On this page
- What routing optimization changes, and what it leaves alone
- The baseline you need before any rule change
- Levers that move approval rate versus cost
- Choosing the first route to change
- Proving a lift instead of an averaging artifact
- What the cheapest route really costs
- When MID health becomes the real ceiling
- FAQ
Key takeaways
- Payment routing optimization is a loop: fix one approval-rate denominator, change one lever, then prove the lift net of cost.
- Two providers can read identical traffic and publish different acceptance rates, so compare routes only after restating both.
- A blended average improves for free when a weak segment's volume collapses, so read each result per merchant identifier (MID), card range and country.
- Routing stops paying once a MID's monitoring ratios or volume cap bind, and no cascade depth repairs either one.
Who this helpsSubscription operators, payments leads and billing engineers who already have routing rules in production and need to prove which of them actually earned their keep.
Teams narrate most routing changes instead of measuring them. A rule ships and the dashboard average moves a tenth of a point. Nobody can say whether the lever moved it or the traffic mix did. This loop settles it: fix one denominator across gateways, change one lever, prove one result net of cost.
What does payment routing optimization actually change, and what does it leave alone?
Payment routing optimization changes which acquirer and merchant identifier (MID) sees each attempt, and in what order. It leaves the card, the customer and the bank's risk appetite alone. The honest claim is narrow. A route change moves where you send the attempt. It does not change whether the bank's answer was recoverable in the first place.
Routing reads a short list of inputs. It reads card type, the bank identification number (BIN) and its issuing country. It also reads amount, currency, merchant category code (MCC), and each route's recent success. Static routing fixes those mappings by hand. Dynamic routing recomputes them per attempt from live outcomes. Writing the mappings themselves is a separate job with its own syntax. The static versus dynamic choice has its own page. This page picks up after you write the rules. Acquirer routing logic optimization is the narrow part of the job. Cross-acquirer optimization is a measurement problem before it is a configuration problem.
Three things sit outside the lever entirely. First, Visa Core Rule 1.7.2.1 forbids a reattempt that manipulates data elements of the original request. That rule names the acquiring identifier, acquirer country, MCC and entry mode. Second, you may not reattempt Category 1 declines (04, 07, 12, 14, 15, 41, 43, 46, 57, R0, R1 and R3) on that credential at all. The third limit is the largest for a subscription book. If the issuer declines the first charge or account verification on a stored credential, the standing instruction never existed. Nothing exists to reroute.
Which baseline do you need before you touch a single routing rule?
You need one approval-rate definition that every gateway and MID reports into. Write down the numerator, the denominator and the eligibility rule. Then restate each provider's dashboard into it. Two providers can measure identical traffic and publish different acceptance rates. Compare their headline numbers and you measure reporting convention, not routing.
Stripe publishes the size of the gap. In its own worked example, the same 103,000 requests read 93.0% as a payment success rate and 94.9% as an authorization rate. The second denominator drops attempts that never reached the bank. A raw rate counts every attempt. A deduplicated rate counts each payment once. One payment with three attempts therefore reads 33.3% raw and 100% deduplicated. Denominator choice alone creates that 1.9-point spread on one month of traffic. Holding that single definition steady across providers is a monitoring job in its own right.
| Denominator | What it counts | What the number does | Where it bites |
|---|---|---|---|
| Payment success rate | Every valid payment attempt you made | 93.0% on 103,000 Stripe requests | Failures before the bank count against you |
| Network authorization rate | Only attempts that reached the bank | 94.9% on the same 103,000 requests | Blocked and failed authentication attempts drop out |
| Raw | Each retry as its own event | One payment, three attempts: 33.3% | Deep cascades read worse than the book really is |
| Deduplicated | Whether the order was approved in the end | The same payment: 100% | The open month reads low until retries run |
| Net versus gross | Net counts orders, gross counts attempts | 97.9% net beside 91.44% gross | A 6.5-point gap can be retry volume alone |
| Declines over total attempts | The ratio the networks watch per window | Retries inflate both sides | Around 15% can put the account under review |
Each row is one definition. Pick one, write it down, and restate every provider's dashboard into it.
Which routing levers move approval rate, and which only move cost?
Two levers move approval rate: routing a segment to the acquirer that actually approves it, and cascading behind a soft decline. Picking the cheaper debit network or the cheaper acquiring domicile moves cost instead. Sort the levers by what they move before you pick one, because least-cost routing is not approval-neutral. Those two outcomes pull against each other, so score them on one line.
| Lever | What it moves | Published effect | What that number is not |
|---|---|---|---|
| Routing by BIN and issuing country | Approval rate | 2 to 5 points across 12 tested setups | A forecast for your card mix |
| Cascading to a second acquirer | Approval rate | Roughly 2 to 5 points in isolation | A reset of the 20-attempt budget |
| Local acquiring in the customer's market | Both | 8.4 points approval, 1.27% versus 3.07% cost | Available wherever you want a local MID |
| Network tokens | Approval rate, not routing | Visa 4.6% and Mastercard 2.1% on card-not-present | Comparable across schemes or portable across providers |
| Least-cost debit routing | Cost only | Declines 0.5% to 2% higher in some setups | Approval-neutral |
| Retry timing by decline code | Approval rate, by schedule | 11.6% average lifetime-value lift on recovered soft declines | A routing change at all |
| Account updater on stored cards | Approval rate | 97 basis points in the first full month | A substitute for fixing the credential chain |
| Percentage rotation between providers | Neither | No consistent effect once BIN and brand are controlled | A lift; the variance is often misread as one |
Every effect here is one vendor's measurement on one book, so treat each as a hurdle to beat rather than a forecast.
