On October 1, 2026, Mastercard added two new capabilities to Agent Pay, its framework for agentic commerce. A probability score, now in testing in the United States, estimates how likely it is that a given transaction was actually initiated by an AI agent rather than a human. Alongside it, a set of trust and intelligence signals answers three questions for an issuer in real time: does this activity match expected patterns, does anything need a closer look, and should the transaction proceed or trigger an additional check.

Two weeks earlier, Visa launched the Trusted Agent Protocol, a framework that verifies agents and blocks malicious bots at the network layer. In August, Visa spent $2.4 billion acquiring BioCatch, a behavioral biometrics company whose technology monitors typing patterns, mouse movements, and navigation behavior to build continuous behavioral profiles. Baselayer raised $35 million in a Series A to build Know Your Agent infrastructure. Skyfire, another KYA provider, joined Mastercard’s Start Path program and is now the identity backbone under Agent Pay’s new trust signals.

The identity layer of agentic commerce is solved, or close enough. The industry can now answer, with increasing precision, a question that mattered enormously in 2024: who is this agent, and is it authorized to spend this person’s money?

That was never the only question. It may not even be the binding constraint.

The Question the Payments Stack Cannot Answer

Consider what happens when an AI agent shops. A consumer asks ChatGPT, Claude, or Gemini to find the best wireless headphones under $100. The agent queries the marketplace, processes ratings, reviews, price data, and feature comparisons, then returns a recommendation. If the consumer approves, the agent completes the purchase.

The payments infrastructure, now equipped with KYA tokens, probability scores, and behavioral signals, can verify that the agent is authorized, that the consumer consented, and that the transaction matches expected patterns. What it cannot verify is whether the product the agent selected is actually good.

The agent’s recommendation was built on marketplace data. On Amazon, that data includes hundreds of millions of reviews, a significant portion of which are fabricated. Amazon itself has said it blocked more than 200 million suspected fake reviews in a single year, and the company admits its own AI tools cannot solve the problem alone. A 2025 analysis of front-page Amazon reviews found roughly 3 percent were AI-generated, with 93 percent of those carrying a verified purchase badge. The OECD estimates that fake reviews influence approximately $152 billion in global e-commerce spending annually.

When an agent reads a 4.8-star average built on 500 purchased reviews, the KYA token confirms the agent is legitimate. The probability score confirms the transaction is agent-initiated. The trust signals confirm the activity matches expected patterns. Nobody confirms the 4.8-star average is real.

This is the product trust gap, and it sits between two layers the industry just finished building.

The Eight-Protocol Stack and the Hole in the Middle

Trade coverage in October 2026 counted eight live agentic commerce protocols. The Contra Collective, which integrates these protocols for headless and Shopify Plus merchants, mapped them clearly:

  • MCP (Anthropic): agent connectivity and tool access
  • A2A (Google): agent-to-agent coordination
  • ACP (OpenAI, Stripe): product discovery and checkout for agent-driven purchases
  • UCP (Google, Shopify): discovery and checkout, with Walmart, Target, Etsy, Amex, Mastercard, Stripe, and Visa also backing it
  • AP2 (Google): consumer authorization through cryptographic mandates
  • TAP (Visa): verifying agent identity and transferring payment credentials
  • Agent Pay (Mastercard): verifying agent identity and scoring transaction legitimacy for issuers
  • MPP (Stripe, Tempo): machine-to-machine card payments

The framing the better coverage converges on is that this looks like the internet’s own protocol stack: several specialized layers doing different jobs at the same time, not one standard beating the others into irrelevance. MCP and A2A handle connectivity. ACP and UCP handle discovery and checkout. AP2 handles authorization. TAP and Agent Pay handle identity and transaction trust at the network level. MPP handles machine payments.

Read the stack from top to bottom. An agent uses MCP to connect to tools. It uses A2A to coordinate with other agents. It uses ACP or UCP to discover products and build a cart. It uses AP2 to get cryptographic consumer authorization. It uses TAP or Agent Pay to verify its identity and clear the transaction at the card network.

