Benchmarking approval rates without the spin

Approval rate is the most quoted number in payments and one of the least comparable. Two providers can describe the same traffic with numbers three points apart, both accurately, because the metric has no agreed denominator.

Where the number moves without anything improving

• Technical declines. Excluding format errors and timeouts raises the rate. Including them lowers it. Both are defensible; only one is usually disclosed.
• Retries. Counting a recovered retry as a success while counting the original attempt in the denominator differs from counting the pair as one — several points of spread on retry-heavy portfolios.
• Pre-auth filtering. Blocking risky transactions before they reach the network raises approval rate and lowers revenue. It looks like an improvement in exactly the metric being optimised.
• Currency and mix. Domestic traffic approves higher than cross-border. A quarter with more domestic volume shows a better rate from mix alone.

Comparability requires cohorts

A single portfolio-level number cannot be compared across merchants, because the mix dominates. A subscription business billing stored credentials in one country and a marketplace taking first-time cross-border cards are not measurable against each other, and the difference says nothing about either's optimisation.

Useful benchmarking compares cohorts: same issuer country, same card product, same MCC band, same token type, same amount band. Within a cohort the comparison means something. Across the aggregate it does not.

Token type deserves its own axis

Raw PAN, network token, and wallet-provisioned device tokens have materially different baseline approval rates from the same issuers. A portfolio shifting toward wallet volume will show a rising approval rate with no optimisation at all.

Reporting the blended figure without the token mix beside it makes an accounting change look like a result. Splitting by token bucket is the minimum honest presentation.

The only claim that survives scrutiny

Benchmarks describe where you sit. They do not establish that a change caused an improvement — for that the comparison has to be against a holdout on your own traffic, in the same period, with the same mix.

It is a smaller and less flattering number than a year-over-year comparison, and it is the one that holds up when someone checks. A vendor unwilling to run their optimisation against a holdout is asking to be measured by a method that cannot fail.