# How Agentic Is Agentic Commerce? A Population-Scale Measurement of x402 Adoption and Authenticity

Crypto.world

> The x402 protocol lets AI agents pay for data, APIs, and compute by settling a stablecoin payment on-chain, and its hundreds of millions of settlements are widely cited as proof that an "agent…

**arXiv**: 2607.12575

**Published**: 2026-07-14

**Authors**: Shengchen Ling, Yajin Zhou, Lei Wu, Cong Wang

## AI summary

## Overview

**Research area:** Blockchain measurement and security economics — specifically on-chain measurement of "agentic payment" protocols (x402) on Base and Solana, with a focus on distinguishing genuine third-party demand from operator-manufactured activity.

**Technical level:** Intermediate. The paper assumes familiarity with stablecoin transfers, meta-transactions (EIP-3009 authorizations, relayers, gas sponsorship), and basic graph analysis; the prose itself is largely non-mathematical.

**Scope (one sentence):** A 280-day, population-scale census of every x402 settlement on Base (with a coarser Solana census), which classifies each settlement as fictitious, operator-internal, or unattributed using a value-flow graph, and then characterizes the operator-driven economy that produces most of the volume.

## What This Paper Is About

The x402 protocol lets AI agents pay for data, APIs, and compute by settling a stablecoin payment on-chain, and its hundreds of millions of settlements are widely cited as proof that an "agent economy" has arrived. The authors argue that settlement count is the one metric an interested party can manufacture almost for free — the facilitator pays the gas, and nothing on-chain records who actually controls a payment. The paper's goal is therefore to measure how much x402 activity is real external demand versus payments an operator makes among addresses it controls, and to bound genuine adoption rather than report a raw headline count.

## Key Contributions

1. **A reproducible measurement methodology and public artifact.** A pipeline that identifies x402 settlements by intersecting the USDC `AuthorizationUsed` event with a curated facilitator allowlist, resolves the true payer through the meta-transaction layer, enforces an explicit counting discipline (every quantity names its base set and measure), and classifies every settlement on a directed value-flow graph. The authors state they will release all data and implementation on publication.
2. **The first population-scale measurement of x402 adoption and authenticity.** A 280-day census (2025-09-17 to 2026-06-23) on both deployment chains, establishing that 84.98% of Base's 136,708,672 settlements are operator-internal, of which at least 21.20% by count (54.08% by value) is provably manufactured.
3. **A bounded estimate of the genuinely independent economy.** Two figures bracket it: $187,861.35 that demonstrably reaches a nameable catalog service, and $20,258,746.09 (45.92% of value) that is not provably manufactured — a two-orders-of-magnitude range the chain cannot narrow.
4. **An anatomy of the manufactured economy.** From signals orthogonal to the payments (structure, timing, capital, price, dynamics), the paper shows the activity is a coherent operator-driven system — a forest of machine-timed stars circulating recycled capital — reproducible in its entirety for about $355,583 in sponsored gas, making settlement count a Goodhart metric.