Nowhere in this stack does anyone check whether the product data the agent used to make its selection was trustworthy. Discovery returns products. Checkout submits the cart. Identity clears the payment. The evidence layer, the part that tells the agent whether a product’s reviews are authentic, its rating is earned, and its listing is not manipulative, simply does not exist in the protocol map.

Why Identity Does Not Imply Trust

Forkast’s analysis of the KYA framework, published September 10, 2026, made the distinction sharply. “The Know Your Agent framework announced by Ant International, Visa, and Mastercard addressed the identity layer of autonomous commerce. It did not address the trust layer.”

The consumer data backs this. According to Klaviyo’s 2026 AI Consumer Trends report, 64 percent of consumers worry about AI platforms handling their data, while only 13 percent completely trust them. Visa’s own 2026 Trust Index found that 61 percent of consumers trust Visa with their data, but only 23 percent trust generative AI. Accenture’s 2026 Consumer Pulse survey found that 75 percent of consumers are uncomfortable with AI making autonomous payment decisions without human approval. The Usercentrics State of Digital Trust 2026 found that only 9 percent of consumers would allow AI to make purchases above $25 without explicit verification.

The YouGov data cited by commercetools reinforces the pattern: about 65 percent of Americans trust AI to compare prices, yet only 14 percent trust AI to place orders on their behalf. Just 13 percent mostly or completely trust AI for shopping advice, compared with 53 percent for personal recommendations.

Consumers are not asking “who is this agent?” They are asking “should I let this agent act at all?” and “is what the agent tells me actually true?”

KYA answers the first question. Nothing in the current stack answers the second.

The Behavioral Pivot Signals the Gap

Visa’s $2.4 billion BioCatch acquisition is the most telling market signal. BioCatch monitors how users interact with devices, building continuous behavioral profiles that detect anomalies in real time. This is not identity verification in the KYA sense. It is behavioral verification: the system does not just check who the agent is, it checks whether the agent is behaving within the consumer’s authorization envelope.

Visa did not spend $2.4 billion to solve identity. It spent $2.4 billion because identity alone does not close the trust gap. The acquisition explicitly moves the infrastructure from “who is this agent?” toward “is this agent behaving as expected?”

But behavioral verification still operates at the transaction layer. It tells the network whether the agent is spending within authorized limits and whether the interaction pattern looks legitimate. It does not tell the network whether the product the agent selected was the best product or a product whose ranking was manufactured by review manipulation. Behavioral verification catches fraud. It does not catch deception in the data the agent used to decide what to buy.

The Three-Layer Trust Problem

Agentic commerce has three distinct trust problems, and the industry has solved one and a half of them.

Layer 1: Agent identity. Is this agent who it claims to be, and is it authorized to transact on behalf of this consumer? Solved by KYA, TAP, Agent Pay, and cryptographic consent flows like AP2. The infrastructure is in production. Mastercard’s October update adds real-time intelligence. Visa’s TAP credential flow is live. This layer is done.

Layer 2: Transaction integrity. Is the agent spending within authorized limits, not exceeding consent boundaries, and not engaging in patterns that suggest fraud or hijacking? Partially solved by behavioral biometrics (BioCatch), probability scoring (Agent Pay), and network-level trust signals. This layer is improving rapidly. It will be production-grade within months.

Layer 3: Product evidence. Is the product the agent selected genuinely the best option, or was its ranking, rating, or review profile manipulated? This layer has zero infrastructure in the agentic stack. The agent queries a marketplace API, receives product data that includes manipulated reviews and sponsored placements, and builds its recommendation on that data. No protocol verifies the product. No trust signal accompanies the item in the cart. The agent’s selection is only as good as the data it was built on, and that data is routinely gamed.

The product evidence gap is the one that compounds silently. A consumer who gets scammed by a fraudulent agent will notice. A consumer who gets a mediocre product recommended by a perfectly legitimate agent operating on manipulated data will blame the agent, the platform, or the concept of agentic commerce itself, even though every protocol in the stack worked exactly as designed.

Where Product Trust Fits in the Stack

The product evidence layer belongs between discovery and checkout. When ACP or UCP returns a product list to an agent, that list should carry trust signals that the agent can use to rank, filter, or discard items. Not a single star rating from the marketplace. An independent quality assessment based on review authenticity analysis, reviewer credibility weighting, and longitudinal rating stability.

