How Agentic Is Agentic Commerce? A Population-Scale Measurement of x402 Adoption and Authenticity
Source: arXiv:2607.12575 · Published 2026-07-14 · By Shengchen Ling, Yajin Zhou, Lei Wu, Cong Wang
TL;DR
This paper investigates the true extent of adoption and authenticity of the x402 agentic commerce protocol, which implements AI agents paying on-chain for services via stablecoin settlements often referenced as evidence for a new 'agent-driven economy.' The authors argue that naive settlement counts are misleading because a large fraction of these transactions are manufactured internally by operators who sponsor gas fees and recycle capital, making settlement counts a poor proxy for genuine demand. They conduct the first population-scale measurement of x402 on the Base blockchain over 280 days, supplementing with coarser data from Solana. Their analysis resolves true payers and clusters wallets into operators and demonstrates the economy is highly concentrated and dominated by synthetic internal activity. Only a small bounded range of settlement value clearly corresponds to independent third-party demand. They show that the bulk of activity (over 84%) is operator-internal, including 21% of strictly fictitious settlements and 64% internal cluster transfers. The operator-driven economy is star-shaped, machine-timed, and fuelled by gas-subsidized capital recycling. Settlement counts reflect manufacturability incentives rather than real adoption.
The work contributes a rigorous, reproducible methodology combining on-chain event identification, meta-transaction role resolution, operator clustering by funding links, and layered value-flow graphs to classify settlements into fictitious, internal, or unattributed categories. Their dataset of 136 million Base settlements worth $44 million reveals extreme concentration (Gini > 0.98) across payers, recipients, and facilitators. While the protocol ecosystem advertises over 25,000 resources, only a fraction (249) have meaningful revenue and just over half of hosts respond with a payment challenge. The manufactured activity can be recreated by a single operator at low gas cost (~$355K), exposing settlement count as a classic Goodhart metric. This challenges industry claims of a realized agentic economy and calls for caution interpreting raw settlement data as adoption.
Key findings
- Base chain carried 136,708,672 x402 settlements worth $44,121,383.81 over a 280-day window; Solana added 49,477,928 settlements worth $9,032,706.12 under a coarser count.
- Payer, recipient, and value distributions are extremely concentrated with Gini coefficients above 0.98 and Nakamoto coefficients of 112 payers and 3 recipients on Base, 271 payers and 1 recipient on Solana.
- 21.20% of Base settlements are classified as fictitious (self-payment or closed-loop clusters) accounting for 54.08% of settlement value, provably performing no economic work.
- 63.78% of settlements are internal settlements within funding-linked clusters (shared operators), accounting for a majority of remaining value.
- Only 15.02% of settlements are unattributed, i.e., potentially genuine third-party demand; only $187,861.35 demonstrably reaches a nameable external service.
- A large proportion of the advertised 25,163 x402 resources collapse into just 811 distinct payees, with only 249 payees earning over $10 lifetime revenue and 52.09% of hosts still returning a live HTTP 402 payment challenge.
- The largest identified operator hubs (lnpay, t54) together control nearly 66% of settlements; their economies form star-shaped, machine-timed cycles recycling operator capital.
- The entire large settlement count could be reproduced for approximately $355,583 in sponsored gas, highlighting its manufacturability.
Threat model
An adversary is an operator or facilitator controlling multiple payer wallets, relayers, and recipient hubs who can manufacture settlements at near zero marginal gas cost by sponsoring fees and recycling capital to inflate settlement counts. They seek to create the illusion of genuine adoption to attract investment or market confidence. The adversary can abuse meta-transaction layering and on-chain indistinguishability of actors but cannot forge or erase on-chain funding traceability with current transparent blockchains. They cannot meaningfully obfuscate all value-flow or recreate genuine external demand patterns perfectly.
Methodology — deep read
The authors establish a multi-step pipeline to measure x402 adoption and authenticity. First, they define the threat and measurement challenge: the facilitator sponsors gas fees and the on-chain artifacts do not directly identify who truly controls a payment, making it easy to manufacture large numbers of settlements indistinguishable from genuine demand.
They collect on-chain data on the Base blockchain over a 280-day window (September 2025 to June 2026). They use the USDC contract's EIP-3009 AuthorizationUsed events combined with a curated allowlist of known x402 facilitators (from Allium's index and x402scan) to precisely identify valid x402 settlements. For Solana, the method is coarser, based on USDC SPL transfers where the fee-payer signer matches a facilitator address.
They resolve the payer (authorization signer), recipient (transfer.to), and relayer (transaction_from) roles for each settlement, as meta-transactions mask the actual payer on-chain.
Next, they cluster wallets into operators by linking those sharing exact funders or vanity address suffixes, leaning on on-chain traces of funding and sweeps to group wallets into coherent entities. This operator clustering merges payers, recipients, and hubs into funding-linked clusters. Operators are named using the x402 protocol’s off-chain resource catalog (the Coinbase CDP Bazaar) and web service domain metadata.
The key analytic innovation is constructing a directed layered value-flow graph with edges for settlements, funding transfers, and hub sweep transfers. Using this graph, each settlement is classified into one of three tiers based on whether it is fictitious (C1: self-payment or fully closed cluster with no net outflow), internal settlement (C2: within a cluster but not fully closed), or unattributed (C3: crosses cluster boundaries). This classification leverages whether the value actually leaves the operator’s cluster, the only signal resistant to automation confounds. They also examine temporal patterns: genuine demand exhibits non-zero cadence variance, while manufactured settlements often follow machine-timed regular intervals.
