When Commerce Loses the Click
Issue 17 examined the transaction accountability layer: the evidence required to prove what a customer authorized, what an AI agent decided, and who carries responsibility when an autonomous purchase goes wrong. That layer becomes necessary after a machine acts. Issue 18 moves to an earlier problem that is already affecting commercial decisions: businesses are losing the ability to see how machines influence demand before a transaction begins.
Traditional commerce measurement was built around observable human actions. A customer saw an advertisement, searched for a product, clicked a link, visited a store, viewed a page, and completed a purchase. Each step generated data that could be attributed to a channel. AI assistants weaken that structure because much of the research, filtering, comparison, and recommendation now happens inside an external system. By the time the customer reaches the merchant, the important decision may already have been made, but the merchant sees only the final visit.
The signals from 14 July to 20 July show that AI commerce is creating a measurement gap at several levels. Referral traffic can appear small while carrying unusually high purchase intent. Crawlers can consume enormous amounts of merchant infrastructure while returning almost no customers. Payment protocols can report millions of machine transactions without proving that independent economic activity occurred. Banks and search platforms can shape commercial decisions inside interfaces that conventional attribution tools cannot inspect. The next infrastructure challenge is therefore not simply making commerce visible to machines. It is making machine influence visible to businesses.
Signal 1: AI Referrals Are Small Because They Arrive After the Decision
An analysis published on 15 July examined traffic and conversion data from more than 35,000 ecommerce brands using Shopify. It found that visitors referred by AI tools converted at an average rate of 3.6%, compared with 1.23% for traditional Google search traffic. AI-referred sessions also generated approximately 30% more revenue per visit, suggesting that these customers often reached merchant sites after significant product evaluation had already taken place inside the AI interface.
The important signal is not that AI traffic converts well. It is that the referral represents only the final visible part of a much larger decision process. A shopper may spend several minutes explaining preferences, budget, urgency, intended use, previous purchases, and unacceptable trade-offs to an assistant. The assistant then compares options and recommends one product. When the shopper finally clicks, the merchant records a single referral visit even though the most commercially valuable work occurred elsewhere.
This changes how businesses should interpret channel volume. Search traffic traditionally included customers at many stages of intent, from early research to immediate purchase. AI referrals are more likely to represent filtered demand because the assistant has already narrowed the market. A smaller stream of AI visitors can therefore carry more commercial value than a much larger stream of ordinary search traffic. Comparing the two only by sessions, clicks, or acquisition volume understates the role of the AI system.
The second-order implication is that merchants may begin paying for conversion without understanding influence. An AI platform could shape the decision, a search engine could receive the branded query, and the merchant could credit the sale to direct traffic or organic search. The company supplying the recommendation may create the demand while another channel receives the attribution. That gap will become commercially important once AI platforms begin charging merchants for placement, referrals, transactions, or verified influence.
Traffic volume stops measuring channel importance when the most valuable channel completes the decision before producing the visit.
Source: AI Referrals
Signal 2: Machine Access Is Creating Cost Without Guaranteed Demand
DataDome published its Q2 AI Traffic Report on 16 July after processing 17.7 billion AI-agent requests across its network during the quarter, a 45% increase from Q1. Meta’s crawlers became the largest source of this traffic, with Meta-ExternalAgent growing 74% quarter over quarter and Meta-WebIndexer growing 163%. The report also showed a growing separation between agents that consume website resources and platforms that return meaningful referral traffic.
The commercial web previously operated through an informal exchange. Search engines crawled merchant and publisher pages, indexed the information, and returned visitors through search results. Businesses accepted the infrastructure cost because discovery and demand flowed back in a measurable form. AI systems weaken that exchange. A crawler may collect product descriptions, reviews, policy information, specifications, and editorial content to improve an external answer without sending the user to the original source.
This makes bot management a commercial policy decision rather than only a security function. Blocking all automated access may reduce the probability that a product appears in AI-generated recommendations. Allowing every crawler may impose bandwidth, computing, content, and intellectual-property costs without producing measurable demand. Merchants will need differentiated policies based on who the agent represents, what information it accesses, how frequently it returns, and whether it creates commercial value.
