Morning. It is quiet in the apartment. The refrigerator knows by itself that the milk has run out. It whispers to the kitchen assistant: “Into the cart also cheese and apples, please.” The buyer is not a person. It is a home agent. It negotiates with a merchant agent near your house. That one instantly checks prices, freshness, expiration date. Finds the best combination. Payment goes out without pressing buttons. A drone lands in the yard. In the box—milk, cheese, apples. The robot-courier nods to the peephole, like a good neighbor, and disappears into the sky.
At this same time, in another part of the city, a corporate agent purchases coffee machines for the office. It bargains with a dozen sellers. It checks the warranty, service intervals, availability of consumables. It looks at the return history. It pays according to permissions, signed beforehand. The coffee machines arrive towards evening. They are met not by managers, but by other robots—the ones responsible for inventory and setup.
People? They set goals, formulate constraints and check the result. Everything else—from search to payment and delivery—agents take upon themselves. And this is not "sometime" science fiction. This is already being designed and assembled brick by brick.
From imagination—to action
In the fall, the industry received an important missing element of this picture—a payments protocol for the agent world. Google introduced the Agent Payments Protocol (AP2). This is an open way for "buyer-agents" and "seller-agents" to agree on a purchase, confirm intent, and securely debit money—and on top of different payment methods, from cards to instant transfers and stablecoins. The protocol is intended as an extension of two already existing "universal languages" for agent systems: A2A (Agent-to-Agent, a standard for agent interaction with each other) and Model Context Protocol (MCP, a standard for connecting agents to data and tools). The idea is simple: if agents can talk and connect to sources, they also need a unified way to pay—transparently, with trusted authority and reporting. This is exactly what AP2 promises to address. a2a-protocol.org
The Google team explains why a new level of rules is needed. Today's payment systems assume that somewhere there is a person who personally presses "buy." With autonomous agents, this assumption breaks. Three key questions arise: who authorized the purchase, does the payment correspond to the user's real intention, and who is responsible if something went wrong. AP2 formalizes the answers to these questions. At the center are "mandates" (Mandates): cryptographically signed instructions that record the intention and the specific cart. This creates a chain of proof: from intention to cart, from cart to payment. Transparent, verifiable and—most importantly—with a unified vocabulary for different platforms.
In addition to the general framework, details are important. AP2 supports different payment methods, including new rails from the crypto world. Together with Coinbase, Ethereum Foundation and other participants, an extension module x402 for agent crypto payments has been introduced. This is a signal: the protocol aims to be neutral to the payment "physics," and therefore—scalable.
A2A, on which agent interaction is built, has already been brought into the orbit of open standards: its documentation emphasizes the role of a "common language" for the joint work of agents from different vendors and frameworks. MCP, in turn, serves as the "USB-C for AI"—a unified way to connect agents to data and external tools. In the combination A2A + MCP + AP2, we get the basic infrastructure: we talk, we connect, we pay. a2a-protocol.org
What this changes in e-commerce
When an algorithm buys, the logic of choice changes. The agent does not look at a banner. It does not give in to a random impulse. It optimizes for given goals: budget, timelines, quality, risks. It compares catalogs, versions, properties. It keeps in memory the history of successful and unsuccessful orders. It evaluates after-sales service. It does all this lightning-fast, and without fatigue.
The more advanced such agents are, the smaller the role of the "hit-or-miss" showcase and the greater the role of data. Choice turns into calculation. Competition—into comparison by parameters and probabilities. This means companies will have to feed the market not just with advertising, but with precise, coherent, and constantly updated knowledge about themselves and their products.
This transformation does not cancel marketing. It shifts the center of gravity. From storytelling—to the structure and completeness of information. From emotion—to the link between what is promised and what is confirmed by facts, certificates, reviews, warranties, delivery times. The agent reads all this. And understands.
