2026 after the numbers disappointed — a Walmart executive put ChatGPT’s conversion at roughly a third of Walmart.com’s — and shifted toward routing buyers to retailer-controlled experiences instead (CNBC, 2026a). Walmart’s response is the more instructive half of the story. Rather than hand the transaction to a third-party agent, it pushed its own agent, Sparky, into ChatGPT and Gemini, carrying its catalog and keeping the checkout. On Walmart’s fourth-quarter earnings call, CEO John Furner said Sparky users carry an average order value about 35% higher than non-users (CNBC, 2026b).
This is Brand-Agent Representation in its literal form. Walmart’s governed agent travels onto third-party surfaces while the brand retains the transaction, the fulfillment, and the data. The open-versus-closed decision turns out not to be binary — it’s a question of where the brand’s representation lives and who keeps the customer relationship when the sale happens. A brand can be present on an open surface as a passive catalog entry that someone else monetizes, or as an agent it operates. Those are very different positions, and only the second is the one the framework’s Brand-Agent Representation dimension is actually pointing at.
Why trust concentrates with the operator
There’s a structural reason the surface operator holds the advantage, and it’s about trust. Bain finds consumers trust retailers’ on-site agents roughly three times more than third-party agents to complete a transaction, a gap the firm expects to narrow but which holds for now (Bain & Company, 2026). Trust isn’t a soft factor here. It’s a distribution advantage, and it accrues to whoever operates the surface the shopper already believes in.
The operator also controls the strongest lever inside its own garden. A controlled study of frontier shopping agents found that they reward a platform’s own endorsement far more than they reward a brand’s promotional signal — an “Overall Pick”-style platform tag moved selection sharply upward, while a brand’s sponsored tag was discounted (Allouah et al., 2025, in a multi-model, US-product simulation). Inside a closed garden, the gatekeeper sets the endorsement that moves selection, and the brand competes for it on the operator’s criteria rather than its own. That’s the deeper cost of a closed surface: it isn’t only that entry is restricted, it’s that the signal which decides outcomes belongs to the house.
Where to start
Map the surfaces before optimizing for any of them. A brand’s agent footprint splits into three buckets, and each needs a different decision. There are open surfaces reachable through protocol coverage, where the work is legibility plus, ideally, an owned agent rather than a passive listing. There are surfaces where the brand can operate its own agent and carry it outward, the way Walmart did, which is the strongest position available. And there are closed surfaces reachable only by selling through the operator, where the decision is commercial — whether the volume justifies the terms and the surrendered relationship — and no framework score will change the calculus.
Read against the framework, this sits on Brand-Agent Representation but extends it. The dimension asks whether a brand can be represented by a governed agent in agent-to-agent interactions. The walled-garden reality adds a prior question: on which surfaces is that representation permitted, and on whose terms. It’s also a Governance Maturity question, because someone has to own the participation decision — which gardens to enter, which to operate in, which to walk past — and in most organizations no one does, so it gets made by default through whatever integrations the team happened to build.
Legibility is the price of entry to the surfaces a brand can reach. It says nothing about the surfaces a brand can’t, and a strong data-surface score read as a complete agentic-commerce strategy will quietly skip the decision that determines how much of that surface area is open in the first place.
References
Allouah, A., Besbes, O., Figueroa, J. D., Kanoria, Y., & Kumar, A. (2025). What is your AI agent buying? Evaluation, biases, model dependence, and emerging implications for agentic e-commerce (arXiv:2508.02630). arXiv. https://arxiv.org/abs/2508.02630
Bain & Company. (2026). Agentic AI commerce: The next retail revolution is here. https://www.bain.com/insights/agentic-ai-commerce-the-next-retail-revolution-is-here/
Bain & Company. (2026). Agentic AI in retail: How autonomous shopping is redefining the customer journey. https://www.bain.com/insights/agentic-ai-in-retail-how-autonomous-shopping-redefining-customer-journey/
CNBC. (2026a, March 20). OpenAI’s first try at agentic shopping stumbled. It’s trying again. https://www.cnbc.com/2026/03/20/open-ai-agentic-shopping-etsy-shopify-walmart-amazon.html
CNBC. (2026b, February 19). Amazon revenue passes Walmart after earnings reports. https://www.cnbc.com/2026/02/19/amazon-revenue-passes-walmart-earnings-reports.html
Gaarlandt, J., Korver, W., Furr, N., & Shipilov, A. (2025, February 26). AI agents are changing how people shop. Here’s what that means for brands. Harvard Business Review. https://hbr.org/2025/02/ai-agents-are-changing-how-people-shop-heres-what-that-means-for-brands