## Main Findings

- **The headline count is large.** Over the 280-day window, Base carries **136,708,672** x402 settlements worth **$44,121,383.81**, with **491,587** payers, **116,781** recipients, and **129** listed facilitator relayer EOAs. Solana adds **49,477,928** settlements worth **$9,032,706.12** (227,133 payers, 23,743 recipients, 23 facilitator fee-payer signers) under a coarser counting unit. The authors resolve **30** facilitators in total.
- **Identification is a bounding problem, not a lookup.** Counting every `AuthorizationUsed` event gives a loose upper bound of **140,274,969** settlements and **$772,090,109.10** gross. Filtering by relayer address instead over-counts by **54.8×**: the **5,587** addresses that ever relayed a settlement sent **$42,348,540,765** in USDC. Only the intersection of the event with a curated facilitator allowlist yields the reported census.
- **Activity is extremely concentrated on every axis.** Base payer, recipient, and value Gini are **0.98702**, **0.99808**, and **0.98872**; Solana's are **0.98463**, **0.99526**, and **0.98054**. Base's largest recipient takes **23.08%** of settlements (recipient Nakamoto coefficient 3) while the largest payer takes only **8.44%** (payer Nakamoto 112). Solana funnels into one hub: largest recipient **64.52%** (Nakamoto 1), largest payer **1.93%** (Nakamoto 271). Facilitator Nakamoto is **1** on both chains (largest facilitator **63.54%** on Base, **67.72%** on Solana).
- **The facilitator layer hides a monopoly behind a round-robin.** Base's 129 relayer EOAs give an HHI of **292**, which would normally read as competitive. But relayers ranked 4th through 12th each carry between **3.24%** and **3.26%** of settlements — nine shares equal to within about two hundredths of a percentage point, which the authors read as one operator load-balancing its own traffic.
- **Advertised supply barely materializes.** From the CDP Bazaar catalog, **25,411** resources across **37,338** payment accepts, of which **25,163** advertise a Base option, deduplicate to only **811** distinct `payTo` addresses across **910** distinct hosts. Two template factories (`lowpaymentfee.com`, 10,035 listings, and `orbisapi.com`, 9,461 listings) produce **77.48%** of listings while earning roughly **$2,234** over their lifetime. **624** of 811 recipients (**76.94%**) ever settled; only **249** (**39.90%** of those) earned at least $10, at a median lifetime revenue of **$3.96**. Of the 624, **423** earn at least $1, **291** at least $5, **134** at least $50, and **97** at least $100.
- **The off-chain surface has decayed.** A single-round, GET-only, no-payment probe of the **910** Base hosts found **805** (**88.46%**) still returning an HTTP response but only **474** (**52.09%**) returning a live HTTP 402 challenge; **105** returned no response and **331** returned a response without a challenge.
- **The catalog's multi-chain and multi-scheme breadth is illusory.** Of the **259** resources advertising the dynamic-pricing "upto" scheme, **185** (**71.43%**) come from a single address that has never settled a payment on Base (zero transactions, an unfunded EOA); the 11 addresses advertising it earn between **$285.88** and **$1,040.63** lifetime. Of all **25,411** entries, **10,681** (**42.03%**) advertise a Solana option, but **99.72%** of those (**10,651**) also advertise Base.
- **A small set of named operators dominates.** Named operators account for **67.76%** of settlements (**92,640,016**): lnpay **23.08%**, t54 **21.74%**, and 2025-11 self-payment activity **20.13%**. Resolving wallets to operators collapses the payer-side Nakamoto coefficient from **112** to **3**.
- **Fictitious activity (C1) is 21.20% of settlements and 54.08% of value.** Self-payment alone is **20.13%** of settlements, essentially a single 2025-11 campaign: **27,089,029** of **27,514,640** self-settlements (**98.45%**) come from three wallets first funded from Binance and Gate that all cease within the same two-second window (last settlement 2025-11-28 04:04:27–29Z). Four fully-closed clusters add to the tier, moving **$779,488.96** with external outflow of exactly **$0.00**; the largest returns its whole **$422,064.04** to its own funder.
- **Internal settlement (C2) is 63.78% of settlements.** Clusters have inflation factors (total settlements divided by those terminating outside the cluster) spanning orders of magnitude, the highest above **7,801×**. lnpay's funder seeds **337** fleet wallets in identical 50.00 or 40.00 USDC amounts; the fleet sends **31,551,433** transactions to the hub and none to an external seller; **98.34%** of the hub's **$415,765** inflow (**$408,848**) returns to that funder. The fleet's own payments are funded by **$1,261,430** of x402 receipts, **99.98%** of which come from other wallets in the same fleet (**$314** from outside). t54/lucyos drives **29,538,389** settlements with only **272,779** (**0.9%**) reaching outside, and **70.33%** of its payers first settled on one synchronized day. ainalyst's four mega-payers drive **8,884,277** settlements against **1,139** external, with the hub repaid **16 seconds** after funding a payer.
- **Unattributed (C3) is 15.02% of settlements.** That is **20,529,628** settlements carrying **$12,835,401.84** (**29.09%** of value). A residual behavioral model trained on a human silver set, applied to the tail's **457,400** payers, has a majority vote flag **87.82%** as "manufactured-behaving" (95% CI [87.73%, 87.92%]) — presented as an estimate of an unresolvable residual, not as validation.
- **Genuine independent demand is bounded two orders of magnitude apart.** At most **$12,835,401.84** (29.09%) if only C3 counts; at most **$20,258,746.09** (45.92%) if C2's **$7,423,344.25** is granted the benefit of the doubt. Within that ceiling, only **$187,861.35** demonstrably reaches a nameable catalog service. Only **234** catalog sellers took in at least $10 of lifetime x402 value (fewer than the 249 in the supply funnel, which counts total lifetime revenue including non-x402).
- **The manufactured economy is structurally a star forest, not a marketplace.** The settlement graph has **536,690** nodes and **718,668** edges; its giant weakly-connected component (**409,360** nodes, **76.27%**) carries **99.2671%** of settlements. Weighted asynchronous label propagation recovers operator-centered stars whose membership is stable across 10 seeds (t54 Jaccard 0.9976–0.9995, median 0.9983; lnpay 0.9227–0.9624, median 0.9444). The largest communities by settlement volume are t54 (**24.96%**) and lnpay (**23.10%**), and the top-20 stars together carry **93.03%** of all settlements.
- **The whole count is nearly free to reproduce.** Because facilitators sponsor gas, the entire **136,708,672**-settlement count is reproducible for about **$355,583** in gas, paid by the facilitators.