The authors evaluate concentration using economic concentration metrics (Gini coefficients, Nakamoto coefficients, Herfindahl-Hirschman Index). They map the on-chain settlements against the off-chain resource catalog to assess supply utilization, revenue, and host liveness via HTTP probes.
They analyze the largest operators’ capital flows, timing, and volume to characterize the manufactured economy’s star-shaped, operator-centric nature and gas subsidies.
The pipeline is fully reproducible with all code and data to be published, enabling future re-measurement over different chains, facilitators, or time windows. Concrete examples include identifying a self-payment campaign in November 2025 with 27 million transactions generated by three wallets funded from exchanges. The clustering methodology and the three-tier classification are detailed and conservative, purposely only declaring clear-cut fictitious cases as manufactured. Overall, the methodology provides a rigorous, population-scale empirical measurement that goes beyond simple settlement counts to separate operator manufacturing from genuine agent demand.
Technical innovations
- A layered value-flow graph combining settlements, funding, and sweep transfers enables clustering and economic substance analysis robust to agent automation confounds.
- A three-tier classification scheme (fictitious, internal, unattributed) grounded on value closure on the value-flow graph—only activity moving value outside funding clusters can be considered independent.
- Identification of true payers within meta-transactions despite facilitator-sponsored gas payments, using AuthorizationUsed event and on-chain signatures to resolve roles.
- Operator clustering using exact shared funders and vanity-address suffix patterns, enabling aggregation of wallet fleets under common economic control.
- Demonstration that settlement counts are a Goodhart metric highly susceptible to manufacture via gas sponsorship and capital recycling, challenging naive adoption claims.
Datasets
- Base x402 settlements — 136,708,672 settlements — Public on-chain data (USDC contract events, facilitator allowlist)
- Solana x402 settlements — 49,477,928 settlements (coarser counting) — Public on-chain data (USDC SPL transfers with registered facilitators)
- x402 off-chain resource catalog (CDP Bazaar) — 25,163 Base resources, 25,411 total resources — Coinbase CDP Bazaar public catalog
- HTTP probes of 910 Base hosts — Liveness probes on known service hosts
Baselines vs proposed
- Counting all AuthorizationUsed events without facilitator filter: 140,274,969 settlements worth $772,090,109.10 (overcount by 3.2× compared to curated set)
- Counting all USDC transfers relayed by known facilitators without event filtering: $42,348,540,765 total value (54.8× overcount compared to event-intersected set)
- Gini coefficients of payer, recipient, and value distributions all above 0.98 for Base and Solana; Nakamoto coefficient 112 payers on Base vs 3 operators after clustering
- Manufactured (fictitious + internal) settlements constitute 84.98% of Base settlements with 54.08% value in fictitious tier alone
- Unattributed settlements ≤ 15.02%, with only $187,861.35 value demonstrably reaching named external catalogs
Limitations
- Reliance on curated facilitator allowlist means unknown facilitators or improper labels could miss or misclassify settlements.
- Clustering methods cannot definitively attribute control due to custodial and shared funding services, leading to conservative assignment of internal settlements.
- Unattributed settlements (C3) remain ambiguous and may include both genuine demand and unresolved manufacture; no ground truth labels exist.
- Study period limited to a 280-day window on Base and a single snapshot on Solana; longitudinal dynamics beyond this period are not explored.
- Measurement focuses on USDC settlements; other tokens or off-chain payments may not be captured.
- Automation blurs timing and value signals, limiting classical economic substance detectors and requiring reliance on value-flow closure which may miss subtle manipulations.
Open questions / follow-ons
- How to definitively distinguish between large custodial platforms' internal ledger entries and wash trading in clustered settlements?
- Can machine learning on metadata and temporal behavior improve classification of unattributed settlements (C3) towards genuine versus manufactured?
- What is the impact of cross-chain or multi-token agentic payments outside USDC on overall agent-driven economy adoption?
- How might future protocol changes integrate stronger cryptographic identity or metadata to prevent facilitator gas sponsorship abuse?
Why it matters for bot defense
This study highlights the fundamental challenge of using raw volume metrics as a proxy for genuine adoption in decentralized, meta-transaction-based agentic payment systems. For bot-defense and CAPTCHA practitioners, it underscores the necessity of deeper provenance and economic-substance analysis when evaluating automated transaction ecosystems. Naive settlement counts can be heavily inflated by operator-manufactured activity subsidized by gas fees, analogous to how bot-generated CAPTCHA solves or traffic can skew user engagement metrics. Practitioners should therefore favor multi-dimensional, graph-based behavioral analysis that captures economic interrelations and clustering to discern true independent users from internally controlled automation. The paper's three-tier classification scheme and value-closure criterion provide a conceptual framework useful for analogous fraud and bot-detection contexts dealing with indistinguishable transactions or interactions subsidized or masked by intermediaries.
Cite
@article{arxiv2607_12575,
title={ How Agentic Is Agentic Commerce? A Population-Scale Measurement of x402 Adoption and Authenticity },
author={ Shengchen Ling and Yajin Zhou and Lei Wu and Cong Wang },
journal={arXiv preprint arXiv:2607.12575},
year={ 2026 },
url={https://arxiv.org/abs/2607.12575}
}