Agents are becoming customers. GEO only covers the first step

In September Michał Sadowski, founder of Brand24, posted on LinkedIn that a new kind of marketing is coming. Its job is to make a product easy to find and easy to test for an AI agent working on the customer’s behalf. GEO, tracking how a brand appears in model answers, is one piece of it. He listed three more: getting the product into the protocols agents use, managing how the brand looks in MCP marketplaces and integration platforms, and publishing an offer an agent can evaluate.

I agree with all of it. But two months on, the picture has moved. That changes where I would start.

What is live in 2026 and what is still a waitlist

Payments got the loudest announcements. They got the least adoption.

OpenAI launched Instant Checkout in ChatGPT with Stripe in September 2025, on the Agentic Commerce Protocol. In March 2026 it stepped back: merchants can use their own checkout, and ACP now carries product feeds for discovery inside ChatGPT. Google went the other way in January with the Universal Commerce Protocol, which supports checkout from eligible US retailers in AI Mode and the Gemini app.

Cloudflare announced Cloudflare Wallets on August 4, 2026. This is Michał’s example. An owner gives an agent a wallet with a spending cap, an approved merchant list and a maximum transaction size. Today you can reserve a handle. Funding and agent wallets are still listed as coming. The seller side moved further. Cloudflare’s Monetization Gateway entered a closed beta for eligible US sellers on September 30, charging agents per request over HTTP 402, with payments settled in USDC on Base. It already runs in production for paid inference, search, market signals and PDF generation.

Discovery is where the volume is. OpenAI moved its app directory into a plugin directory in July 2026, shared by ChatGPT and Codex. The official MCP Registry verifies that a server published under your domain is really yours. Integration platforms such as Composio and Pipedream put thousands of tools behind one endpoint, and an agent picks among them by reading descriptions.

Where I would start, in order

1. Publish an offer a machine can evaluate

An agent comparing five tools reads your pricing page first. “Enterprise: contact us” is an invitation to a person. To an agent it is missing data, and missing data usually ends the comparison.

You do not have to give up sales-led pricing. Publish the structure: plans, what each includes, limits, unit prices, and where custom pricing starts. Put it in structured data on the pricing page, in documentation an agent can read without rendering JavaScript, and in an llms.txt that points to both.

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Example SaaS, Business plan",
  "offers": {
    "@type": "Offer",
    "price": "49.00",
    "priceCurrency": "EUR",
    "priceSpecification": {
      "@type": "UnitPriceSpecification",
      "price": "49.00",
      "priceCurrency": "EUR",
      "unitText": "per month, up to 10 users"
    },
    "url": "https://example.com/pricing#business"
  }
}

Add the terms an agent needs to limit risk for the person it works for: trial length, cancellation, refunds, data location. An agent is told to avoid surprises, so the vendor with predictable terms gets picked.

2. Treat your tool listing as a product page

A plugin listing, a registry entry or a connector on an integration platform is read by a model before any person sees it. The tool name, its description and the description of each parameter decide whether the agent understands what your integration does and chooses it over a competitor’s.

This is copywriting for a different reader. A tool described as “Manage records” loses to one described as “Create an invoice for an existing customer and return its PDF link”. Models choose tools from their descriptions, so the gap between a tool an agent calls readily and one it skips can come down to two sentences and clearly named parameters. Test it: give an agent your tools and a competitor’s, and log which one it calls.

3. Get into the payment rails only where you sell per request

If your product is an API, a dataset or a paid tool, per-request payment over x402 is worth a pilot now, because it removes account creation and API keys from the agent’s path. If you sell annual seats, payment protocols matter less this year than a trial an agent can start without emailing sales.

Measure it separately or you will not see it

GA4 sees people with browsers. An agent calling your API or MCP server never shows up there. Before you optimise any of this, log agent requests at the server, count MCP tool calls by tool, and track 402 responses and completed payments if you charge per request. Otherwise you are tuning a channel you cannot see, which is where most teams already are with GEO.

What to track	Where it lives
Mentions and position in model answers	GEO monitoring
Plugin and registry installs, tool calls per tool	Your MCP server logs
Agent requests to pricing and docs	Server or CDN logs, by user agent
Trials started by agents, paid requests	Product analytics and billing

Why I am writing about this

I have run SEO, conversion and pipeline work in B2B SaaS, most recently as Head of Growth & Web Development at Symfonia and as a fractional GTM lead at CodiLime, where much of the work was building assets for technical buyers. When buyers change how they reach a product, I see it in revenue first. My own marketing stack runs on agents connected to my tools through MCP. This site runs on Cloudflare. I also publish model comparisons here, built on repeatable tasks. So I see the problem from both sides: as someone who writes for buyers, and as someone watching which tool an agent actually calls.

Gartner expects AI agents to intermediate 90% of B2B purchases by 2028. I would not plan a budget on that number. I would plan on this: the first evaluation of your product is increasingly done by software, and it reads different signals than a person does.

If you want to see how your product looks to an agent before your competitors do, book a call. We start with a short audit of the five points above.