Billthrough is the Metric.
Auth Rate is the Distraction.

Authorization rate is easy to measure, easy to compare, and easy to report to a board. It is also the wrong primary metric for subscription merchants. This is not a contrarian position. It is what the mechanics of subscription billing actually reveal when you look closely.

Auth rate measures whether a single payment attempt succeeded. Billthrough measures whether a subscription ultimately collected payment across all retry attempts over a billing window. Those are different questions, and optimizing for the first one does not automatically improve the second. A PSP that looks weak on auth rate can still move the needle on billthrough and net revenue when it reaches subscribers on a rail your primary processor cannot access, costs nothing per failed attempt, or connects to a different issuer relationship that succeeds where the primary one failed. The auth rate number does not tell you any of that.

I watched this play out repeatedly over thirteen years working on subscription billing across global markets. The context matters here. In local payments and carrier billing, the PSP is not a secondary fallback. The subscriber chose that payment method. Carrier billing is what they have. There is no card to chain to. The merchant is evaluating the billthrough performance of that rail on its own terms, and grading it against card auth rate benchmarks that have nothing to do with how carrier billing works. Account managers would bring auth rate comparisons into merchant reviews. Merchants would penalize PSPs with lower auth rates without asking whether those rates were even comparable across rails, or whether the billthrough over the full retry window told a different story. The higher auth rate sometimes masked a worse billing outcome.

The distinction matters most in local payments, carrier billing, and emerging markets. Which is exactly where subscription growth is fastest. In these contexts, network rails are different, error codes are different, and the retry economics are entirely different from card-centric models. Auth rate comparisons across these rails are often not even measuring the same underlying thing.

The payments industry built its entire reporting infrastructure around card authorization rate because cards were the dominant payment method when that infrastructure was designed. That is no longer the world we operate in.

The reason this distinction matters now more than it did five years ago is the structural shift underway in global payments. Digital wallets have crossed a threshold from alternative payment method to dominant payment infrastructure. The optimization frameworks most subscription analytics teams use have not kept pace.

Digital wallets are not uniformly distributed. Southeast Asia, India, and Latin America have adoption rates exceeding 85% in some markets, driven by QR code infrastructure, mobile-first banking, and the leapfrogging of card rails entirely. Thailand, Vietnam, and the Philippines are each projected to hit 75% digital wallet adoption by 2026, per Juniper Research. India's UPI processes over 16 billion transactions per month.

In these markets, an e-wallet like GCash, GoPay, or a UPI-linked method is not a supplementary payment option. It is the primary payment rail. And the failure modes, error taxonomies, and retry mechanics for these rails are entirely different from the card-centric models that most subscription analytics frameworks assume.

QR code payments are the dominant transaction type within digital wallets globally. They are projected to account for over 40% of all digital wallet transactions by volume in 2026, reaching 380 billion transactions, per Juniper Research. Authorization rate comparisons drawn from Visa and Mastercard network data tell you almost nothing about your billing performance in these markets.

Brazil is worth naming specifically. Pix, the instant payment system launched by the Banco Central do Brasil in 2020, processed over 63 billion transactions in 2024, a 52% year-over-year increase, surpassing the combined volume of debit and credit card transactions by 80%. It is used by more than 174 million Brazilians and now accounts for 40% of Brazil's e-commerce payment volume. For subscription merchants operating in Brazil, Pix is not an emerging option to consider adding. It is the primary rail your subscribers expect to use.

Pix Automático, the recurring payments feature developed by the Banco Central, launched on June 16, 2025. It allows consumers to authorize a recurring payment agreement once, after which future charges are pulled automatically on the scheduled date. The failure modes are still different from cards. There is no network token. There is no chargeback mechanism in the traditional sense. Bank-side processing limits and account balance failures replace the issuer decline codes that card-based retry logic was built around. And Pix Automático's design gives subscribers the ability to cancel a mandate up to one minute before the next charge, which changes the behavioral dynamics of churn in ways that card-on-file recurring billing simply does not face. Perhaps most importantly for billthrough: Pix Automático unlocks subscription access for the roughly 60 million Brazilians without credit cards who could not previously subscribe to digital services at all. That is an expansion of addressable subscribers, not just a change of payment rail.

The most common failure mode I encountered in billing optimization programs was not a wrong strategy. It was operating without adequate signal. In carrier billing and local payment rails, 15 to 30% of transactions returned a generic error, only to reveal a specific, actionable error code when we pulled issuer log data. Generic errors are not a payments problem. They are an instrumentation problem. And you cannot optimize a billing window that you cannot read.

The following are mechanisms I observed directly in subscription billing operations across global markets. They are presented not as anecdotes but as analytical patterns. Each one represents a distinct category of optimization lever, and each one requires different data and different intervention logic.

The single PSP model is a structural constraint on billthrough optimization in card-based markets. Every PSP has a different network relationship with issuers, a different retry cost structure, and a different authorization profile for the same underlying transaction. When one PSP fails on a transaction, a second PSP, routed to automatically, may succeed for reasons that have nothing to do with the subscriber's willingness or ability to pay.

