When the buyer is not a person but their agent

Throughout the AI100 corpus, we have been talking about a world in which the user asks a question and the system forms an answer. That is the world of the answer environment: the brand competes for the right to be named, cited, and recommended. But the next wave is structured differently. An AI agent does not merely answer a question — it acts: it searches for products, compares options, checks prices and availability, and in some cases already completes the purchase without showing the person an intermediate list of alternatives.

This is not a distant future. ChatGPT Instant Checkout has been operating since September 2025, allowing users to make purchases directly inside the conversation through Shopify and Etsy partners [1]. In January 2026, Google announced its own Universal Commerce Protocol (UCP), joined by Walmart, Target, and more than 20 partners [2]. Amazon is building a closed ecosystem — Rufus AI and Alexa+ — inside its own platform. Three competing systems are already live, and none of them asks brands for permission to participate.

Salesforce data from Cyber Week 2025 shows that the volume of tasks completed by AI agents on shoppers’ behalf grew 70% versus 2024, and 84% on Black Friday [3]. Gartner forecasts that by the end of 2026, 25% of enterprise software procurement will involve an AI agent [4]. McKinsey estimates that personalization and autonomous AI-assisted purchasing could create an additional $1.2 trillion in value for global retail [5]. According to Human Security, agentic AI activity grew 6,900% over the past year, and 64% of consumers planned to use AI for holiday shopping in 2025 [6].

For a brand, these figures point to a simple but painful reality: one of the most important “buyers” in the near future may not be human at all. And if it is not human, then website design, emotional positioning, visual identity, and banner ads are irrelevant to it. What it needs is structured data it can read, compare, and verify.

Today, agentic commerce is taking shape around three incompatible protocols. ACP (Agentic Commerce Protocol) is the OpenAI–Stripe protocol that works through ChatGPT Instant Checkout. OpenAI takes 4% from each transaction plus the standard payment processor fee [1]. UCP (Universal Commerce Protocol) is Google’s coalition protocol, connected to AI Mode and Gemini, with support from Walmart, Target, and Shopify [2]. Amazon’s closed ecosystem — Rufus AI, Alexa+, and Buy for Me — works only inside the Amazon marketplace.

For a brand, that means it must be visible in all three worlds. A company that has optimized its data for only one platform loses demand on the other two. And none of these worlds operates by the rules of classic SEO.

A shopper asks ChatGPT: "Find me a smoothie blender under $150 that's quiet and has at least a two-year warranty." The agent queries catalogs via Shopify. Brand A is a well-known manufacturer with a polished website, but its product feed lacks a "noise level" attribute, and the warranty is mentioned only on an FAQ page in PDF format. Brand B is less established, but its Product schema includes all twelve attributes, including decibel level and warranty_months. The agent picks B. Not because B is better, but because B is machine-readable.

In the enterprise world the mechanics are the same, only the stakes are higher. A company is looking for procurement software; one vendor has SSO, ISO 27001, deployment model, and SLA listed in a structured catalog with machine-readable attributes. Another has all of that buried in a PDF sales deck. An agent comparing options on a CTO's behalf will pick the first — not because the product is better, but because the data is available for computable comparison.

An AI agent does not read marketing copy the way a human does. According to MIT Sloan, 82% of executives cite data quality as the main barrier to achieving generative AI goals [7]. The agent works with structured attributes: name, category, price, availability, specifications, return terms. If that data is missing, inaccurate, or out of sync across platforms, the agent cannot compare the product correctly — and it chooses the competitor with cleaner data instead.

That is fundamentally different from the answer environment. In a model-generated answer, a brand can compensate for weak data with a strong reputation or vivid expert content. In an agentic scenario, those compensations disappear. The agent makes its decision on the basis of attributes, not impressions. If your product feed does not contain machine-readable data on compatibility, warranty, size, or delivery times, the agent may remove you from the comparison before it even gets to assessing quality.

There is another unpleasant feature of the agentic world: classic web analytics stop working. When a user asks ChatGPT about a product category, the agent compares, recommends, and completes the purchase — all inside the conversation. The brand’s site gets no visit, no click, no session [1]. GA4 does not see that traffic. A marketer who measures success by site visits does not notice that the sale happened through another channel.

That does not mean analytics has become useless. It does mean, however, that new metrics must appear alongside the familiar ones: citation in agentic recommendations, inclusion in agentic comparisons, conversion through agentic checkout. The tools for this are only beginning to emerge, but the need itself is already obvious.

Agentic choice has not yet become the dominant form of purchasing. But the pace of growth is such that within 12–18 months it could become a mandatory channel for any company selling goods or services online. And preparing for it does not begin with a new tool. It begins with fundamental work: clean structured product data, synchronized across platforms, machine-readable, and current.

For a business owner, this means yet another shift in mindset. First, you had to learn to be visible to the search engine. Then, to be understandable to the answer system. Now you have to be usable for the agent — which does not ask, does not read, and does not click, but simply chooses on the basis of data.