When the Buyer Is a Machine: Agentic AI and the End of the Human Shopper

The shift from recommending to transacting

For two years, the story of AI and commerce was a story about influence. The engine read the reviews, weighed the options, and named a product. A human still did the buying. The agent was an advisor standing next to the shopper, not the shopper.

That line has now been crossed. In 2026 the agent does not just recommend the product. It selects it, adds it to the cart, and completes the payment. The buyer is becoming a machine, and the machine is buying on someone else's behalf.

The proof is no longer hypothetical. During 2025's Cyber Week, AI and AI agents influenced 20% of all orders, accounting for $67 billion in global sales, according to Salesforce. Agent task completion volume surged through the peak. HUMAN Security measured AI agent traffic jumping 144% on Black Friday alone. And the retailers who put agents on their own properties grew US sales at seven times the rate of those who did not, 13% versus 2%, in the seven weeks into Cyber Week.

The agent does not send you a curious browser. It sends you a completed transaction, or it sends nothing at all because it never selected you in the first place.

This is the inversion at the heart of the new commerce. Search optimization was built to win a human's click. The human scanned a page, formed an impression, and decided. Agentic commerce removes the page, removes the impression, and increasingly removes the human from the decision loop entirely. What remains is a machine reading structured data, scoring options, and acting.

The strategic consequence is blunt. Being selected by a person and being selected by an agent are not the same problem. A person responds to a brand story, a hero image, a feeling. An agent responds to whether your price, availability, rating, and specifications are present, parseable, and verifiable. The shopper you spent a decade learning to persuade is being replaced by a buyer that cannot be persuaded, only qualified.

Who is building the buying machine

This is not one company's bet. Every layer of the commerce stack, from the assistant to the payment rail, is racing to make agents into buyers. Five fronts matter.

OpenAI moved first on in-chat purchasing. It launched Instant Checkout in September 2025 and rolled "Buy it in ChatGPT" out to all US users on February 16, 2026, powered by the Agentic Commerce Protocol it co-developed with Stripe. Etsy sellers came first, with over a million Shopify merchants including Glossier, SKIMS, Spanx, and Vuori queued behind them. Then OpenAI did something instructive: in March 2026 it scaled back the standalone in-chat checkout and pivoted toward discovery plus merchant-owned checkout, listing Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, and Wayfair as ACP discovery integrations (Digital Commerce 360). The buy button moved. The discovery layer stayed.

Google is building the payments substrate. It announced the Agent Payments Protocol (AP2) on September 16, 2025 with 60-plus launch partners including Mastercard, PayPal, Coinbase, American Express, and Salesforce (Google Cloud). AP2 uses three cryptographically signed mandates, Intent, Cart, and Payment, carried as W3C Verifiable Credentials, so an agent can prove a human authorized a purchase. By May 2026, Google had donated AP2 to the FIDO Alliance for open governance.

Amazon is building a walled garden. It runs Rufus, its on-site shopping agent, and "Buy for Me," which purchases from other brands' sites on a shopper's behalf. Then it went to court to keep rival agents out, which we return to below.

Perplexity is pushing autonomous shopping through its Comet browser, which can scour the web, compare prices, and click "purchase" for the user.

Visa and Mastercard are building the trust layer for payments. Visa introduced its Trusted Agent Protocol in October 2025 with more than ten partners, an open framework built on the HTTP Message Signature standard that lets merchants distinguish a legitimate authorized agent from a malicious bot. By 2026 Visa was working with more than 100 partners, with over 30 building in its Intelligent Commerce sandbox. Mastercard's Agent Pay runs in parallel, and in January 2026 Mastercard publicly committed to supporting AP2 alongside its own rails.

The market has effectively conceded that agentic commerce will be multi-protocol. The fight is not over which standard wins. It is over which brands are legible to all of them.

Notice what every one of these players is solving for. Not persuasion. Plumbing. They are building the rails that let a machine identify a product, trust a merchant, authorize a payment, and complete a transaction without a human touching a checkout button. The brand's job is no longer to be chosen at the end of that pipeline. It is to be eligible to enter it at all.

The protocols that decide who can transact

The protocol layer is where eligibility is now defined, and it is worth understanding because it is the part most brands ignore.

A protocol is the contract that lets an agent and a merchant exchange a transaction safely. Four matter today, and they are converging rather than competing:

• ACP (Agentic Commerce Protocol) — OpenAI and Stripe's open standard for in-chat purchasing. It works across payment processors, integrates without backend changes, and keeps the merchant in control of the customer relationship. There are no fees on purchases that start in ChatGPT.
• AP2 (Agent Payments Protocol) — Google's signed-mandate framework for proving a human authorized an agent's purchase, now under FIDO Alliance governance with 60-plus backers.
• Trusted Agent Protocol — Visa's framework for merchants to verify that an inbound agent is authorized and not a malicious bot, built on existing web standards with minimal UX change.
• Agent Pay — Mastercard's network-specific implementation of agent-authorized payments, now committed to interoperate with AP2.

