The New First Buyer

One of my earliest jobs was at a YC-backed e-commerce startup. I later spent a few years at Amazon. That made me pay attention to how online shopping actually works. Small changes in search, ranking, inventory, and fulfillment can reshape how millions of people experience a purchase.

Agentic commerce feels different from a normal change in checkout or distribution. It changes who, or what, the merchant has to persuade.

For most of the consumer internet, the merchant built for the human at the screen. Now the merchant may have to satisfy the shopper's agent first. The agent wants structured data, reliable inventory, deterministic pricing, provable identity, and low transaction risk.

SEO is an imperfect analogy, because the decision-maker changes. Search engines helped humans find options. Agents increasingly decide which options are worth showing. A search result can tolerate uncertainty, because the human interprets, compares, and takes the risk. An agent has a different incentive. If it recommends stale inventory, ambiguous pricing, or a broken checkout path, the user blames the agent for a bad recommendation.

The website does not disappear, but its role changes. It carries the brand, handles direct traffic, and shows the human what they came for. Whether the human's agent arrives at all depends on other parts: the feed, the manifest, the checkout endpoint, the identity layer, and the payment permissions.

Online merchants have spent years on persuasion: landing pages, photography, recommendations, reviews, and checkout UX. Those matter when a human is in the loop. An agent first needs to know that it can evaluate the transaction and complete it safely.

What agents need

People built most commerce websites for other people to read. Agents need data explicit enough to act on. That means product attributes, inventory, delivery windows, return policies, discounts, substitutions, bundles, and payment constraints. In human search, imperfect information can still lead to a click. In agentic commerce, it may keep the option out of consideration.

No single interface exists that every merchant can implement once and forget. Each agent platform asks the merchant to expose products, checkout, identity, payments, and policy in a slightly different shape. Merchants have to decide which platforms matter before demand is obvious.

Trust is harder, because agentic commerce lets software browse a catalog and, more importantly, commit money, identity, and intent. Payment protocols try to solve this with scoped credentials and authorization records. Stripe's shared payment tokens, for example, let an agent pass limited payment credentials to a seller. The underlying card never moves. Google's Agent Payments Protocol builds the payment layer around mandates. Anyone can check a mandate to see what the user authorized and what the agent bought.

Those details sound technical, but the merchant is still on the hook when the agent gets it wrong. The failure modes include fraud, chargebacks, bad discounts, and incorrectly indexed products. They also include confused substitutions, policy violations, and disputes over who authorized a transaction. A merchant can tolerate an inefficient new channel. It is harder to tolerate one that applies the wrong discount to a catalog or approves the wrong substitution at machine speed.

The merchant does not want a protocol war

From the merchant's perspective, the problem is not that protocols exist. It is that each protocol points merchants toward a different platform bet. One centers on ChatGPT. Another centers on Google and Shopify. Another centers on payment networks. Another centers on merchant-hosted agents that negotiate directly with buyer agents.

The Agentic Commerce Protocol is one path, built around programmatic checkout between buyers, agents, and sellers. Universal Commerce Protocol, which Shopify and Google developed together, is another. Stripe's agentic commerce docs now point sellers toward either UCP or ACP, depending on the integration.

The cost resembles earlier standards wars. The platform with the most distribution can ask the ecosystem to conform. The smaller merchant has to decide whether the integration is worth the work. If there is one winner, the problem eventually simplifies. If the market stays split, merchants either maintain multiple integrations or buy an abstraction layer that makes the split less painful.

The translation problem

The useful product here may be narrower than "make my website AI-ready." It looks more like a translation layer. Take the merchant's existing product, inventory, fulfillment, pricing, and payment data. Then expose it in whatever shape the relevant agent requires. Implementation speed is part of the value. It also cuts the number of platform bets a merchant has to make before demand shows up.

Whichever product merchants pay for first will say what they are actually worried about. If they buy protocol mapping, the pain is engineering cost. If they buy an agent-readiness audit, it is visibility, and if they buy delegated payment controls, it is liability. I would watch that last one most closely, because the downside there is a transaction the merchant never meant to accept.

The fixed price starts to look less fixed

Agents may negotiate with each other. UCP describes commerce capabilities as something merchants and agents can declare and negotiate. That matters because commerce is full of conditional behavior: discounts, fulfillment constraints, loyalty status, substitutions, bundles, pickup windows, payment methods, and return policies. Once machines can read those conditions, price and terms stop being whatever the page prints. They become something software can reason about.

At first, this looks mundane. An agent asks whether a merchant can apply a loyalty discount, find a substitute, bundle items, or meet a delivery window. But the direction is more interesting than the first use case. A buyer agent can represent willingness to pay, timing, preferences, brand constraints, and substitution tolerance. A seller agent can represent margin, inventory pressure, promotions, and fulfillment capacity. Between the two, the listed price starts to look like an opening number.

That creates problems for brands. What does fairness mean when some consumers have better agents than others? What happens to comparison shopping when the comparable object is no longer product plus price, but product plus negotiated terms?

Online retail has had dynamic pricing for a long time, but agentic commerce makes the negotiation explicit. Agents will keep checking and haggling in ways people usually will not. A person may not want to open twelve tabs, read return policies, compare loyalty programs, and haggle over bundles. An agent might. That could open categories where the friction used to be too high: services, B2B purchasing, configured goods, and high-consideration purchases.

Where consideration starts

Not every purchase will move into an agent interface. People still browse, brands still need websites and product pages, and many purchases are emotional or aesthetic in ways that resist full delegation.

The shift matters most when agents become the first filter for high-intent purchases. In those cases a merchant has to be eligible before it can be visible. The data has to be readable, the transaction has to work, and the risk has to be acceptable. Most merchants have never had to prove any of that to a piece of software before.