How AI agents find and buy services

As of September 2026, AI assistants and agents help people find, compare, and book services, but no major assistant sells a professional service inside the chat. On March 24, 2026, OpenAI stopped offering in-chat checkout to new merchants (opens in a new tab); Google's checkout in AI Mode and Gemini excludes services (opens in a new tab); and the other in-chat programs are built for retail goods. When an agent pays for a service, it pays on the seller's own checkout or through a payment API, and a person approves the payment.

This guide is part of AEO HQ's work on AI search optimization, the practice of being found and recommended in AI answers. It describes what each assistant and payment provider documents today, what research says about how agents choose, and what a service business can do. Facts link to their sources, most of which are the platforms' own documentation; several features described here are in preview and change often. Recommendations are marked as ours.

Scope and definitions

This guide is for businesses that sell services, from accounting and consulting to bookable appointments, and want AI agents to find them and, where possible, complete a purchase. It covers ChatGPT, Google's AI Mode and Gemini, Microsoft Copilot, Perplexity, Claude, and Stripe's payment tools for agents, as documented on 27 September 2026. It does not cover advertising inside assistants or building an agent of your own.

• AI agent. Software that takes actions for a person, such as searching, opening websites, filling in forms, or paying, rather than only answering a question.
• Agentic commerce. Buying and selling in which an AI agent carries out some of the shopping or checkout steps for a person.
• In-chat checkout. Paying inside the assistant, without visiting the seller's website. OpenAI called its version Instant Checkout.
• Merchant of record. The business that sells to the customer and is responsible for the transaction. In the assistants' programs, the seller keeps this role: OpenAI's launch post promised to keep merchants "in control of the customer relationship as the merchant of record" (opens in a new tab), and Microsoft says merchants "remain the merchant of record" (opens in a new tab).
• Model Context Protocol (MCP). "An open-source standard for connecting AI applications to external systems" (opens in a new tab). A business can run an MCP server that agents use, for example, to search availability and manage bookings (opens in a new tab).
• Commerce protocols. Shared formats for agents and sellers to exchange catalog, cart, and checkout data. The Agentic Commerce Protocol (ACP) is an open specification that OpenAI and Stripe maintain, currently in beta (opens in a new tab). Google describes the Universal Commerce Protocol (UCP) as an open protocol that "standardizes the programmatic exchange of information (via API, MCP or A2A) between the AI Agent and the Merchant's backend" (opens in a new tab).
• Browser agent. An agent that operates a web browser, such as ChatGPT's cloud browser or Gemini in Chrome's auto browse.
• Shared Payment Token (SPT). Stripe's scoped grant that lets an agent use a customer's payment method, with usage and expiration limits (opens in a new tab).

Where AI agents can buy today

Each platform documents a different route. The table summarizes what exists and whether a service business can use it.

A note on history. OpenAI launched Instant Checkout on September 29, 2025, for single-item purchases from U.S. Etsy sellers, and merchants paid "a small fee on completed purchases" (opens in a new tab). OpenAI's help center still says that for "some eligible products and merchants, ChatGPT may also show an Instant Checkout option" (opens in a new tab), and its plugin guidelines say Instant Checkout is in beta and "available only to select marketplace partners" (opens in a new tab). None of these pages offers it to new applicants.

How agents find services

An agent finds a service business the same way an assistant does, and a few newer routes add to that.

1. Search. Assistants that search the web draw on search indexes: OpenAI says ChatGPT search "sometimes partners with other search providers," and it names Microsoft and Shopify (opens in a new tab), and Google's generative AI features are "rooted in our core Search ranking and quality systems" (opens in a new tab). Google says "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (opens in a new tab), and that llms.txt files "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them" (opens in a new tab) (official documentation). For a service business, search is the main route (our reading), and How ChatGPT, Gemini, Claude, Perplexity, and Copilot find and cite sources describes each index.
2. Product feeds. Feeds are how retailers reach shopping features, and they are built for products. OpenAI's merchant page answers "Do I need a product feed if ChatGPT already crawls my site?" with "No, but feeds give you greater control" (opens in a new tab), and OpenAI says product results use structured metadata from first-party and third-party providers (opens in a new tab). The Google and Microsoft checkouts above both run on merchant feeds.
3. Apps and connectors inside the assistant. A user can add an app or connector that searches a directory. In April 2026, Clutch, a B2B services marketplace, launched an app in ChatGPT that surfaces "verified profiles, review highlights, pricing signals, and real client feedback" (opens in a new tab) (company announcement). MCP is supported by "AI assistants like Claude and ChatGPT" (opens in a new tab), among many other clients.
4. Directories for agents. Stripe runs a directory for agents, in public preview (opens in a new tab), that helps agents find businesses' tools and services; a business applying for a listing is asked for "a link to your llms.txt file" and "One to three example prompts" (opens in a new tab). This is the only documented use of llms.txt for agent purchases found for this guide. Most such files go unread: 97% of llms.txt files received zero requests in May 2026 (opens in a new tab) (137,210 domains; vendor server-log study).

