Agentic payments: integration and readiness. 6 chapters and a final quiz.
The operating guide for merchants and PSPs: make your catalog readable by agents (feeds, schema.org, MCP), identify and accept an agent (Web Bot Auth, KYA, delegated tokens), manage fraud and disputes, deploy x402, then roll out a roadmap and governance.
Structure a catalog that agents can read: product feeds, schema.org, MCP servers
Set up cryptographic agent identification (Web Bot Auth) and a KYA policy
Work with delegated tokens and verifiable credentials in payment flows
Map out liability for fraud or disputes involving an agent
Chapter 1. A machine-readable catalog.
An agent doesn’t see your website. Not your retouched photos, not your optimized checkout funnel, not your banners. It consumes structured data. If your products don’t exist in machine-readable form, they don’t exist for ChatGPT, Gemini, or Claude, however good your SEO for human visitors is. Agentic visibility is won across three complementary channels.
🔑
A new baseline
In 2026, machine readability determines whether an agent recommends a product, and everything hinges on complete attributes and fresh prices and inventory. A beautiful product page with no GTIN and no structured availability stays invisible to conversational commerce.
The 12 non-negotiable attributes
Title, description, brand: the semantic foundation the agent compares
GTIN (EAN/UPC code) and MPN: to match your offer with other sellers’ offers
Standardized category (a product taxonomy such as Google’s): to target queries
Price, sale price, currency: the agent compares down to the cent
Availability and condition (new/refurbished): an undeclared stockout means a rejected order and damaged trust
Image URL and canonical product URL: for display within the conversation
A standardized feed pushed to agent platforms (OpenAI’s feed specification for ChatGPT, Google Merchant-style feeds). Freshness requirements apply to prices and inventory, in near real time.
🏷️
On-page schema.org
JSON-LD markup on every product page. This passive channel is essential for agents that browse the open web rather than private feeds.
🔌
MCP server
The Model Context Protocol (an open standard Anthropic published in late 2024) exposes catalog search, inventory, and checkout as tools the agent can call. This channel is active and conversational.
In 2026, it pays to cover all three channels, because the mix of agentic traffic is shifting too fast to bet on just one. Schema.org markup is the floor: low cost, maximum reach. The ACP/OpenAI feed unlocks Instant Checkout. The MCP server lays the groundwork for deep integrations, including B2B, where the buyer’s agent queries your ERP directly.
🎯 Quick question
Why does structured markup matter more than site design in agentic commerce?
Chapter 2. Agent identification: Web Bot Auth and KYA.
Before accepting a payment from an agent, you need to know who it is. The HTTP User-Agent is self-declared, so it can be spoofed; IP address lists are brittle. The answer now being standardized, Web Bot Auth, has the agent cryptographically sign every request (HTTP message signatures, Ed25519 keys). The server then verifies it against a directory of public keys.
Verifying an agent with Web Bot Auth
Agent
Signs its HTTP request
Signature and Signature-Agent headers pointing to its key directory
➜
Edge / WAF
Fetches the public key and verifies the signature
An unknown or improperly signed agent is treated as an ordinary bot
➜
Merchant
Applies its policy
Allow, restrict (rate limit, limited catalog), charge, or block
➜
Payment layer
Matches the web identity with the payment identity
Visa, Mastercard, and Amex are working with Cloudflare to link Web Bot Auth to their trusted-agent programs
Cloudflare and its partners push Web Bot Auth toward standardization.
March 2026
Production launch
Agent signature verification at Cloudflare’s edge.
April–August 2026
IETF milestones
Specifications sent to the IESG (April), key management best practices (August); an RFC is expected in 2027.
KYA: Know Your Agent
Technical identification isn’t enough. You also need to vet the agent, the way KYC vets a bank customer. KYA (Know Your Agent) cryptographically binds the agent to a registered operator (its developer) and to the delegating user, with metadata: type of activity, use case, intent. A merchant or PSP can then grade its level of trust.