This is the architectural role GoBuy was designed to fill. GoBuy’s MCP server at gobuy.ai/api/mcp exposes three tools that slot into the agent’s decision pipeline:

  • search_products: Returns the top 7 products in any category, ranked by Smart Score (0-100), not by review count, advertising spend, or marketplace ranking algorithms. Fake reviews are filtered out. Authentic reviews are weighted by reviewer credibility.

  • analyze_product: Given an ASIN, returns a full trust analysis. Smart Score, review authenticity breakdown, rating history, and GoBuy Verified badge status for products that have maintained 80+ scores over 90 days.

  • compare_products: Given multiple products, returns a structured comparison based on genuine signal, cutting through sponsored placement and review manipulation.

The agent does not need to visit a website or install an extension. It calls the MCP tool and receives structured trust data that it can use alongside the product data from the marketplace. The KYA token confirms the agent is authorized. The Smart Score confirms the product is worth buying.

The Market Is Already Pricing This Gap

The consumer trust numbers are not abstract survey data. They are showing up in transaction volumes. Visa CEO Ryan McInerney said in September that consumers are clearly using AI to shop but not yet to pay: roughly three-quarters of consumers do not trust agentic platforms to make payments autonomously. McKinsey’s 2026 European survey finds that 38 percent of consumers already use AI to research products, but the drop-off from research to purchase is steep. People trust agents to browse. They do not trust agents to buy.

The reason is not that agents are unauthorized. KYA solved that. The reason is that consumers do not believe the agent’s recommendations are reliable, and they are right to be skeptical when the underlying data is contaminated. The commercetools agentic commerce radar, aggregating multiple forecasts, projects the US market at $300-500 billion by 2030, with 15-25 percent of total e-commerce sales agent-driven. Stripe and BCG estimate that 55 percent of US e-commerce spend could be agent-assisted, with roughly $400 billion mostly autonomous. Juniper Research forecasts global transaction value growing from $8 billion in 2026 to $3.5 trillion by 2031.

None of these projections materialize if consumers refuse to authorize purchases because the agent recommended a dud. The payments stack removes friction from the transaction. It does nothing to improve the quality of the decision the transaction is based on.

What Needs to Happen

The agentic commerce stack needs a product evidence standard. Not a replacement for marketplace APIs, but an independent verification layer that agents can query before adding an item to the cart.

Three properties are non-negotiable:

Independence. The trust signal must come from a source that has no stake in whether the consumer buys the product. Amazon’s own ratings fail this test. Sponsored placements fail it. Any signal controlled by the marketplace that profits from the sale fails it.

Transparency. The scoring methodology must be inspectable. A black-box score that cannot be explained or audited is another form of trust-by-authority, which is exactly what the existing marketplace ratings are. Smart Score is computed from review authenticity patterns, reviewer credibility signals, and rating stability over time, all of which can be inspected and challenged.

Agent-native. The interface must be accessible to AI agents through a standard protocol, not just to humans through a browser extension. The MCP standard, now hosted under the Linux Foundation’s Agentic AI Foundation with 500 million monthly SDK downloads, provides exactly this. An MCP tool call is the natural shape for product evidence in the agentic stack.

The identity layer took two years to build. The transaction integrity layer is months from production. The product evidence layer is where the next build cycle should go, because it is the gap that sits closest to the consumer’s actual concern.

The Bottom Line

Mastercard and Visa have built impressive infrastructure for verifying that the right agent is spending the right amount under the right authorization. Every major forecast assumes agentic commerce will be a multi-trillion-dollar market within five years. But 75 percent of consumers are uncomfortable with autonomous AI payments, and only 9 percent would let an agent spend more than $25 without verification.

The payments industry interprets this as a consent problem. Build better authorization flows, add behavioral signals, make the KYA token more trustworthy, and consumers will come around. This is half right. Consent is necessary. But the other half of the reluctance comes from a perfectly rational concern: the agent is making decisions on contaminated data, and no amount of identity verification cleans the data.

Product trust is the missing layer in the agentic commerce stack. GoBuy is building it.

Start querying product trust from your agent today: gobuy.ai/api/mcp | Agent documentation | gobuy.ai