A follow on blog to my Intent data post yesterday. Where intent is needed for authorization, measurement is needed by every “specialist” participating in an agentic interaction. As background I was founder/CEO of Commerce Signals, focused on measurement and card transaction data. Measurement is a powerful business. In fact, I would say Google started out as a measurement company with the PageRank algorithm. By keeping track of what users clicked on which link for which search word, they created the directory of the internet. Let’s dig a little deeper into why measurement is key in agentic, and for all federated models.

Google is not building a monolithic “central brain” to disintermediate the ecosystem. Instead, as discussed in my UCP Blog (also see Ask Macy’s Case Study), they are fostering a world of specialist collaborative models that interact across three specific technical layers:

1. The Capability Negotiation Layer (UCP): Collaboration begins with Dynamic Discovery via the Universal Commerce Protocol. When a “Super Agent” (orchestrator) receives a request, it queries a decentralized registry to find specialists. The handshake is a “contractual” one: the orchestrator learns a specialist’s capabilities (e.g., loyalty.redemption ) without seeing the underlying model logic.
2. The Context & Tooling Layer (MCP): The exchange of task-specific data happens via the Model Context Protocol. Think of this as the “USB-C” for agentic commerce. The specialist model acts as a server, exposing specific tools (e.g., “Apply Dynamic Discount”) to the orchestrator. This preserves the Specialist Knowledge Gap—the retailer or bank keeps its “secret sauce” while providing the orchestrator with the necessary outcome.
3. The Trust & Mandate Layer (AP2/ACP): This is where models coordinate on Non-Repudiation. The orchestrator generates an Intent Mandate (the user’s goal) and passes it to the specialist, which resolves it into a Cart Mandate (specific SKUs/pricing). These are bundled into a “High-Trust Bundle” signed by a Credential Provider.

See Google’s Enterprise Agent Platform (the evolution of Vertex AI). 

The Access and Knowledge Gap

Within a federated specialst model environment, interaction cannot operate as a public utility; it requires a commercial construct. For a specialist model to participate, it requires four things:

• Access Control/Permissions: A Chase agent, for example, will not participate in a coordination flow or reveal sensitive financial tools without knowing its customer has explicitly permissioned that participation.
• Pricing & Incentives: Specialist models hold unique, siloed knowledge of customer preferences. Retailers like Macy’s or Walmart (as seen in Strategic Innovation Era) will not provide deep-tier pricing or “best offers” to a generic scraper. 
• Commercial terms and governance (ex prevent leakage of insights into the orchestator)
• Measurement (ie what happened)?

Agentic is Breaking the “Closed Loop”

For a decade, the “Data Games” were stable. Retailers shared SKU-level data with Google/Meta to “close the loop” on ad efficacy, as outlined in my 2021 analysis. Agentic commerce breaks these flows (ex agentic tokens).

Each specialist has unique insights into customer preferences, but how that knowledge is applicable to a specific collaboration is currently unknown. Prior to participation, measurement of that specialist’s contribution is required for attribution (value share) and for tuning the specialist’s response. This creates a Measurement Gap. For example Macy’s agent provided “preference” logic that helped close the sale, but how do we value that contribution compared to the orchestrator’s discovery work? These same challenges exist in the advertising world today, where an individual is “touched” by overlapping messages on TV, YouTube, Display, and eMail. Who gets the credit? In agentic commerce, things are much more deterministic, as a fully authenticated consumer directly permissioned a qualified agent to take action. The unknown is no longer who influenced the intent, but which specialist adds value to the sale.

The Economics of Inquiry: Tolls and Bounties

Because specialist compute is expensive, “Capability Negotiation” cannot be free. I see two primary models:

1. Pay Before Inquiry (The Toll – X402): A micro-payment to “wake up” the specialist model. This ensures the orchestrator has “skin in the game” and prevents specialists from being used as free data-scraping tools. 
2. Pay for Success (The Bounty): The specialist absorbs the compute cost but pays an attribution fee upon conversion.

The success event (the purchase) is the ultimate measurement of the orchestrator’s effectiveness and the value of every specialist that contributed. To unlock market forces, each contributor needs to understand the incremental value they created.

The Attribution Gap: The Need for a Neutral Referee

This leads to the most significant point of friction: Measurement. In a “Pay for Success” world, the seller cannot rely on self-attestation of the super agent orchestrator. Google should not be able to  “grade its own homework” by being the sole arbiter of success for its ecosystem, nor should merchants be the only source of truth. By the same logic, neither can the merchant (which is the case today in merchant checkout).

For agentic commerce to scale, measurement and attribution must be handled by a neutral intermediary. This entity must be able to link the initial Intent Mandate to every specialist contributor and the purchase.

Networks are the Logical Orchestrators

In my view the card networks (Visa, Mastercard, Amex) are the only entities that hold the identity, commercial constructs, and governance across all parties today. They are uniquely positioned to act as the neutral attribution engine that maps the transaction back to the specialists who provided the winning logic.

This is all part of the Gordian knot of agentic monetization. Remember that the major retailers are insisting on control of checkout flows. This need for control is not just for cards on file, by owning checkout they also own the commercial terms, lock up the data (on what was purchased), and control the collaboration (who else gets paid).

While Google might be hesitant to have a third party grade its performance, a bi-lateral approach (where every orchestrator negotiates individual data and pricing deals with every bank and merchant) will never scale. The networks provide the only logical place for these operating rules and commercial settlement rails to live. 

This will take two years to iron out. In the meantime, while the technology of data sharing is easy, the competitive and incentive problems remain the largest stumbling blocks. Once resolved, however, this will represent the single largest expansion of Value-Added Services (VAS) for the networks in recent history.