How trust works between agents
AP2 introduces the concept of a "mandate"—like a digital authorization. It proves that the owner of the wallet (or card) wanted exactly this purchase on these terms. There are "person nearby" scenarios—when we confirm the cart that the agent collected. There are "person not nearby" scenarios—when we set rules in advance: buy tickets if the price drops below X, or order a tire change if a sensor reports a puncture. In both cases, the seller and the payment provider have the means to resolve questions of authorization, authenticity, and liability. This is the key to scale: without a trusted, verifiable "history of intention," autonomous purchases will run into legal and regulatory barriers.
A wide circle of participants is already gathering around the protocol: payment networks, providers, marketplaces, technology companies. This is important not for the sake of logos, but for the sake of interoperability. The more nodes understand the same "language of intentions and carts," the less friction there is between agents and platforms. Google Cloud
To paraphrase Google's position: a common vocabulary is needed so that agent purchases are secure, correspond to the user's will, and fit into existing and new payment rails. Partners from the ecosystem echo: standards are a condition for trust, and trust is a condition for mass adoption. This is what consensus looks like.
What data buyer-agents will use
A buyer-agent assembles a "portrait" of an offer from many layers of data. On the first level—the classics: price, availability, parameters, dimensions, compatibility, delivery time, delivery cost, return policy. On the second—reliability of the seller and the product: warranty, service centers, failure indicators, reviews, certification. On the third—context compliance: whether the product is suitable for a specific task and place, whether it fits within corporate policy, whether there are any regulatory restrictions.
Another layer is time. The agent can wait for a discount, monitor currency exchange rates and logistics windows. It compares a direct purchase from a vendor with a purchase through a marketplace. It bargains with a "seller-agent" if that agent is ready to create bundle offers. AP2 specifically describes how such deals are finalized, when the price "reaches" the right point and the cart is agreed upon by both sides.
But there is a special type of data on which everything else will depend. This is the system's knowledge about the brand's identity and product lineup. How does the AI understand that "this model" is the successor to last year's? How does it distinguish "Pro" from "Plus," regional SKUs, local bundles? How does it connect reviews, tests, instructions, service bulletins to the correct object? If in the model's memory or in its connected sources there are gaps and confusion here—the agent will make mistakes, and thus, will not choose you.
Brand as data
In a world where an algorithm makes the choice, a brand is not a logo and not a slogan. It is a knowledge graph, filled with facts. Names, codes, attributes. Compatibility, versions, life cycles. Update history. Correspondence of models in different markets and languages. Map of accessories. Instructions and service standards. Certificates and permits. All of this must be correct, uniform, and constantly kept up-to-date.
A "data as a product" discipline is needed. Detailed specification of the assortment. Unified identifiers. Updates "to a single source of truth." Publication in formats that agents understand—from structured markup on a website to APIs for partners and marketplaces. Otherwise, the model will fill in the gaps on its own. And filling in on its own—is a chance for an error.
And here GEO comes into play—Generative Engine Optimization. This is not SEO in the classic sense. It is work to ensure that generative systems—assistants, agents, models—answer about you correctly, confidently, and consistently. So that in their answers you are represented where needed, and disappear where it is harmful. So that they know current lineups, prices, tolerances. So that they name you correctly in all languages, in every region. GEO is not a trick, but infrastructural work with brand data.
GEO: what it is and why
Generative Engine Optimization is about influencing the answers of generative engines and agents through the quality and coherence of data, and not through "gaming the algorithm." The goals are measurable:
- Share of Voice in AI answers: how often your brand pops up when a user (or agent) needs your class of products.
- Completeness of coverage: does the system know all your models, series, configurations. Does it confuse generations, does it "lose" regional SKUs.
- Correctness of attributes: do the properties in the answers match your specifications. Are there no dangerous "hallucinations."
- Currency: how quickly the AI "digests" updates—new models, discontinuations, changes in prices and conditions.
- Tone and associations: what emotional and business connotation the brand carries in the answers (important for consumer and B2B segments).
