Agentic Commerce Has a Product-Fit Problem

Visa’s Commerce Benchmark puts a neglected question at the center of AI shopping: can an agent establish what the customer needs before recommending what to buy? For brands, the answer changes what it means to be ready for agentic commerce.

A shopper wants a new television. An AI assistant finds a well-reviewed model, checks the price and recommends it. The product appears to satisfy the request.

But the shopper has not mentioned that videos from their child’s phone look washed out on the current television. They do not know that this detail could determine which replacement is suitable. Unless the assistant asks the right questions, an apparently strong recommendation can miss the requirement that matters.

This is the kind of problem explored in Visa Research’s October 2026 whitepaper, Can an AI agent shop with you? Its Commerce Benchmark evaluates a demanding part of the buying journey: turning an incomplete request into an informed selection, or an informed decision not to buy.

For brands and retailers, the implication extends beyond the quality of the shopping model. A recommendation depends on the agent obtaining the right evidence, connecting it to the customer’s situation and carrying those requirements through to the final choice.

Being discovered opens that process. Establishing fit determines whether the product deserves to survive it.

The buying brief is still being written

Visa deliberately starts with a buying request that is still incomplete. Some requirements emerge only when the assistant asks a question, requests a measurement or suggests consulting another household member. The agent must help establish the brief while researching the market. [1]

That reflects a familiar commercial reality. A customer may know the problem without knowing the specification that solves it. They may request a quieter appliance without understanding which operating conditions matter. They may ask whether clothing will fit without recognizing that the label’s sizing system differs from their own.

The product can be excellent and still be wrong for the particular buyer. Equally, it can be suitable and lose consideration because the evidence needed to establish that suitability never enters the conversation.

For a merchant, the second failure is a lost commercial opportunity: a product that could serve the customer never receives a properly informed assessment. Improving its visibility alone may leave the deciding question unanswered.

A correct recommendation has to be earned

One of the benchmark’s most consequential rules is that selecting a suitable offer by chance does not count as success. The assistant must have established the customer-held facts needed to justify the decision. A run can also fail if the assistant misinforms the customer, even when it reaches the right selection. [1]

This raises the standard for evaluating AI shopping. A plausible answer, a confident explanation or a familiar brand name cannot establish that the decision was sound. The evidence must support the recommendation.

Visa also treats a justified refusal as success when no suitable offer exists. That matters commercially. An agent that recommends an unsuitable product may appear to have helped a merchant win, while creating the conditions for disappointment, cancellation or return. These are potential business consequences beyond the benchmark’s scope.

For a merchant, the useful goal is to win purchases the business can successfully serve. That includes identifying when a product fits, when it fits subject to a condition and when there is no suitable option.

The decisive fact may sit beyond the search result

Visa’s controlled market reflects a difficult information environment. Search summaries tend to present products favorably. Important qualifications may appear only in full listings, seller policies or a small number of reviews. The assistant has limited time and tool calls, so inspecting everything is impractical. [2]

The whitepaper describes failures such as accepting a missing listing value, failing to ask about an unstated requirement or declaring that no suitable product exists while leaving a relevant offer unexamined. Its working hypothesis is that failures often occur before interpretation: the agent never asks the question that would reveal the fact. [3]

For merchants, this creates a practical distinction between publishing information and getting it into a decision.

A delivery restriction can be present on a policy page and still be missed. A compatibility condition can be accurate and still be difficult to connect to a particular use case. A sizing chart can exist without resolving the customer’s uncertainty.

Clear listings, accessible policies and structured product data remain essential. An additional opportunity is to make product expertise available when a buying agent encounters a question that those materials have not resolved.

Where an on-site product expert can help

This is the role AIVO Assist is designed to address: an on-site product expert that answers product and policy questions, asks for missing specifics, supplies deciding facts and sources, and explains when an option is unsuitable.