## Methodology in Plain English

The authors start from a problem of definition: no transaction says "I am x402." A settlement's only on-chain trace is a generic EIP-3009 `AuthorizationUsed` event, submitted by the facilitator's relayer rather than by the paying agent. Counting all such events is too loose; filtering by relayer address is far too loose. So the authors take the **intersection** of the event with a facilitator allowlist built by merging two independent curations (Allium's on-chain index and the x402scan explorer), resolving 30 facilitators and 129 relayer EOAs.

They then resolve roles before any grouping: the payer is the authorization signer (cross-checked against `Transfer.from`), the recipient is `Transfer.to`, and the relayer is `transaction_from` — so the party that consented to the payment, not the party that broadcast it, is what gets counted.

To test authenticity, the authors survey the standard wash-trading, Sybil, and invalid-traffic detection toolkit (self-dealing, closed loops, absence of economic substance, value anomalies, coordinated timing, structural exclusivity) and argue that agentic payments neutralize most of it: the legitimate actor is itself a bot, dust-sized round payments are normal, and custodial agent-wallet platforms legitimately fund many payers from one treasury. What survives is **economic substance** — whether value reaches a party outside the transactor's control — because a facilitator relays a payment to the named seller and never returns funds, so a hub sweeping receipts back to the treasury that funded its payers is a second, deliberate transfer closing a loop.

They build a directed, layered value-flow graph whose edges are three kinds of USDC transfer: settlements (payer to hub), funding (a wallet's inbound capital), and sweeps (a hub's outbound transfers). Addresses are merged into clusters by two conservative rules — an exact shared funder, or a vanity-suffix address batch — with exchange, router, and relayer hubs held out as opaque boundaries so a high-degree shared address cannot falsely merge unrelated operators. Fund-cycling is independently confirmed by strongly-connected-component and cycle enumeration. Every settlement is then sorted by what its endpoints prove: **C1 fictitious** (self-payment, or a provably closed cluster with zero external outflow), **C2 internal settlement** (payer and payee in one funding-linked cluster that is not provably closed), or **C3 unattributed** (the payee lies outside any resolved cluster, or the payer is in no cluster). Verdicts attach per settlement and only assert the direction each tier supports — C3 is never claimed to be genuine.

Finally, they characterize the manufactured economy from signals orthogonal to the settlements themselves — graph connectivity, timing regularity, capital flow, price, and dynamics — and cross-check the off-chain advertised surface by matching CDP Bazaar catalog listings to on-chain recipients and by a minimal, GET-only, no-payment liveness probe of listed hosts.