Underpinning all of them is MCP (Model Context Protocol), the open standard Anthropic introduced in November 2024 and now stewarded by the Linux Foundation with support from Anthropic, OpenAI, Google, and Microsoft. MCP lets a brand publish its catalog, cart, and commerce functions as a machine-callable interface. As one Shopify MCP guide put it, once your server is live, any agent that speaks the MCP language can interact with your brand instantly.

The strategic point is the same across every protocol. They reward brands whose product data is structured, complete, and exposed through a clean interface, and they ignore brands whose product truth is locked inside JavaScript-rendered pages and marketing copy. An agent transacting through ACP or AP2 does not read your homepage. It reads your feed, your schema, or your MCP endpoint.

The protocols are not a payments story. They are an eligibility story. They define the format your product must be in to be transactable at all, and that format is machine-readable, not human-readable.

For a brand, this collapses a decade of separate concerns, SEO, conversion-rate optimization, payments, into one question: is your product expressed in a structure an agent can read, trust, and act on? If the answer is no, the most advanced checkout infrastructure ever built routes around you.

How agents actually pick

Persuasion does not move an agent. Structure does. The research on agent selection is consistent and, for marketers raised on brand-building, uncomfortable.

The most rigorous examination, a Yale, Columbia, and University of Chicago study, found that AI agents show a 20–40% reduction in selection probability when a single key product attribute is missing. The agent does not fill the gap with goodwill or brand affinity. It downgrades or drops the product. The study also found that a small rating difference, 4.4 versus 4.1, frequently flipped which product the agent ranked first, and that agents assign "trust bonuses" to well-reviewed products, sometimes selecting them even when they cost more.

The 2026 commercial data confirms the pattern at scale. Stores with 99.9% attribute completion are seeing 3–4x higher visibility in AI recommendations than stores with sparse data (commercetools). The mechanism is mechanical: agents operate on confidence scores, rating every product on how well it matches the request. When attributes are incomplete or vague, the agent recommends a competitor instead.

The signals that move an agent cluster into four groups:

• Structured product data. Price, availability, dimensions, materials, compatibility, all explicitly marked up. This is the single most controllable lever and the one most retail pages fail. A page written for human scanning, with image-based feature callouts and client-side pricing, is frequently invisible to the agent deciding what to buy.
• Reviews and ratings. The verification layer. Agents lean on user-generated content because it separates brand claims from real experience. Small differences in average rating change outcomes.
• Trust and reliability signals. Return policies, fulfillment accuracy, retailer reliability. Agents favor merchants they can trust to deliver, and in 2026 those signals must be machine-readable, not buried in a footer.
• Contextual comparison. Agents do not score products in isolation. They score them relative to alternatives, at machine speed, which makes competitive context more decisive than it ever was for human shoppers.

Clean metadata beats creative marketing. The agent favors clarity over branding and precision over persuasion. You cannot tell it a story. You can only tell it the truth, in a format it can read.

There is one critical caveat brands must internalize. SparkToro's research found that AI recommendations are highly inconsistent in their ordering: less than a 1-in-100 chance of getting the same brand list twice, and less than 1-in-1,000 of the same list in the same order. The reconciliation with the selection studies is the most important strategic insight in this report. The consideration set, which brands appear at all, is stable and optimizable. The ranking within it is essentially random. Chasing a "position" in agent output is a fool's errand. Earning a permanent seat in the consideration set is the entire game.

Agent-eligibility is the new frontier, and it has teeth

If being chosen by a human was about reputation, being chosen by an agent is about eligibility. And eligibility is becoming something brands and platforms now fight over in court.

In early 2026, Amazon sued Perplexity, alleging its Comet browser agent concealed its automated nature and masqueraded as a human shopper to slip past bot detection and shop on Amazon on customers' behalf. A federal judge in San Francisco granted Amazon a preliminary injunction in March 2026, blocking Comet from accessing password-protected sections of Amazon to shop for users (GeekWire). The Ninth Circuit then temporarily stayed that injunction on appeal, letting Perplexity's agents back on, for now (Winbuzzer).

The case is a landmark, and it tells brands two things. First, the largest retailer on earth would rather build a walled garden, controlling the agent, the data, and the ad revenue through Rufus and Buy for Me, than let third-party agents transact freely on its turf. Second, "agent-eligible" is not a metaphor. Platforms are now drawing hard technical and legal lines about which agents may transact, which is exactly why protocols like Visa's Trusted Agent Protocol exist: to give merchants a sanctioned way to distinguish an authorized agent from an unauthorized bot.