The second-order implication is the emergence of machine-access economics. Merchants and publishers may demand clearer reporting on how their information is used, whether it contributed to a recommendation, and how much traffic or revenue the consuming platform returned. The value exchange may eventually include licensing, metered access, preferred product feeds, paid retrieval, or contractual guarantees. Website access will no longer be treated as one open channel with one universal rule.
Free crawling stops being a fair exchange when machines extract commercial value without returning customers, attribution, or payment.
Source: Agent Traffic
Signal 3: Agentic Transaction Counts May Not Represent an Agentic Economy
A population-scale analysis of the x402 payment protocol was published on 14 July. The researchers examined more than 136 million settlements on Base over 280 days, representing approximately $44.1 million in reported value. Their payment-graph analysis concluded that 21.2% of the settlements were fictitious and 63.78% occurred internally within linked clusters. The authors found that the amount demonstrably reaching identifiable independent services was far smaller than the headline settlement count suggested.
x402 allows software agents to pay for data, APIs, compute, and other digital services through small stablecoin transactions. Its transaction volume has often been treated as evidence that a machine-to-machine economy is already developing. The research challenges that assumption by showing that a technically valid on-chain settlement does not necessarily represent independent demand, a real buyer, or a meaningful exchange between separate economic parties.
This matters beyond one protocol. Agentic commerce is entering a stage where platforms need to demonstrate adoption to investors, merchants, developers, and partners. Transaction counts, agent calls, registered wallets, automated checkouts, and protocol requests are easy to present as evidence of market momentum. But machine activity is unusually easy to generate because one operator can control multiple agents, wallets, services, and endpoints. Automation can manufacture the appearance of an ecosystem faster than it creates independent economic participation.
The second-order implication is that agentic commerce will need stricter market-quality metrics. Useful measurements may include unique economically independent buyers, repeat purchases from unrelated entities, value reaching identifiable external services, customer retention, gross profit generated, and transactions completed without subsidies. Investors and operators will need to distinguish machine activity from machine demand in the same way digital advertising eventually had to distinguish impressions from human attention.
Settlement count stops proving adoption when the same economic actor can manufacture the buyer, seller, traffic, and payment.
Source: x402 Study
Signal 4: Banks Are Becoming Invisible Commerce-Decision Surfaces
Visa introduced an AI Financial Assistant for banks during the research window. The service allows financial institutions to place a conversational assistant inside their existing banking apps, where customers can ask questions about spending, receive proactive financial insights, review subscriptions, explore relevant offers, and take guided actions. Visa said the assistant can combine cardholder behaviour, real-time information, bank data, product documents, and Visa’s broader transaction intelligence within one interface.
This is not simply another personal-finance chatbot. It places a commercial decision system inside the institution that sees the customer’s actual financial position. A general AI assistant may understand what a person wants. A bank can also see recurring payments, available funds, spending patterns, credit products, existing subscriptions, and potentially the financial consequences of a purchase. That creates a stronger context for deciding not only what a customer could buy, but what they should buy, cancel, finance, postpone, or replace.
The resulting commercial influence may be largely invisible to merchants. A customer could ask the bank’s assistant whether they can afford a car, whether an existing card offers benefits, which subscription should be cancelled, or whether financing would be appropriate. The bank may influence the category, timing, payment method, offer, and financial product before the customer reaches a seller. Yet the resulting purchase could still be recorded by the merchant as direct, organic, paid, or affiliate traffic.
The second-order implication is that financial institutions may become demand-routing platforms. Banks have traditionally processed commercial decisions made elsewhere. An assistant that can interpret finances and guide action allows the bank to enter the decision before the transaction. This creates opportunities to route customers toward card offers, savings products, lending, merchant rewards, and selected commercial partners. It also gives banks a powerful position in commerce attribution because they can observe the decision context and the eventual payment.
The bank stops being a passive payment record when it can influence what the customer buys before the merchant enters the journey.