Level
What you know
Typical policy
Anonymous
Nothing verifiable (self-declared User-Agent)
Treat as a bot: block or strict rate limit
Identified
Valid Web Bot Auth signature, known operator
Catalog access, no checkout
Verified (KYA)
Registered operator + attestations about the agent
Checkout allowed within limits
Mandated
KYA + user mandate credential (AP2-style)
Full checkout, proof archived
Trust levels for an incoming agent
⚠️
Don’t block blindly
Blocking all bots now means turning away customers, because agent traffic converts better than average. The goal is to differentiate: welcome verified agents, and charge or turn away the rest, rather than put up a wall.
🎯 Quick question
What does Web Bot Auth add compared with the traditional HTTP User-Agent?
Chapter 3. Delegated tokens and verifiable credentials.
One rule governs the entire architecture. An agent must never hold a PAN, nor unlimited access to a payment method. All the engineering work of 2025–2026 is about handing it narrow, revocable, traceable powers. Two families of objects do this: delegated tokens protect the instrument, and verifiable credentials prove the mandate.
Agentic Token or Shared Payment Token; the issuer sees the “agent” signal
➜
Archiving
Mandate + token + timestamps retained
The evidence file for any later dispute
🔑
The credential is your proof of consent
Timestamped and non-repudiable, a verifiable credential signed by the cardholder works as a digital sign-off and makes any tampering detectable. Archived with the transaction, it turns an “I never authorized this” dispute into a simple check of the signature and the limits.
Who issues whatVisaMastercardStripeGoogle PayPayPal
For a PSP, the target architecture resembles network tokenization with one extra dimension: each token is indexed not only by merchant and instrument, but also by agent and by mandate. Revocation has to work on all three axes. A compromised agent, an expired mandate, and an instrument that needs to be blocked are each handled independently.
🎯 Quick question
How does a delegated token’s role differ from a mandate verifiable credential’s?
Chapter 4. Fraud and disputes: allocating liability.
The classic cardholder-merchant-issuer triangle gains a fourth corner. The agent platform enters the liability chain, whether it is OpenAI, Google, or an assistant developer. Disputes change in nature: fewer stolen cards, more “my agent exceeded its mandate” or “the product delivered doesn’t match what the agent ordered.”
× 2.4
dispute rate on agent transactions vs. comparable human card-not-present transactions
TrustSphere, 2026
78 %
of financial institutions expect fraud involving shopping agents to rise
Industry survey cited by Solutions Numériques, 2026
Jan. 2026
Opinion attributed to the CFPB: the existing dispute framework applies to agent transactions
TrustSphere, 2026 (CFPB text not published)
Scenario
Who bears the risk in practice
Your defense
The agent buys outside its mandate (“hallucination” or drift)
The agent platform, if the archived mandate proves the overreach; otherwise the merchant absorbs the chargeback
Require and verify the mandate (limits, expiration) before completing the order
The customer regrets a purchase that did fall within the mandate
The consumer: their recourse is bounded by the properly signed mandate (the reading attributed to the CFPB)
Archive the signed credential + timestamps + cart details
Compromised user or agent account
Depends on who was negligent: the platform (poorly protected keys) or the user; a merchant that verified the agent in good faith is better protected
Only accept Web Bot Auth / KYA agents, with low default limits
Merchandise not delivered or not as described
The merchant, as in standard e-commerce—card chargebacks still apply
Delivery and proof processes unchanged, SLAs on agent orders
Disputed x402 payment
The buyer: on-chain settlement is final, with no chargeback
Your own contractual refund policy, if you sell via x402
Dispute scenarios and indicative allocation of liability
⚠️
The archived mandate is your defense file
In representment (chargeback defense), the key document is no longer the 3-D Secure proof but the signed mandate credential, together with the execution log. A merchant that can’t produce the mandate tied to an agent transaction starts out on the losing side. Keep these records for as long as your dispute windows run: 13 to 18 months depending on the scheme.
Tag every agent transaction as such in your systems (a dedicated flag, an “agentic” channel) so you can run separate reporting and fraud rules.
Set specific limits per agent and per agent operator, separate from cardholder limits.
