What is the x402 Bazaar, and how do agents find your endpoint?

What the Bazaar actually is

The Bazaar is Coinbase's discovery index for x402 endpoints. According to Coinbase Developer Platform, a seller lists an endpoint with its schema, price and description, and an agent finds the service programmatically. That is the product. No human storefront, no editor deciding what ranks.

The gap it fills is narrow. The x402 protocol, documented at x402.org, charges no fee of its own: payers cover network fees only, and it works across EVM chains and Solana. What the protocol does not do is tell a buyer which seller to choose. Payment and discovery arrived separately, and the Bazaar is the discovery half.

What a listing carries

A listing is a small record. The fields are unglamorous.

Coinbase frames the Bazaar as the machine-readable layer: sellers publish, agents query, matching happens with no human in between.

How an agent picks between two endpoints

An agent does not browse. It gets a task, queries an index, receives candidates, and chooses. That choice happens before any payment is built. The machinery in the x402 protocol repository (the specs/ directory, the exact payment scheme that transfers a specific amount, the PaymentPayload and PaymentRequirements objects, the facilitator role, the TypeScript, Python and Go SDKs) runs after the agent has decided what to buy. The normative x402 specification documents are hosted in the x402 protocol specifications (Coinbase (x402 repository)), covering the exact payment scheme and the payload and requirement objects that define what a payer signs and what a facilitator verifies and settles.

So the selection step carries the weight. Two endpoints serving the same task, priced at the same figure, returning the same JSON shape, are identical on every axis except one. Price does not break the tie. Schema does not break the tie. The description does.

Why the description decides

A schema tells an agent whether it can call the endpoint. A price tells it whether the call fits a budget. Neither says whether the response answers the question.

That leaves the description doing the work, and most sellers write it badly. "Crypto data API" loses to "returns the top 100 holders of an ERC-20 token by balance, refreshed every 15 minutes, historical snapshots on request". Same capability, different outcome.

This is the same mechanism at a lower layer. AEO compared with SEO makes the argument for pages: an answer engine retrieves a chunk, and the chunk it retrieves is the one whose language matches the query. An x402 listing is that chunk. If your description lacks the words an agent's task planner generates, the endpoint stays invisible however cheap it is.

Cloudflare's crawl data shows the shape of the imbalance. For the week of 19 to 26 June 2025, Cloudflare measured crawl-to-referral ratios of nearly 71,000 to 1 for Anthropic, 1,600 to 1 for OpenAI, 202.4 to 1 for Perplexity, 40 to 1 for Microsoft and 9.4 to 1 for Google. Those are 2025 figures and have improved by more than an order of magnitude since. Agents consume far more than they credit.

The demand underneath the listing

The market this index serves is thin, and that is the honest starting point.

According to TRM Labs, x402 has settled roughly $52.7 million across 198.9 million settlements on Base, Solana and Polygon since May 2025. Screened down, $25.62 million looks like genuine commerce, and 0.6% to 7.5% of that appears plausibly agentic. 99.6% of the value settled in USDC.

Chainalysis counted well over 100 million cumulative x402 transactions on Base through Q1 2026, and found payments of $1 or more rose from 49% of value in early 2025 to 95% by early 2026. The payments worth competing for are now the large ones, and those buyers run a selection step.

CoinDesk, citing Artemis, put real daily x402 volume at around $28,000 with an average payment near $0.20, against roughly 131,000 daily transactions.

A listing is cheap to publish and the current return is close to zero. Publish one anyway. The discovery layer is being standardised now, and the endpoints that populate it early are the ones the index will know about.

What to put in the description

Write for the task planner, not for a person.

• Name the output, not the category. "Top 100 ERC-20 holders" beats "on-chain analytics".
• State the refresh cadence and the historical depth. Agents filter on freshness.
• Name the chains and the asset classes explicitly.
• Give the input format in the same sentence as the output format.
• Drop brand language. An agent has no prior on your brand.

The discipline is the same one that applies to answer engines, covered in where AI engines cite crypto. If you have done that work for your site, you have most of the listing written already.

For the broader decision about whether to open an endpoint at all, Should you charge AI crawlers, block them, or leave them alone? covers the trade-off, and Cloudflare's Monetization Gateway: pricing the web for agents covers what changes when a CDN becomes merchant of record.

An endpoint an agent cannot find earns nothing. The description is the only field in the listing that decides whether it gets found.

Kunzum's own corpus is auditable against this claim. We publish 1,500 raw AI responses and 12,518 citations from our measurements, so the pattern of what gets cited and what gets ignored can be checked rather than taken on trust.