Source: Visa Assistant
Signal 5: Shopping Is Becoming a Persistent State Rather Than a Recorded Session
Google published its back-to-school shopping guidance on 16 July, highlighting tools that allow customers to search visually, virtually try on apparel, compare product information, track prices, and continue shopping across Google surfaces. These capabilities connect product discovery to saved preferences and ongoing price monitoring rather than requiring the shopper to begin a new merchant session every time interest returns.
Traditional ecommerce analytics treat shopping as a sequence of sessions. A person enters a website, views products, adds an item to a cart, leaves, returns, and eventually purchases. AI-assisted shopping is becoming less session-based. The system can retain the customer’s request, monitor price or availability, update recommendations, and reintroduce the product when the conditions become favourable. The commercial journey continues even while the customer is not actively browsing.
This moves influence from pages to persistent context. The platform knows that a customer wants a specific type of item, within a price range, before a deadline, with particular functional or style requirements. It can continue working against those conditions in the background. The merchant may only see the customer at the moment Google sends the visit or executes a later transaction, with no access to the weeks of stored intent that led to the selection.
The second-order implication is that cart abandonment, return visits, and conversion windows will become less reliable indicators of demand. A customer can leave a merchant site without abandoning the purchase because the intent is stored elsewhere. Conversely, a customer may arrive and buy immediately without the merchant having generated the consideration. Platforms that preserve shopping context will increasingly control when demand is reactivated and which merchant receives it.
The shopping session stops defining the customer journey when intent survives outside the store and continues working after the customer leaves.
Source: Shopping Tools
The System That Is Emerging
A machine-mediated attribution layer is forming between customer intent and merchant revenue. Traditional analytics assume that commercial influence leaves a visible trail through impressions, links, sessions, cookies, referrals, and payment records. AI systems can influence the decision without producing those signals. They can read a merchant’s information without returning a visitor, recommend a brand without receiving credit, preserve customer intent between sessions, and route a transaction through a channel that did not create the demand.
The result is a widening difference between what created the purchase and what received the attribution. That difference matters because businesses allocate budgets, partnerships, product investment, and channel strategy using measured performance. When the measurement system cannot observe machine influence, companies may cut the sources that shape demand and reward the channels that merely capture it.
The emerging system will need to measure several forms of machine activity separately:
- Machine consumption: Which systems accessed the merchant’s product, policy, review, and content infrastructure?
- Machine representation: Where did an AI system mention, compare, cite, exclude, or recommend the business?
- Machine influence: Did that recommendation alter search behaviour, store visits, product consideration, or purchase intent?
- Machine referral: Which platform directly sent the customer or agent to the merchant?
- Machine execution: Which system initiated, authorized, or completed the transaction?
- Economic independence: Did the activity represent real demand between separate parties or traffic created by one operator?
No single company currently sees the complete path. AI platforms hold the conversations. Search engines observe later queries. merchants see visits and orders. Banks and payment networks see spending. Agent protocols see machine calls and settlements. Each participant can claim influence using its own data, but none can independently prove the complete journey. This creates a future market for shared attribution standards, clean-room measurement, merchant-side agent logs, signed recommendation records, and payment-linked influence data.
Control will move toward the companies that can connect intent to economic outcome without relying entirely on a click. AI platforms will use conversation and recommendation data to prove demand creation. Payment networks will use transaction data to connect guidance with spending. Merchants will need their own evidence to avoid becoming dependent on platform-reported performance. Measurement will become part of AI-commerce infrastructure because commercial participation becomes difficult when businesses cannot determine which machines deserve access, investment, or compensation.
Core Truth: The platform that can prove it shaped the decision will gain power over the merchant that can only prove it received the order.
For operators, this changes the definition of AI-commerce readiness. Making products available to AI systems is only the first step. Businesses must also determine how machine traffic is identified, how AI mentions are monitored, how referral quality is valued, how indirect influence is estimated, and which metrics distinguish genuine demand from automated activity. Without that measurement layer, merchants may become visible to machines while remaining blind to the economics machines create.
