Approval rate by segment: the cuts that lead to action

A single approval rate tells you that something is wrong and very little about what. If the monthly number falls by a point, the cause might be one issuer in one country, one provider, one card type or a change in who is buying. Breaking the rate down finds the cause. Break it down too far, though, and you get a table of 400 rows in which half the numbers are noise.

This article sets out which segments to cut, what each one can lead you to do, how far to trust a small one, and how to turn a gap into a decision. It assumes you already have the formula, covered in how to calculate approval rate, and the data, covered in building a decline dashboard.

Why does one approval rate hide the problem?

Because it averages groups of payments that behave very differently. Stripe’s analytics documentation says that domestic success rates are on average higher than cross-border, that credit success rates are on average higher than debit, which in turn are higher than prepaid, and that digital wallets typically beat card payments because they are tokenised and device-authenticated. Its guide adds that authorisation rates can be 10% lower online than in person.

So a change in mix can move your headline rate when nothing has got better or worse. A campaign that brings in more foreign prepaid cards will lower the rate with every provider working as before. The mix trap in our approval rate guide shows how this can even make the weaker provider look stronger.

Segmenting answers two questions: which group moved, and whether the move was performance or mix.

Which segments should you cut first?

Use a fixed order, from the coarsest cut to the finest, and only go deeper where the volume supports it. Keep a segment if a different result would change what you do.

Segment	What it can reveal	What you can do about it
Payment method	Cards against wallets against local methods	Change what the checkout shows and in what order
Issuing country	Cross-border penalty, one market slipping	Add a local provider, adjust routing, review 3-D Secure
Card type	Debit, credit and prepaid differences, funds declines	Set expectations, suggest another method, route prepaid separately
BIN or issuer	One bank refusing you	Route by BIN, raise it with the acquirer
Amount band	Limits, exemption edges, high-value caution	Adjust rules at the band’s edge, offer another method at high amounts
Device and browser	Broken flows, failed device data, wallet availability	Fix the checkout, test the flow on that device
Provider	One route weaker on the same traffic	Change a rule, run a split, talk to the provider
New or returning customer	Thin history on first purchases	Review fraud rules, send better data

Issuing country

Use the country of the issuer, not the customer’s location. Stripe’s card country pivot means the country of the card issuer, and says this may differ from where the customer physically is. A card from one country used by someone abroad is still a foreign card to the acquirer. Compare each country’s rate with its own history before you compare it with other countries, because local card mix differs. Our article on local and cross-border acquiring covers the usual fix when a large country sits persistently low.

Card type

Split debit, credit and prepaid inside each major country. Stripe says lower debit and prepaid rates usually reflect a higher chance of insufficient funds. That means a drop in this segment may be nothing to do with your setup. Check the decline group before you act.

BIN and issuer

The BIN is the first six to eight digits of the card number, as Checkout.com’s documentation describes it, and Checkout.com lets you filter acceptance by issuing country, issuing bank and issuing BIN. Start with the issuer and rank by declined volume rather than by rate. A bank at 40% on 30 payments matters less than a bank at 70% on 30,000. Group the long tail into “other”, as Stripe’s reports do for low-volume data points. Our guide to BIN routing covers what to do with a BIN that stands out.

Payment method

Keep cards and each local method family apart. Stripe’s reports for local methods pivot on payment method, buyer country and currency, because the concept of an issuer does not carry over. It also says decline reasons vary by method. A voucher that expires unpaid is not a decline, so define success per family.

Amount band

Equal-width bands hide what matters. Set the edges where rules change instead. In Europe, Stripe’s guide lists the low-value exemption below EUR 30 and transaction risk analysis thresholds at EUR 100, 250 and 500, so a band ending at EUR 50 can mix payments that are treated differently. Add your own edges: provider limits, the amounts where instalments appear, your largest order size. Stripe’s guide notes that a USD 10 hold is more likely to succeed than a larger one, and a falling rate with amount is normal to a degree. Look for a sudden step at one amount, which points to a limit or a rule.

Device and browser

Device explains a different kind of failure. Stripe’s input method pivot shows wallets separately, and wallets depend on the device. Adyen warns that the redirect to authentication can lower conversion through technical errors or shoppers dropping out. We found no reliable published figures for approval by device, so treat this as a diagnostic: a gap between desktop and one mobile browser usually means something in the flow is broken, and the fix sits with your team.

Provider

Compare providers only within one segment and one definition. Checkout.com’s documentation notes that some PSPs call its acceptance rate their approval rate, and it advises using acceptance rate to benchmark providers. Use one internal formula for all of them, as our approval rate guide recommends.

How small is too small?