## Why This Matters

The paper reframes a metric that the ecosystem treats as evidence of adoption. Because gas is sponsored and nothing on-chain marks who controls a payment, settlement count measures the incentive to transact, not demand — and any producer can inflate it at near-zero marginal cost. The authors' three-tier scheme gives a defensible lower bound (only what the trace proves) and an explicit upper bound, replacing a single headline number with a stated range.

**Real-world applications:**

- **Protocol and standards design:** the analysis points at concrete gaps (no on-chain marker distinguishing x402 from any EIP-3009 transfer, no authoritative facilitator registry, no scheme-level separation of `upto` from `exact` settlements) that a standards body or facilitator could close.
- **Investor and analyst diligence:** anyone citing x402 settlement counts as market traction now has a method, and specific public numbers, for discounting them before valuing an agent-payments business.
- **Platform and marketplace integrity:** the value-closure test generalizes to any gas-subsidized or custodial payment rail where a treasury funds many wallets, including custodial agent-wallet providers who need to distinguish their own internal ledger from merchant adoption.
- **Catalog and discovery curation:** the finding that two template factories produce 77.48% of listings while earning roughly $2,234, and that only 52.09% of hosts return a live 402 challenge, is directly actionable for resource-discovery catalogs.

**Industry relevance:** agentic payments are being marketed on exactly the metric this paper shows to be manufacturable. The $355,583 figure for reproducing all 136,708,672 settlements, and the two-order-of-magnitude gap between $187,861.35 and $20,258,746.09, are the numbers a due-diligence or benchmarking audience needs.

## Future Directions

- **Re-measurement as the ecosystem changes.** The authors explicitly frame the pipeline as reusable so new windows, facilitators, and chains can be re-measured; whether the C3 tail grows as agent frameworks mature is an open question the single 280-day window cannot answer.
- **Shrinking the two-order-of-magnitude bound.** The gap between $187,861.35 and $20,258,746.09 exists because C2 clusters cannot be adjudicated on-chain — a custodial platform's internal ledger and a wash pool look identical. Any off-chain corroboration (service-side logs, custodial disclosures) that resolves even a sample of C2 value would tighten the bound substantially.
- **Validating or replacing the residual model.** The 87.82% "manufactured-behaving" estimate for the C3 tail depends on a human-annotated silver set, and the authors themselves flag that genuine agentic payments and manufactured traffic share signatures, so annotators and any model trained on their labels can carry stereotype bias. Testing that estimate against independent ground truth is left open.
- **Designing a non-manufacturable adoption metric.** If settlement count is a Goodhart metric, the natural next step is a metric that resists gas-subsidized inflation — for example one anchored to value that provably exits a transactor's cluster, or to catalog services with sustained revenue rather than one-off activation.

## Target Audience

Most valuable to blockchain measurement and security researchers; to protocol designers and standards participants working on agentic payments; to analysts, investors, and journalists who report x402 or "agent economy" adoption figures; and to engineers building custodial agent-wallet platforms or resource-discovery catalogs, who will recognize their own architectures in the C2 clusters. Readers need only moderate familiarity with stablecoin transfers and on-chain events — the graph methodology is explained conceptually, and the paper is careful to state its base sets and measures throughout. Note that the supplied text is truncated partway through §5.1.2, so the full characterization of the manufactured economy along timing, capital, price, and dynamics (§5.2–5.5), the discussion and related-work sections (§6–7), and the conclusion (§8) are only summarized here at the level the abstract and introduction report.

## Authors’ abstract

AI agents are said to be forming an economy in which they pay, on their own, for the data, APIs, and compute they consume. x402, which settles a stablecoin payment on-chain for each purchase, is the most widely deployed protocol for this, and its hundreds of millions of settlements are read as proof that the economy has arrived. We show the count cannot be read as adoption: it is the one metric an interested party can manufacture almost for free, since the facilitator…

[The full abstract on arXiv](https://arxiv.org/abs/2607.12575)

[Read the original paper](https://arxiv.org/abs/2607.12575)

---

Canonical: https://crypto.world/paper/how-agentic-is-agentic-commerce-a-population-scale-measurement-of-x402-2607.12575