Order matters. The most-quoted lever in that table is not a route at all. Visa reports a 4.6% authorization lift on tokenized card-not-present traffic versus raw card numbers. That lift is the hurdle an incremental routing change has to beat. Consistent transaction-ID hygiene on recurring charges sits in the same category. Do the data work first, then measure routing against the new baseline. Otherwise you credit the token rollout to your routing rule.
How do you choose the first route to change when every acquirer looks about the same?
You stop reading the blended number and start reading segments. Acquirers look identical in aggregate because the aggregate is a mixture. An overall 92% can hide 85% on one acquirer for cards issued in the European Union. A top market at 97% sitting beside a market 15 points lower is routine. That spread is where the first rule change belongs.
- Enrich every attempt with BIN attributes: issuing country, card brand, debit or credit, and issuer name.
- Cut approval rate by issuer country, card type, amount band and MID, using your one fixed denominator.
- Record each segment's share of total attempts next to its rate, so a mix shift is visible later.
- Drop segments too small to decide on; a BIN with 40 transactions cannot justify a route change.
- Remove the declines routing cannot reach, starting with the roughly 1 in 5 hard declines caused by stale card data.
- Rank what is left by payments lost per month, and change the route for the top segment only.
How do you prove a routing change worked instead of reading an averaging artifact?
You run the old route as a concurrent control arm and read the result per segment. Split traffic live between the current route and the new one. Both arms then face the same day, the same card mix and the same issuer moods. A before-and-after read on the same traffic compares two different environments. That comparison is the common merchant mistake.
Then size the test before you ship the lever. At a 70% baseline, detecting a move to 80% takes about 330 payments per variant. Detecting 70% to 71% takes about 33,000 payments per variant. Routing lifts are the second kind. Start at a 90/10 split to cap the downside, hold it 7 to 14 days, and fix the sample size first. A test read once at its planned end carried a 5.2% false-alarm rate. Read twenty times during the run, the same test carried 24.3%.
Payment routing optimization software can run the split and hold the sample size for you. It cannot pick your denominator. A payment routing optimization platform also adds a cost line that the lift has to clear. One analysis prices that trap. A 0.3% authorization gain worth about $45,000 a year sits against $306,000 of implementation, platform and processing cost. The result is a $261,000 annual loss.
One averaging artifact can still make an honest test read wrong. The artifact has a name, Simpson's paradox. Its cause is a shift in who showed up, not a change in how anything performed. On a MID whose volume collapsed, the weak segments stop contributing attempts. The blended approval rate then improves for free. The detection test is mechanical. Recompute the new period's overall rate on the prior period's segment mix. If that hypothetical number barely moves, you are looking at a mix shift.
The same trap sits in the search metric that operators read next to revenue. Google records a position only where the page registered an impression. A page whose weak-query impressions collapse then shows a better average position with no ranking change. Both artifacts share one fix: read the rate beside the count, per segment, every time.
Work the arithmetic before you promise anyone a result. Say you run Northwind Audio, a subscription book taking 60,000 card attempts a month. You expect a one-point approval gain from routing cards issued in Canada to a local acquirer. A 50/50 split needs about 33,000 attempts per arm. A clean read therefore takes more than a month of your whole book, let alone one country. The honest options are to test a bigger lever, pool several months, or accept that this change ships unproven.
What does the cheapest route cost you in interchange, cross-border and scheme fees?
The cheap route costs you the difference between interchange programs it no longer qualifies for. It also adds a per-attempt network fee on every extra authorization. On United States consumer credit card-not-present traffic, the gap between the cheapest and the most expensive interchange outcome is about 182 basis points. Qualification decides which end you pay. Qualification is mostly data, not route choice.
Measure the prize before you chase it. Visa caps the card-not-present interchange discount from authorization-side work at 15 basis points. The split is 5 for a token, 10 for authentication data, 15 for both. The scheme then charges for the same capabilities. A 0.015% digital commerce services fee nets the token incentive to about 0.035%. A 0.05% authentication data fee halves the 0.10% incentive to 5 basis points.
Cross-border is where the numbers stop being small. Adyen reports domestic transactions at an average 1.27% against 3.07% for cross-border routed through a single home-market acquirer. On a foreign-issued card acquired in the United States the stack is explicit. Visa charges a 1.00% international service fee, which rises to 1.40% when settlement is not in dollars. A 0.45% international acquirer fee sits on top of that. Mastercard adds a 0.85% global acquirer support fee alongside its own 0.60%. You cannot simply point volume at a cheaper local MID either. Visa assigns the outlet country from your principal place of business. The trade-offs of a real local setup are their own decision.