A rate from a small sample is a rough estimate. The table shows the range you should expect around a measured 80%, using the Wilson method described in NIST’s Engineering Statistics Handbook at 95% confidence.

Example figures, for illustration only.

Payments in the segment	Measured rate	95% range	Roughly either way
100	80.0%	71.1% to 86.7%	8 points
200	80.0%	73.9% to 85.0%	6 points
500	80.0%	76.3% to 83.3%	3.5 points
1,000	80.0%	77.4% to 82.4%	2.5 points
5,000	80.0%	78.9% to 81.1%	1 point
10,000	80.0%	79.2% to 80.8%	under 1 point

Volume has to grow about fourfold to halve the range. NIST notes that limits based on the normal approximation may not be accurate enough when the sample or the number of failures is very small, and it gives an exact method for those cases. For a dashboard, the Wilson range is enough, and NIST says its performance does not depend strongly on the sample size or the rate.

Two things make real ranges wider than the table. The method assumes independent payments, and one customer trying five times is not five independent events, so count intents. And it says nothing about changes over time. A segment measured over a month containing a sale and an outage is not one steady population.

Why do many segments produce false alarms?

Every extra comparison is another chance to be fooled. NIST’s handbook makes the point for multiple comparisons: run many pairwise tests and the overall significance level is no longer what it was for a single test.

Take 20 segments in which nothing has really changed, and test each at the 5% level, the threshold Stripe uses for its own A/B tests. The chance that at least one looks significant is about 64%, if the tests are independent. A monthly review of 20 countries will produce a false alarm most months.

Four habits contain this:

1. Choose the segments in advance. Write down the cuts you review, and do not hunt through new ones until something looks interesting.
2. Set a minimum volume. Grey out or hide rows below it, so nobody acts on them.
3. Ask for a repeat. A real gap survives into the next period. Noise tends to fade.
4. Check that the size matters. A significant gap of half a point on small volume may not be worth a routing rule.

How do you turn a segment gap into a decision?

Work through the same steps each time.

1. Confirm the volume. Is the segment above your minimum, and do the ranges still separate?
2. Check the mix. Did the share of prepaid, foreign or new customers change inside the segment?
3. Look at the decline groups. Funds declines mean something different from generic or technical ones. Our decline dashboard article sets out the groups.
4. Check the timeline. Line the change up against releases, rule changes and provider incidents.
5. Size the prize. Multiply the gap by the segment’s payments and average order. A large gap on a tiny segment waits.
6. Test before you commit. Send a share of the same segment to the alternative. A random split removes the differences a routing rule creates. Stripe’s A/B test for its own features, for example, divides eligible volume into two equal groups, runs for 30 days and waits a further 7 days for retries to finish before it reports. Our guide to traffic splitting covers the set-up.
7. Write down the result. Note the rule you changed and the date, so the next review can see the effect.

Worked example: same gap, different verdict

Example figures, for illustration only.

A shop sells into one European market through two providers. A review shows that Provider A approves 80% of debit cards on one bank’s BIN and Provider B approves 74%. The analyst proposes moving the BIN to Provider A.

Cut	Provider A	Provider B	Do the ranges overlap?	Verdict
One BIN, one month	80 of 100 (71.1% to 86.7%)	74 of 100 (64.6% to 81.6%)	Yes	No conclusion
All of the bank’s debit BINs, one quarter	800 of 1,000 (77.4% to 82.4%)	740 of 1,000 (71.2% to 76.6%)	No	Worth testing

The gap is the same six points. On 100 payments it could easily be noise. On 1,000 payments it is unlikely to be, although the two providers may still have seen different customers. So the team splits that bank’s debit cards evenly between both providers for a month. The gap holds at about five points, the decline groups show Provider B returning more generic declines, and the team writes a BIN rule. It reviews the BIN list after a quarter, because issuers and providers change.

What are the common mistakes?

• Ranking by rate, not by lost volume. The worst rate is often on the smallest row.
• Reading rates without counts. Every rate needs its denominator beside it.
• Cutting by everything at once. Country by card type by BIN by provider by device leaves almost nothing in each cell.
• Mixing definitions. Attempts in one row and intents in another cannot be compared.
• Treating a routing outcome as a provider’s quality. A provider that receives your hardest traffic will look worse.
• Acting on one good week. Wait for a repeat, or test it.

How does Omno help with segment analysis?

Omno’s transaction trail records every attempt, cascade step, provider response and webhook, and its reports and widget analytics sit on that data, so a rate can be traced back to the payments behind it. When a segment needs action, routing rules can use card BIN, country, currency, amount, browser information or customer segment, and traffic splitting by proportion lets you compare providers on the same segment. These are rules you set, not predictions. The reporting and analytics page shows what is included.