The practice of GEO is helped by standardization. If your site is marked up with structured data, if catalogs are uniform, if the brand "map" matches in your own and partner sources, agents build understanding faster and more accurately. MCP adds a channel to your data and tools—this is another way to convey the "truth from the original source" to the model. A2A helps agents exchange results and delegate tasks. AP2 finalizes that the result of this exchange can be a purchase—secure and verifiable. Together, this is the new commerce. modelcontextprotocol.ioa2a-protocol.org
How to sell to agents
Imagine a product card that is read not by a person. It does not need a banner and a "brand story" in a poetic style. It needs:
- precise name and identifier;
- an exhaustive list of attributes and their units of measurement;
- links to accessories and compatible models;
- maintenance standards, consumables, frequency of maintenance;
- legal conditions: warranty, return, compliance with regulations;
- logistics and SLA: timelines, delivery windows, restrictions;
- quality telemetry: fault tolerance, frequency of returns, rating of service centers;
- "social proof," but structured: confirmed reviews, results of independent tests, links to certificates.
And all this—synchronized and consistent in a single graph. Then a seller-agent can confidently assemble a bundle offer. And a buyer-agent—can check that it fits within the intention and budget. AP2 describes how to finalize this "meeting in data" in the form of mandates and a cart, in order to then pay without friction.
Logistics: robots and drones are not an exotic thing, but a routine layer
When payment and cart approval become machine-driven, logistics also moves toward automation. "Last mile" robots, drones for urgent deliveries, automated warehouses. For a buyer-agent, this is simply an expansion of the menu: it chooses not "beautiful packaging," but a trusted delivery scheme with a known probability of success and arrival interval. In terms of data, this is another matrix: schedules, weather windows, weight and noise restrictions, narrow streets, city policy. With the development of agent protocols, logistics will become a transparent set of parameters, not a "black box." Choice is again calculation.
The role of the human
The human remains the main one. They formulate rules and boundaries. They approve exceptions. They resolve conflicts in data and contradictory goals. But where intuition and experience used to work, now statistics and verifiable transaction histories will work. This will increase the requirements for brand management. From marketing "in the broad sense"—to managing knowledge about the brand and products. From creative—to data infrastructure.
What brands should do right now
- Bring the assortment to a "unified specification." Same naming rules, units of measurement, canonical identifiers.
- Connect products in a graph: successors, versions, compatibility, accessories.
- Ensure authenticity and accessibility. So that MCP clients and partner agents can receive the "truth" from the original source. modelcontextprotocol.io
- Deploy monitoring of "how AI sees us." What do large models answer? What is the Share of Voice? Where are the errors and gaps?
- Set up an update process: releases, discontinuations, prices, conditions—everything must "flow out" to the world quickly and without distortion.
- And—importantly—stop thinking of GEO as tricks. It is an operational discipline. It is about the cleanliness and synchronicity of data. About agents being able to answer about you better than any salesperson on the floor.
Aureol comes to help
The agent world requires measurements. A dashboard is needed that shows what generative systems know about you. What they confuse. Where you are in the answers, and where you are not. How the share of mentions changes. What the tone of the context is. Which products have "fallen" out of view.
This is exactly where analytical tools, like Aureol, are useful. They help to see the Share of Voice in AI answers, measure the completeness of assortment coverage, check the correctness of attributes and currency. They help compare markets and languages. To highlight where the brand "map" is broken, and where, on the contrary, everything is ideal. This is not a replacement for data and processes—it is a radar. It warns about slumps and records progress.
Instead of an epilogue
We are standing on the threshold of a world where it's not clicks that buy, but agents. Where a brand's "speech" is first and foremost its data: connected, verifiable, alive. Where logistics is a set of parameters, not a phone call. Where payment is a chain of verified mandates, not a form on a website.
This picture is already taking shape. "Buses" for agent communication have appeared (A2A). There is a unified "port" for connecting to data (MCP). A payments protocol (AP2) has appeared, which addresses the main risks and questions of trust and is compatible with different payment rails, all the way to stablecoins. Next—scaling, standards, and daily routine. And those who start with the main thing will win: putting knowledge about themselves in order, and making it accessible to machines. Because now the algorithm makes the choice. And it does not choose a banner. It chooses facts.