For a brand outside the walled gardens, the implication is sharper still. Your eligibility to be transacted by an agent depends on three machine-facing assets:

• Structured data completeness. The 20–40% selection penalty for missing attributes is the price of an incomplete feed. Product, Offer, Review, and AggregateRating schema, kept dynamic so price and availability never go stale, is the floor.
• Machine-readable trust signals. Reviews, ratings, return policies, and reliability data the agent can parse and weight. Trust that lives only in brand copy is invisible to the buyer.
• A clean agent interface. Feeds, APIs, or an MCP server that lets agents access your catalog directly rather than scraping a rendered page. This is the path that skips the infrastructure entirely and delivers your product straight to the point of decision.

Being agent-eligible is not a marketing posture. It is an infrastructure standard. Either an agent can read, trust, and transact your product, or it cannot, and there is no charisma that closes the gap.

The brands that treat this as a backend chore are mispricing it. In an agent-mediated market, your data quality is your brand marketing. It is the only part of your brand the buyer can actually perceive.

The market size, and the honest version of it

The forecasts are large, and they vary enough that the honest reading matters more than the biggest number.

McKinsey's research projects agentic commerce could orchestrate as much as $1 trillion in US retail revenue by 2030, and $3 trillion to $5 trillion globally (Digital Commerce 360, McKinsey). That is the orchestration figure, the total revenue agents touch or influence. The narrower "agents actually doing the buying" estimates are more modest and more contested:

• Morgan Stanley estimates agentic shoppers could represent $190 billion to $385 billion in US e-commerce spend by 2030, capturing 10% to 20% of the market.
• Bain & Company projects 15–25%, or $300–$500 billion.
• Gartner projects 20% of digital commerce transactions will be executed through AI platforms or agents by 2030, and that by 2028, 90% of B2B buying will be mediated by AI agents.

The consensus that emerges across the credible forecasts: by 2030, roughly 10–25% of US e-commerce will involve AI agents in a meaningful way, on the order of $200 billion to $500 billion in direct agent transactions, sitting inside a far larger pool of agent-influenced revenue.

Two near-term anchors keep this grounded. Gartner predicted traditional search engine volume would drop 25% by 2026 as answer engines substitute for queries, and that 60% of brands will use agentic AI for one-to-one interactions by 2028. And Salesforce's 39% of consumers, over half of Gen Z, already using AI for product discovery shows the demand side is not waiting for the forecasts.

The biggest number is not the point. The point is the slope. Agents went from zero share of holiday orders to influencing 20% in roughly two years. A channel does not need to be a majority to be the one deciding your next margin.

The discipline here matters. Gartner also predicts over 40% of agentic AI projects will be canceled by end of 2027, and OpenAI's retreat from standalone in-chat checkout shows the buy buttons will keep shifting. The durable bet is not a specific checkout integration or platform. It is being the product an agent can find, trust, and transact, on whatever surface and through whatever protocol the buyer happens to be using.

The execution implication

The work splits cleanly into what you own, what you influence, and what you measure. In an agent-mediated market none of it is optional.

Make your product machine-eligible first. This is the foundation and the fastest win, because the selection penalty for missing data is steep and the fix is concrete.

• Implement complete Product, Offer, Review, and AggregateRating schema, and keep it dynamic. Stale or contradictory data gets the product dropped, not forgiven.
• Render price, specs, availability, and reviews server-side. If they only appear after JavaScript executes, assume the agent never sees them.
• Drive attribute completeness toward 100%. The 3–4x visibility gap between complete and sparse catalogs is the highest-leverage number in this report.
• Stand up a feed or MCP endpoint so agents can access your catalog directly rather than scraping a rendered page.

Earn the trust signals an agent weights. Reviews, ratings, and reliability data are the verification layer, and the agent treats third-party validation as more credible than brand claims. Build genuine review depth, surface return and fulfillment data in machine-readable form, and keep ratings current.

Optimize for the consideration set, not the ranking. Because ordering in agent output is near-random, chasing a position wastes effort. The winnable goal is permanent presence in the set of products an agent considers at all. That is a function of data completeness, trust signals, and contextual competitiveness, not of a single keyword or placement.

Plan for the protocols and the walled gardens. Get readable through the open protocols, ACP, AP2, MCP, so you are transactable wherever agents operate. And recognize that platforms like Amazon are building closed systems where they control the agent; decide deliberately where you accept those terms and where you protect a direct relationship.

Measure selection, not traffic. Classic analytics tell you who arrived. They cannot tell you whether the agent considered you in the first place, and the decision now happens before any click. Track share of voice in agent answers, the prompts you appear and do not appear in, and which competitors are selected when you are not.

The old job was persuading a person who could be moved by a story. The new job is qualifying for a machine that can only be moved by structure. Visibility was a placement. Eligibility is an architecture.

The brands that win the agentic transition will not be the ones with the best campaigns. They will be the ones whose product truth is the most complete, the most verifiable, and the most machine-readable, because that is the only version of the brand the buyer can now perceive. The buyer became a machine. The brands that get read are the ones that get bought.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.