Readable and payable: the two halves of selling to AI agents

Retrieval decides whether the agent wants the page

An agent that cannot find your page never reaches the moment where it could pay for it. Most writing about x402 skips that. The protocol is interesting, but a payment rail with nothing attached to it settles nothing.

The rail is becoming standard. The Linux Foundation announced the x402 Foundation on 2 April 2026 at the MCP Dev Summit North America in New York, with Coinbase contributing the protocol. It launched operationally on 14 July 2026 with 40 member organisations. The 17 premier members include Adyen, AWS, American Express, Circle, Cloudflare, Coinbase, Fiserv, Google, Mastercard, Ripple, Shopify, Stripe and Visa. According to x402.org, the protocol charges no fee of its own, payers cover network fees only, and it works across EVM chains and Solana.

Now look at what flows through it. According to TRM Labs, roughly $52.7 million settled across 198.9 million x402 settlements on Base, Solana and Polygon since May 2025. Screened down to payments that look like commerce, that falls to $25.62 million. Between 0.6% and 7.5% of the screened total appears agentic. $52.47 million of $52.68 million, or 99.6%, settled in USDC.

198.9 million settlements carrying $52.7 million tells you the average settlement is small and much of the activity is not commerce. BlockEden.xyz records daily x402 transactions falling from about 731,000 in December 2025 to about 57,000 in February 2026, a decline of over 92%. 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. Chainalysis counts well over 100 million cumulative x402 transactions on Base through Q1 2026, and notes payments of $1 or more rose from 49% of value in early 2025 to 95% by early 2026.

So the payment layer works. The demand layer is thin, and it is shifting toward larger, fewer payments.

What the corpus shows about what engines actually cite

Kunzum measured the retrieval half directly. 1,500 responses, 12,518 citations, 1,396 domains, published under CC BY 4.0 with the scoring scripts and no email gate. How the corpus was measured covers the collection. Where AI engines cite crypto covers where the citations land.

Retrieval rates in edition one:

• Perplexity retrieved on 100% of answers (120 of 120).
• Claude retrieved on 65% (78 of 120).
• ChatGPT retrieved on 54% (65 of 120).

Google AI Overviews always retrieves, so its web_search_ran column is unset. It is not 0%. Nobody should report a retrieval rate for it.

Those rates are the top of the funnel. 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. Date them whenever you quote them.

The two halves ask different things of the same page

Both halves reward a page that states plainly what it is. Vague copy gets skipped by a retriever and reads as risk to an agent deciding whether to spend.

Where they pull apart is what happens after the fetch. Retrieval rewards being copied into someone else's answer. Payment rewards being consumed on your own terms. If a model can lift the whole answer into a summary, the agent has no reason to buy the page. That tension is the design problem, and it does not resolve itself by adding a price tag.

Payable without retrievable settles nothing

Cloudflare is the cleanest illustration of how much refusal is already happening. W3Techs puts Cloudflare in front of 25.8% of all websites as of the September 2026 survey. Cloudflare's chief strategy officer, quoted by CoinDesk at Consensus in May 2026, described over a billion HTTP 402 responses a day on the network. That is a quoted remark rather than a published statistic. Treat it as a signal about scale, not a measurement.

The tooling exists. Cloudflare's Monetization Gateway, published 1 July 2026, lets customers charge for any protected asset: pages, datasets, APIs or MCP tools, settling in stablecoins over x402. Pay Per Crawl lets publishers price crawler access with Cloudflare as merchant of record, using HTTP 402, and is in private beta. AI Crawl Control lets publishers see and sort AI crawlers by purpose instead of blocking them wholesale.

None of it helps if the agent never considered your page. An agent that routes around you has no reason to send a 402, and no reason to pay one. You have built a checkout on a street nobody walks down.

Do retrieval first, and why

Order matters because the two halves have different cost structures. Payment is a configuration job you can finish in an afternoon. Retrieval is editorial, and it takes weeks to show up in a model's answer mix.

1. Make the page extractable. One claim per section. Numbers with their sources attached in the sentence.
2. Make the page citable. Name the source and link it, so the claim survives being lifted as a chunk.
3. Measure whether engines actually retrieve it.
4. Only then put a price behind it.

AEO compared with SEO covers the measurement side of step three. The short version: if you are not in the answer, the payment option is decoration. You can have a perfect 402 response and a settled wallet and still earn exactly nothing, because the agent solved its problem somewhere else.

That is the whole reason Kunzum publishes the corpus rather than the conclusion. Retrieval is measurable today, on your own pages, with your own queries. Payment demand is not, at least not yet. Start with the half you can check. The free playbook walks through the retrieval checks in order.