Whichever route an agent takes, it reads the facts a business publishes. Perplexity says its answer engine is "more likely to recommend merchants who provide deeper product details such as availability, reviews, pricing, and specifications" (opens in a new tab) (official documentation, about products).

How agents choose: what the research shows

Research on AI agents as buyers is new, and most of it uses simulated stores selling consumer goods. The findings below are consistent in direction; their sizes vary by model and change when models change.

Our reading: accurate, specific offer facts and genuine ratings are the levers the research supports. Placement, speed of response, and wording also move simulated agents, but by amounts that differ between models, and one study found that agents buying expert services were drawn to low prices.

How the purchase happens

In every flow documented for this guide, a person approves the payment. ChatGPT's cloud browser asks for confirmation before a payment (above), and OpenAI adds that some tasks require "a final step you complete yourself" (opens in a new tab). In Stripe's Link agent wallet, the customer approves each spend request, and Link returns a one-time-use payment credential (opens in a new tab). Business buyers are cautious as well: in G2's June 2026 survey of software buyers, 61% used or planned to use AI agents in software buying, but 9% were comfortable letting agents execute purchases within approved guardrails, and 2% without pre-approval (opens in a new tab) (review-platform survey).

Three documented routes lead to a paid order:

1. A browser agent completes the seller's own checkout. OpenAI notes that some websites restrict automated browser agents and that "each website ultimately decides whether to allow cloud browser traffic" (opens in a new tab). To let operators tell its agent from other bots, ChatGPT's cloud browser signs its requests with Web Bot Auth, using the HTTP Message Signatures standard (RFC 9421) and a Signature-Agent header set to "https://chatgpt.com" (opens in a new tab).
2. An agent books and pays through the seller's API or MCP server. Stripe's guide for service businesses shows this pattern and says "Stripe doesn't require a specific endpoint structure, tool naming convention, or payload schema" (opens in a new tab).
3. An agent pays per request over HTTP. In Stripe's machine payments, the server "returns an HTTP 402 response with payment details" (opens in a new tab), and the agent pays with a Shared Payment Token; for card payments through SPTs, the minimum amount is 0.50 USD (opens in a new tab). Stripe's guide to charging agents through an MCP server, which it calls "experimental," (opens in a new tab) names service bookings among its use cases, and the same URL can redirect a person's browser to a normal checkout page (opens in a new tab).

What stops agents. Stripe asks agents that use its wallet to report what happened on each attempt, with tags that include captcha, anti_bot_script, cdn_block, waf_block, rate_limited, login_required, and 3ds_challenge (opens in a new tab). Each tag names something a seller controls on its purchase path, except the card issuer's security check (our reading).

Machine-readable pointers are optional. Schema.org defines checkoutPageURLTemplate as "A URL template (RFC 6570) for a checkout page for an offer," and says "Parameter naming and standardization are not specified here" (opens in a new tab). No assistant's documentation found for this guide says it reads that property or llms.txt when buying. Use structured data to repeat visible facts, as Google asks: make sure structured data "matches the visible text on the page" (opens in a new tab).

How to make a service easy for agents to find and buy

These steps are recommendations. Each draws on the documentation and research above.

