The market for machine attention

Meta’s successful launch of Muse in September showed how eager customers are to try these tools. Brands and retailers are scrambling to catch up.

The extent of AI’s influence depends in good part on whether a shopper begins their search with a need or a brand. In our survey, 42% of people who use AI for shopping say they usually start by describing a need and asking for recommendations; 33% begin with a brand already in mind. When the need comes first, the agent can shape which brands enter consideration before the shopper names one.

A brand controls only part of the evidence that AI models use. In our AI-shopping simulation, 88% of cited sources were independent of the brands being recommended. Reviews, marketplaces, publishers, and other third parties can shape the answer alongside a company’s own product information.

That shifts the basis of influence from persuasion to easily identifiable proof that both agents (and humans) can find. While brand building still matters for shaping what consumers know, value, and ask for, once AI begins comparing options, claims need hard facts behind them. The qualities a company wants to be known for — be it price, performance, reliability, sustainability, service, or something else — increasingly need to be supported by credible evidence.

Making the information readable is only the first step. Existing checkout processes, including checkout forms, carts, options, identity checks, payment controls, and confirmation flows, were built around a person clicking through screens. Many merchants also deliberately use bot-detection and anti-fraud tools to identify nonhuman interaction patterns and block them. Agentic commerce requires a more precise distinction, one that enables authorized agents to get to and through the checkout while still blocking malicious automation. Without this, an agent may know exactly what to buy and be unable to complete the transaction.

In our simulated journeys, 79% encountered at least one barrier after the recommendation. Some barriers are essential, especially verification and authorization. But if a legitimate agent cannot prove its authority, understand the required step, or proceed, that’s a failure.

AI’s impact on brand value may vary sharply by category. Where products are easy to compare and switching costs are low, agents can make incumbency less valuable by reopening the choice every time. Insurance renewals are an obvious example: An agent could continuously compare price, coverage, and service rather than defaulting to the existing provider. In other categories, AI may reinforce the value of distinctive brands. In luxury, for example, iconic brands may benefit when heritage, authenticity, scarcity, and resale value are part of the brief, while less differentiated aspirational brands may face more pressure. In travel and service-oriented industries, some elements may become easier to compare and commoditize, while reputation, trust, experience, and service quality become more important precisely because they are harder to reduce to price alone.

Companies should ask which parts of their advantage an agent can compare directly, which it can verify through evidence, and which still depend on trust, experience, or identity. Loyalty will increasingly have to be earned anew at every repeat purchase — but what wins that competition will differ by category.

In the agent-led checkout models now taking shape, these three trust questions converge at the merchant’s front door. Emerging payment protocols are beginning to turn that mandate into portable, cryptographically signed evidence of the customer’s intent, including the purchase or spending boundaries the customer authorized. If unauthorized activity gets through, the seller may still bear operational consequences such as chargebacks, disputes, and fraud losses. That evidence therefore needs to travel with the transaction and remain available if someone later disputes what the agent was allowed to do.

The exposure runs well past payment authorization into account opening, personal data, returns, subscriptions, and obligations written on the assumption that a person was present. Consumers are giving companies deploying AI agents little room for error: 61% say they would either leave the company or stop using its AI features after one serious AI mistake involving their money or personal data.

Several providers may coexist because consumers want different degrees of delegation for different purchases. People express the greatest willingness to hand authority over to AI agents for routine purchases such as groceries and food delivery. They want to retain more control over discretionary items such as air travel and consumer electronics, and are least willing to delegate major decisions such as buying a car or taking out a loan or mortgage. Nearly half are already willing to delegate some spending authority: 49% would let an AI agent spend at least some amount on their behalf without checking before each purchase. That authority can increasingly be encoded into the payment itself. Emerging agentic payment credentials can restrict an agent by spending limit, merchant, purpose, or time window, allowing consumers to give an agent a defined “license to spend” rather than unrestricted access to their money.

Customers may use different agents depending on the type and cost of purchase, which provider they already trust, or how much control they want to retain.

Consumers also care about whose interests the agent serves. Only 24% believe an AI shopping assistant is really working for them when it gives a recommendation. Separately, one in four say they would hold back from delegating some purchases because they worry AI may put brands’ interests ahead of their own. A branded agent can still earn trust; its incentives and mandate will shape how much authority customers are willing to give it.

Sellers need to prepare for both sides of this market: serve customers through their own AI experiences and make products accessible to the third-party agents customers choose. That access cannot be assumed. In September 2026, Amazon blocked Muse from shopping on Amazon.com, telling users that access by an “unauthorized AI agent” violated its terms. The episode points to another layer of competition in agentic commerce: Merchants will decide which agents they recognize, what credentials they require, and what recognized agents are allowed to do. Emerging protocols already contemplate different permissions for activities such as browsing and payment.

The market may therefore develop something resembling a system of digital IDs for agents, mandates, and access rules. An agent can prove who it is; the customer can prove what it is allowed to do; and the seller can decide whether those credentials are sufficient for entry. More commerce may eventually happen through interactions between a customer’s agent and a seller’s agent, including negotiation over service, delivery, loyalty benefits, contract terms, or price within preset guardrails.

A critical strategic question is how to participate in a multi-agent market. Companies need to decide where they want to own the customer interface, which outside agents they will permit or support, what those agents can access, and which commercial and risk rules govern the transaction. Those choices will determine how much of the customer relationship they control as machine buyers become a meaningful route to market.