Consider an illustrative exchange. A buying agent asks whether a dress is appropriate for a customer who normally wears a UK size 12 and needs delivery in London. A useful product expert needs to establish the relevant sizing evidence and check the delivery destination. If the merchant does not ship to the UK, the correct answer is to say so. Resolving the question may prevent a purchase from proceeding.

In another case, the destination is supported but sizing remains uncertain. The useful response is to request the relevant measurements and compare them with the available garment information. A confident conversion without sufficient evidence would merely move the uncertainty further down the journey.

The buying agent is responsible for discovering the shopper’s requirements and comparing alternatives across merchants. Assist contributes the merchant’s product and policy expertise, and can request details needed to assess fit. It cannot resolve a customer requirement the buying agent never discovers or shares. The exchange works when the agent supplies the relevant context and the merchant supplies dependable evidence.

Measuring the opportunity AIVO can address

AIVO’s AI Win-Rate approach focuses on final recommendations across buying conversations. The benchmark suggests a useful companion question: was the recommendation justified by the customer’s requirements?

Combining commercial outcome with decision quality makes the diagnosis more actionable.

This proposed framework complements Visa’s model evaluation with a merchant perspective. Suitability alone does not entitle a product to win: another suitable offer may provide better value. Diagnosing an evidence-related loss requires support from the conversation trace or a controlled comparison.

For AIVO, it connects measurement to remediation. A missing delivery policy calls for a different intervention from an unanswered compatibility question. A misunderstood product attribute requires a different response from a genuine absence of a suitable offer.

Meridian’s diagnostic capabilities and AIVO’s remediation work can help identify and address these evidence gaps. Assist adds a way to answer questions during the investigation itself. Repeated measurement can then assess whether the change affects the final recommendation and its justification.

Assist’s trackable answer codes can support this analysis by identifying specific responses. Establishing that a response caused a recommendation or sale requires additional observation and attribution; the code alone cannot demonstrate causality.

The improvement has to be demonstrated

The practical test is whether access to an on-site expert improves decisions across repeated, comparable buying journeys. Making Assist explicitly available tests whether its expertise helps; allowing agents to navigate normally tests whether they discover and use it. Both matter. Evaluation should track justified recommendations, unsuitable recommendations, correct declines and unresolved questions, and examine whether the necessary facts were established and applied. Visa’s emphasis on repeated runs is relevant: success on one attempt does not establish reliability. [4]

Readiness extends into the conversation

Visa’s benchmark evaluates models in a controlled environment with simulated customers. It stops at selection or refusal, before checkout, payment, order placement or fulfillment. Its findings concern decision quality within that environment, rather than live conversion performance. [1, 5]

Its contribution is to make the decision problem explicit. Shopping begins with incomplete information. The right questions reveal requirements. Product and policy evidence determine whether an offer meets them. A capable agent must preserve that understanding through the final choice.

For merchants, agentic readiness includes being able to answer the questions that determine whether a product belongs in the final recommendation. A suitable product can lose when the deciding evidence never reaches the conversation; an unsuitable product can win when a material constraint goes unchecked.

AIVO connects measurement and remediation to that challenge. Assist brings the merchant’s product expertise into the buying conversation. The commercial task is to establish fit, resolve uncertainty and measure whether better evidence leads to better decisions.

Source notes

Visa Research, Can an AI agent shop with you? Visa Commerce Benchmark, October 2026. Page references below use the printed page numbers in the supplied whitepaper.

1. Objectives, scope and correct-decision rules: pp. 3–5 and 11–12.
2. Market structure, information access and research limits: pp. 6–9.
3. Difficulty, error attribution and failure patterns: pp. 14–15. The account of where failures tend to occur is explicitly presented by Visa as a working hypothesis.
4. Reliability and repeated evaluation: p. 13.
5. Evaluation practice, current limitations and planned extensions: pp. 16–17.

Disclosure: The author is CEO and co-founder of AIVO, Inc. The discussion of AIVO and AIVO Assist is the author’s interpretation of the research and a description of AIVO’s approach. Visa’s benchmark does not evaluate or endorse AIVO or AIVO Assist.