1. Publish the offer as plain facts. State what is included, the price and currency, the turnaround, who does the work, the refund terms, and the inputs you need from the buyer, as text in the HTML the server sends. In December 2024, none of the major AI crawlers rendered JavaScript (opens in a new tab) (network measurement).
2. Keep the same facts everywhere. Match the site, marketplace profiles such as Clutch, and any feed or directory listing. In lab trials, consistent rather than contradictory claims raised a source's odds of being cited first in at least four of six models (opens in a new tab) (peer-reviewed; laboratory setting).
3. Let the buyer, or the buyer's agent, finish without a sales call where the service allows it: one purchase URL, guest checkout, labeled form fields, and every input collected before payment.
4. Keep the purchase path open to verified agents. Check that bot protection does not challenge the pages an agent needs, and verify signed agent traffic, such as ChatGPT's Web Bot Auth signatures, instead of blocking all automation.
5. Deliver without manual steps after payment where the service allows, and send a receipt and a status link the buyer's agent can read back to the buyer.
6. Decide whether an API or MCP server is worth building. Stripe documents booking and payment through either, with the customer approving each payment; volumes are unknown, and payers are limited to the United States and Canada for now.
7. Earn real ratings. Ratings were the one cue that raised selection across models in the research above. Collect them within the FTC's rule, which prohibits fake reviews and testimonials, including AI-generated ones, and incentives conditioned on a positive or negative review (opens in a new tab).
8. Measure. Log agent traffic on purchase pages, including the Signature-Agent header; watch for ChatGPT's utm_source=chatgpt.com on referral links (opens in a new tab); and ask every buyer how they found you. How to track AI referral traffic in GA4 covers the analytics setup.

What the evidence shows and does not show

Antipatterns

Each antipattern below is common and fails for a documented reason.

Checklist

Frequently asked questions

Can ChatGPT buy a service for me?

Not inside the chat. On paid plans, ChatGPT's cloud browser can work through a seller's website and asks for confirmation before a booking or a payment, and some tasks need a final step you complete yourself (see above). ChatGPT's shopping results link to merchants' sites, where purchases are completed.

Can I sell my service inside ChatGPT?

Not through an OpenAI commerce program today. OpenAI stopped offering Instant Checkout to new merchants on March 24, 2026 (opens in a new tab), its merchant application lists retail categories, and its plugins may sell only physical goods (above). A service business can still be found through ChatGPT search (opens in a new tab) and be bought on its own site. How to get recommended by ChatGPT covers the search side.

What happened to ChatGPT Instant Checkout?

It launched on September 29, 2025, for single-item purchases from U.S. Etsy sellers. On March 24, 2026, OpenAI said that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we're allowing merchants to use their own checkout experiences while we focus our efforts on product discovery" (opens in a new tab). OpenAI's help center still mentions Instant Checkout for some eligible products and merchants, and its plugin guidelines limit it to select marketplace partners (above).

Does Google's agentic checkout work for services?

No. Google's checkout in AI Mode and Gemini lists services among ineligible products, and Google's Shopping policies exclude offers of "labor, time, effort, expertise" (above). Google's AI Overviews and AI Mode can still cite a service business's pages; How to rank in Google AI Overviews and AI Mode covers how.

What is agentic commerce?

Buying and selling in which an AI agent carries out some of the shopping or checkout steps for a person. In September 2026 it runs mostly on retail product feeds and on browser agents that use sellers' own websites, with a person approving payment in every documented flow.

What is the Model Context Protocol, and does a service business need an MCP server?

MCP is "an open-source standard for connecting AI applications to external systems" (opens in a new tab). A business does not need one to be found: assistants find businesses through search. An MCP server helps when the service can be booked by software, since Stripe documents agents booking and paying through one (above).

Do I need llms.txt for AI agents?

No assistant documents reading llms.txt when it recommends or buys. Stripe's agent directory asks for a link to one when a business applies for a listing (above). llms.txt: what it is and whether it matters covers the evidence in full.

Will AI agents replace B2B buyers?

Not on current evidence. In G2's survey, most software buyers used or planned to use agents, but 9% were comfortable letting agents execute purchases within guardrails, and a review of 47 studies found little research on agents that actually complete checkout (above).

Next steps

The Instant AEO Audit is AEO HQ's $499 automated audit, one of the fixed-price packages on the pricing page.

Change log

• September 28, 2026: First published.