Monitor dispute reasons: a high share of “not as described” often signals a problem with your feed quality, not fraud.
Put contracts in place with agent platforms: commissions, liability, dispute procedure, access to logs.
A reminder for Europe: the AI Act (phased in from August 2026 to 2027) doesn’t specifically address shopping agents. Consumer protection law and PSD2 remain your frame of reference.
🎯 Quick question
Why should a high share of “not as described” disputes on agent orders worry a merchant?
Chapter 5. x402 in practice: HTTP 402 + stablecoin.
x402 beats cards when the buyer is a machine and the cart is tiny. Pay-per-call APIs, single news articles, datasets, AI inference, premium content for crawlers: in all these cases, creating an account and paying a 30-cent fixed card fee makes no sense. x402 settles in USDC, in under a second, for a fraction of a cent.
Express server: protecting a route with x402 (illustrative)
Middleware on your routes, or the edge layer (Cloudflare, AWS CloudFront)
➜
Machine client
Receives the 402, signs the payment, retries the request
x402 SDK on the agent side: the protocol handles the negotiation
➜
Facilitator
Verifies the signature and submits the on-chain settlement
You don’t run a blockchain node or monitor the mempool
➜
You (the seller)
Serve the resource and collect the USDC
Optional conversion to euros through your crypto provider
169M
x402 transactions in the first year
Coinbase, 2026
2 weeks
between Cloudflare’s and AWS’s x402 integrations at the edge
InfoQ, July 2026
20+
members of the x402 Foundation (Linux Foundation): AWS, Cloudflare, Anthropic, Circle, and others
x402 Foundation, 2026
⚠️
Compliance comes before the first line of code
Accepting stablecoins in Europe brings you within the scope of MiCA. Conversion and custody go through a licensed crypto-asset service provider (CASP), accounting and VAT treatment must be settled, and the absence of chargebacks calls for a contractual refund policy. Many merchants choose a provider that settles in euros and absorbs the crypto layer.
Content publishers have a complementary use case. Cloudflare’s pay-per-crawl, whose dashboard has been unified across all plans since April 2026, lets you charge AI crawlers for your pages without writing code. You set a price per request and Cloudflare collects payment as merchant of record, a model the Monetization Gateway extends, via x402, to any resource behind its network.
🎯 Quick question
For which sales profile is x402 most relevant?
Chapter 6. Integration roadmap and governance.
There’s no need to build everything at once: the market is moving too fast for heavy bets. The right approach is incremental. Become visible to agents first, then sellable, and finally industrialized, with governance guardrails at every step.
Weeks 1–2
Machine-readability audit
Crawl your own site the way an agent would: schema.org completeness, feed accuracy, how your WAF treats verified bots.
Weeks 3–6
Structured data brought up to standard
All 12 attributes on 100% of the catalog, near-real-time price and inventory refresh, feed quality monitoring.
Weeks 7–10
Choosing and plugging in protocols
Feed + ACP if your PSP supports it (one line of code with Stripe), an agent access policy (Web Bot Auth) at the edge, x402 if you sell API access or content.
Weeks 11–13
Controlled pilot
Flagged agentic channel, low limits, restricted product scope, tracking of conversions and disputes.
Next quarter
Industrialization
Broader catalog, contracts with agent platforms, VC mandates built into anti-fraud tooling and representment.
Governance: the permanent guardrails
🎚️
Limits and scopes
Limits per agent, per operator, and per mandate, separate from cardholder limits. By default they stay low and widen as history builds up.
🧾
Evidentiary logging
Signed mandates, tokens, timestamps, and responses archived for the full dispute window. Together they form your evidence file.
🛑
Kill switch
Immediate revocation of a compromised agent, operator, or mandate, without shutting down other channels.
Getting ready for agentic payments is 20% protocols and 80% fundamentals: product data quality, rigorous identification of counterparties, archived proof of consent, and revocable limits. Merchants that treat the agent channel as a sales channel in its own right, with its own KPIs, contracts, and governance, will capture 2026 demand without taking on its risks.
🎯 Quick question
What is the recommended order for a merchant’s agentic integration?