PSP chaining is a real strategy and worth pursuing where it is available. The honest caveat is that it is not universally available. In local payments and carrier billing markets, the subscriber chose a specific rail because it is the one they have access to. There is often no alternative PSP to chain to, no card on file to fall back on, and no second network relationship to exploit. In those markets, the optimization work lives entirely within the retry window on a single rail: decline taxonomy, timing intelligence, and user communication. Chaining is a card-market solution. It does not transfer cleanly to the local payment markets where a lot of subscription growth is actually happening.

The routing logic above is directional. The specific decision depends on your PSP cost structure, your subscriber's available payment instruments, and the decline type. What it illustrates is that routing is not a fallback. It is a primary optimization strategy for subscription billthrough, and it requires a taxonomy of decline types that most merchants have not built.

The calculation that makes routing economically rational is straightforward: if a second PSP charges nothing for failed attempts and a small success fee, then the marginal cost of routing a failed transaction is near zero. Recovery rates on issuer-specific declines through an alternative PSP relationship can be 10–30% depending on the market.

Everything in the previous five sections describes the same underlying situation, looked at from different angles. The subscriber signed up. The relationship exists. The content was consumed, or the service was delivered, or the seat was provisioned. The commercial transaction already happened in every meaningful sense. Your system logged a decline code and moved on.

That decline is not a lost sale. It is deferred revenue sitting in a queue, waiting for someone to go get it. Most subscription businesses are not going to get it, because they are looking at an authorization rate dashboard that tells them how many transactions succeeded on the first attempt, and they have built their optimization programs around improving that number.

The harvest framing matters because it changes what the work feels like and therefore what gets prioritized. Improving authorization rate feels like optimization work. Technical, incremental, marginal. Harvesting deferred revenue from a known population of subscribers who already said yes feels like leaving money on a table and deciding whether to pick it up.

The subscribers in your declines file are not leads. They are not prospects. They are customers who have already converted. The acquisition cost has already been paid. The content relationship already exists. The only open question is whether your billing infrastructure is capable of completing the transaction that the subscriber already agreed to.

The harvest metaphor is intentional because it implies timing. Revenue that sits in the declines file does not wait indefinitely. After 7–14 days, the subscriber's payment instrument may have changed, their engagement with your product has dropped, and their inclination to update payment details if prompted has declined substantially. The window to recover deferred revenue is real, and it closes.

The teams that treat billing recovery as a systematic operational discipline recover materially more of what is sitting in that file than teams that treat it as an edge case handled by a dunning email and a single retry. The difference is not algorithmic sophistication. It is the decision to take the work seriously.

The practical alternative to auth rate as the primary metric is a small set of measurements that together describe the actual billing outcome for your subscriber base. None of these are novel. Most are not consistently tracked.

Everything in this article assumes a human at some point in the loop. A subscriber who signed up, consumed content, maybe called their carrier, maybe responded to an SMS. The retry window assumes a person whose financial situation might change. The grace period exploit assumes a person gaming the system. The pay period calendar assumes a person getting paid.

That assumption is already breaking down. Agentic commerce, where AI systems make purchasing decisions and execute payments autonomously on behalf of users, is moving from demos to production. Visa launched its Trusted Agent Protocol in October 2025. Mastercard announced Agent Pay in April 2025. The question of how subscription billing handles a world where the "subscriber" is an AI model acting on behalf of a human is not a thought experiment anymore.

The payment failure problem looks completely different in an agentic context. Human-initiated failures have an emotional texture. Frustration, confusion, inertia. Recovery levers exist because there is a person who can respond to a prompt, update a card, or re-authorize an account. Agentic failures are deterministic. The AI either has a valid payment credential in its context or it does not. It either has authorization to complete the transaction or it does not. There is no ambiguity to exploit with a well-timed retry.

This matters for billthrough in specific ways. First, the cost tolerance for failed payment attempts is different. An AI agent executing on behalf of a user is likely running on a tight decisioning loop. A payment failure that returns a generic error might cause the agent to abandon the transaction entirely rather than wait for a retry window. The soft failures that subscription billing teams have built entire optimization programs around may not be soft anymore. The agent just moves on.

Second, fraud signals built on behavioral data stop working. No mouse movements, no session history, no device fingerprint in any conventional sense. Issuer models that learned to trust a returning subscriber based on behavioral continuity have no continuity signal to work with when the actor is an AI. This means false decline rates on agent-initiated transactions are likely to be significantly higher than on human-initiated ones, at least initially. Which means the billthrough problem for agentic subscribers may actually be worse than for human ones, not better, even though the human ambiguity is gone.

Third, the cost accounting changes. When a human subscriber churns because of a failed billing attempt, the cost is LTV. When an AI agent's transaction fails, the cost may also include the lost value of whatever autonomous task the agent was completing on the user's behalf. A purchasing agent that fails to complete a transaction may trigger a cascade of downstream failures in whatever workflow it was part of. The cost of a single declined transaction is no longer bounded by the subscription value alone.