Unstable

Published stablecoin statistics can describe the same transfers differently because their filters, address labels and purpose models answer different questions. Visa’s adjusted series covers several use cases. Two commercial payment estimates produce nearby totals while using different observation periods and category boundaries. A consumer purchase can also produce a later stablecoin settlement without becoming two consumer payments.

Following these distinctions through transaction research, provider methods and institutional disclosures shows how a contract record becomes an estimate of economic activity. Technical reconstruction establishes what executed; attribution and classification connect it to parties and purpose under stated assumptions. The analysis uses published results and disclosed methods; it does not reproduce the commercial classifiers or supply an original transaction census.

Consider an illustrative invoice for 100 units. A buyer sends 100 stablecoins through a router, which passes 100 to the merchant inside the same blockchain transaction. Summing the two transfer amounts gives 200. Counting transactions gives one. If the invoice and counterparties are known, the purchase is 100. These numbers describe different parts of one process.

Now consider a single transaction that pays two independent invoices for 100 and 80. Keeping only its largest transfer gives 100 even though the two purchases total 180. A largest-transfer filter can remove a duplicated route in the first example and omit a legitimate payment in the second. The examples establish no frequency or aggregate bias. They show why a filter needs an economic target and a validation design.

The first observation requires care too. An ERC-20 Transfer event records a contract’s report of a transfer. The standard requires events for zero-value transfers, so the presence of an event alone does not establish a positive movement or a net change in balances (Vogelsteller and Buterin 2015). A useful reconstruction identifies the token contract, successful execution, amount, surrounding calls and final state. Even a well-reconstructed balance change leaves beneficial ownership and purpose open.

Suppose the same amount passes between two addresses as an invoice payment, an internal treasury movement or collateral for a loan. Address and amount fields alone may look alike. Contract calls and traces can identify a lending operation, while business records or a validated attribution model may be needed to identify the obligation behind it. Public data can support economic inference. The inference depends on what the selected fields and additional evidence actually reveal.

Measurement validity concerns this relationship between an operational measure and its intended concept (Adcock and Collier 2001). Here, a count can be exact for its technical definition while remaining an uncertain estimate of payment activity. More precise arithmetic cannot repair an unstated purpose assumption.

The relevant distinctions concern overlapping questions about an event. A payment can also settle an obligation, and redemption can generate a transfer. Treating the terms as mutually exclusive classes would create another classification error.

The BIS-hosted glossary supplies institutional definitions of payment and settlement, while a token dataset supplies records from a particular execution environment (Committee on Payment and Settlement Systems n.d.). Moving between them requires identifying the relevant obligation and parties. The label payment stablecoin does not supply those observations for every use of the token.

Schär and colleagues (2026) study USDT, USDC and PYUSD on Ethereum during calendar 2025. Their distinction between transactions and the events nested inside them is directly relevant: complex transactions account for 31.6% of their transaction population and contain 59.96% of its stablecoin transfer events. Their inputs include an archive node, event signatures and protocol factory contracts. This paper uses their reported findings without rerunning the extraction.

The study establishes that a transfer-event series and a transaction series can describe materially different structures in the same population. It does not label every simple transfer a purchase, or make its three-token Ethereum population a global stablecoin-payment census. Complexity is an observed transaction characteristic. Purpose is a further interpretation.

Later research by Aerts and colleagues (2026, sec. 4.3) compares USDT holdings across Ethereum and Tron. It also states an instructive limit: externally owned accounts and smart-contract accounts are imperfect proxies for economic roles. The former cannot by themselves distinguish payments, remittances, exchange custody and saving; the latter can contain bridges or custodial arrangements. A cross-chain difference in address types therefore requires a bounded economic interpretation. The study’s computations are reported evidence here, rather than a reproduced dataset.

Technical reconstruction identifies which events belong to one transaction. Classification then asks what those transactions mean economically, with account type and complexity supplying evidence under the limits stated above.

Visa’s methodology makes the classification rule itself an object of research. The page retains a largest-transfer rule and activity thresholds for unlabelled addresses. It describes probabilistic address labels and separates use cases, with payments as one category among others. The October version places centralized-exchange flows and minting/burning in Unadjusted, changing their treatment from the July page. A comparison of the two series must therefore preserve the method version as well as the observation period (Visa and Allium n.d.).

The changed categories do not make Adjusted synonymous with payments. The page separately describes DeFi, trading, store-of-value and unclassified activity, and says categories can occur in adjusted or unadjusted series. Its public explanation provides assumptions to inspect. It does not release a versioned production label set, executable classifier or independently labelled purpose-validation sample.

The two commercial payment studies make different purpose bridges. McKinsey and Artemis report approximately $390 billion at an annual rate based on December 2025. Their B2B method applies a 20% allocation to selected business custody and treasury activity; their remittance method applies a 10% assumption to selected trading flows. Other categories use tagged infrastructure or exchange-advertisement data. The estimate is a modelled result, with its annualization specified in the publication’s footnote (Higginson et al. 2026).

BCG and Allium report $350–550 billion of observable bilateral goods-and-services payments during calendar 2025. They filter technical activity, infer purpose from behaviour and leave an unclear category outside the estimate. The newly retained report provides a fuller basis for inspecting those choices than the earlier summary-page capture. It still leaves its underlying database, classifier and validation sample unavailable for independent reproduction (Batra et al. 2026).

BCG describes its estimate as conservative. Excluding ambiguous observations can reduce the selected total, but that choice alone cannot certify a mathematical lower bound on true payments. Such a bound would also need justified treatment of false positives, omitted populations and classification error. A range produced by a model should be read with the model’s uncertainty and coverage, rather than as a complete interval for an otherwise observed global total.

The numerical proximity of the two publications is useful evidence of where their estimates lie. It does not establish independent replication. One annualizes a month; the other estimates a calendar year. They use different category boundaries and unavailable labels. Nor can either estimate be derived by subtracting a set of assumed non-payment categories from a universal raw denominator. The relevant denominator must first be defined.

Visa reported $14.2 trillion of payment volume for the fiscal year ended 30 September 2025. Its filing identifies purchases of goods and services on Visa products, using nominal volumes reported by financial-institution clients and reviewed by Visa. Cash activity belongs to a separate total-volume measure (Visa Inc. 2025a).

A calendar-year stablecoin transfer aggregate does not become comparable by sharing the word volume. A meaningful comparison would align period, economic function, covered population, gross/net convention, currency conversion and treatment of reversals or repeated legs. The technical possibility of compiling such an estimate is open. The published unmatched figures do not perform that work.

Visa’s own stablecoin programme shows how two valid records can coexist. Its December 2025 announcement reported a USDC settlement run rate exceeding $3.5 billion annually. Participating issuers and acquirers use the asset to settle obligations to Visa, while the consumer card experience remains unchanged. A card purchase and the later institutional settlement concern different obligations. Adding them as two consumer payments would misread both records (Visa Inc. 2025b).

An analyst comparing these records must choose whether the question concerns purchases, inter-institutional settlement, the asset used to discharge an obligation, or gross network activity. Each requires sources that identify the relevant obligation and population.

USDC supplies a related institutional example. Circle’s audited 2025 filing reported that reserve income accounted for 96.0% of its revenue. That establishes a historical revenue composition. It does not identify why a particular user transferred USDC or establish an equity valuation (Circle Internet Group 2026).

The October USDC terms distinguish non-EEA Circle Mint account holders from other holders and condition direct redemption on eligibility and account registration. The separate EEA policy concerns another legal regime. Terms governing access, timing, fees and suspension identify a holder’s position against an issuer; reserve disclosures identify assets supporting an obligation. Neither is a payment-purpose classifier (Circle Internet Financial 2026a; 2026b).

The GENIUS Act supplies another definition at another level. Its payment-stablecoin category concerns design and the issuer’s fixed-value obligation. It does not assign an economic purpose to each observed transfer. Implementing documents also require dates and actors: Treasury’s state-certification procedural interim final rule became effective on 30 September, while certifications awaited information-collection approval. That procedural milestone cannot be treated as a universal start date for every stablecoin or payment operator’s duties (U.S. Congress 2025; U.S. Department of the Treasury 2026).

A legal product category, a reserve statement and a transfer record can all concern USDC while describing different relations. The holder’s rights and the issuer’s revenue do not classify the purpose of an observed transfer.

A useful payment estimate begins with a stated target: for example, purchases of goods and services by economically distinct parties over a defined period. It then specifies the technical population and the route from those records to that target.

Contract selection, token variants, chains, transaction status, decimal normalization and repeated transfer legs establish the underlying observation set. Address attribution adds another model. Purpose classification needs an explicit rule for ambiguous cases, a validation sample independent of the training labels, and uncertainty appropriate to the population. Sensitivity to alternative thresholds and category boundaries shows which parts of the result depend on the chosen definitions.

Validation can make inference from public records credible without making every individual purpose certain. A released classifier with independent labels would supply evidence for the purpose bridge, and could support a stronger estimate than the sources inspected here. It would not erase the difference between the input record and the economic concept. Transparent retrospective designs can also contribute; prospective registration is one safeguard against changing a rule after seeing its results, rather than a requirement for every measurement study.

A causal account of market consequences is a separate project. Establishing that a volume comparison influenced an investor, acquisition or financing decision would require evidence of exposure, reliance and an appropriate counterfactual. None of the sources used here identifies that chain. The present contribution concerns the construction and interpretation of measurements.

This is a comparative documentary analysis of measurement methods and selected issuer/network disclosures. It performs no original on-chain census, reconstructs no proprietary provider database and independently reproduces neither commercial payment estimate. Published studies retain their own tokens, chains, periods and units. The USDC discussion is jurisdiction- and terms-specific.

Method changes and newly classified observations affect estimates in different ways. A changed filter alters the definition of a series; correcting an extraction alters its reported value under that definition. An independently validated classifier would strengthen purpose attribution, while revised contractual terms would change the issuer-holder relationship. Each comparison must identify the proposition and period that changed.

A payment estimate must connect its technical records to the obligation and parties it intends to count. The published methods examined here make parts of that connection visible while leaving proprietary labels and purpose validation unavailable. Releasing those inputs would let a further study test the classification and distinguish agreement between estimates from a shared measurement result.