- Each attempt in a cascade carries a network fee: Visa charges $0.0195 per United States credit authorization and $0.0395 on a foreign-issued one.
- Mastercard's network access fee is $0.0195 for United States cardholders and $0.0295 for others, charged per authorization record.
- Visa's excess reattempt fee starts on the 21st attempt in 30 days, at $0.10 domestic and $0.25 cross-border.
- Mastercard charges $0.50 per declined authorization once 10 declines in 24 hours or 35 in 30 days are passed.
- One published decline advice fee moved from $0.05 to $0.78 in February 2026, so a repeated decline is no longer rounding error.
- Recovery per extra attempt falls fast: 15% to 25% on the first retry, 8% to 12% on the second, 3% to 6% on the third, under 2% after that.
- Score the whole change on cost per approved payment: one worked least-cost example turns $15 saved per 100 transactions into a $135 net loss once two extra declines are counted.
When does routing optimization stop paying, and MID health become the real ceiling?
Routing stops paying the moment a MID's monitoring ratio or its underwritten volume cap becomes the binding constraint. At that point the acquirer, not the issuer, decides your volume. Some acquirers require a merchant to stay under a 0.40% ratio, well below the published scheme line. No cascade depth argues with that number. There is no standard monthly volume cap either. Read the second limit off your own underwriting profile.
Your own volume builds the ratio that binds. The Visa Acquirer Monitoring Program (VAMP) ratio counts fraud reports (TC40) plus disputes (TC15) over settled card-not-present transactions (TC05), by count. The merchant excessive line has been 1.50% since 1 April 2026, down from 2.20%. A floor of 1,500 counted events a month applies. Identification costs $8 per counted event. Acquirer tiers sit far lower, at 0.50% above standard and 0.70% excessive.
Measurement and the ceiling meet on the denominator. Mastercard divides this month's chargebacks by last month's sales on that MID. Routing volume off a MID therefore raises its ratio with dispute behavior unchanged. The same month can read 1.75% on the prior month's volume and 1.17% on same-month volume. Spreading volume across MIDs to dilute a ratio is not a lever at all. Visa treats altering merchant data to circumvent monitoring as a data-quality violation. That violation carries a $25,000 per-merchant, per-month assessment.
The schemes' own numbers show the ceiling too. Visa reports that acquirers which remediated their ratio saw approval rates and volume rise, worth about $1.5 billion in incremental volume. Acquirers that did not remediate stayed close to flat. Aggressive retry patterns can trip a separate card-testing ratio of 20% with a 300,000-transaction minimum. Merchant decline-rate monitoring near 15% puts the account itself under review. Reading MID health next to approval rate tells you when the lever is finished.
- Write one approval-rate definition: the numerator, the denominator, and which attempts are eligible.
- Restate every gateway and MID dashboard into that definition before you compare any of them.
- Cut the baseline by issuer country, card type, amount band and MID, and record each segment's share of attempts.
- Pick exactly one lever, and fix the sample size for the lift you expect before launch.
- Run the current route as a concurrent control arm, starting at 90/10, for 7 to 14 days minimum.
- Read the result per segment, then recompute the blend on the prior period's mix before claiming a lift.
- Price the change on cost per approved payment, including per-attempt fees and interchange qualification.
- Check the MID's monitoring ratios and volume headroom before routing more volume onto it.
Frequently Asked Questions
How do I measure whether a payment routing change actually improved approval rates?
Run the old route as a concurrent control arm and compare segments rather than months. Split traffic live, start at 90/10 to cap the downside, hold it 7 to 14 days, and fix the sample size first. At a 70% baseline, proving a one-point lift takes about 33,000 payments per variant.
What is the difference between static and dynamic payment routing?
Static routing is a fixed mapping you write by hand, such as sending debit cards issued in Canada to one acquirer. Dynamic routing recomputes the choice per attempt from live outcomes. Static is auditable and stable. Dynamic reacts faster, but it makes attribution harder, because the route moves while you are measuring it.
Does least-cost routing lower total payment costs or just processing fees?
It lowers network fees on the transactions it moves, and it can raise declines by 0.5% to 2% in some setups, because some issuers run stricter fraud rules on certain debit networks. One worked example turns $15 saved per 100 transactions into a $135 net loss once two extra declines are counted.
How many processors do you need for intelligent payment routing to be worth it?
Two active providers is the working minimum, and roughly 10,000 transactions a month is the floor where processor-level patterns become readable. Segment decisions need hundreds of transactions per cell, so splitting 200 monthly payments across two gateways optimizes noise rather than approval rate.
How do you optimize payment routing for efficiency?
Fix one approval-rate definition across every gateway and MID, then change one lever at a time. Cut the baseline by issuer country, card type and amount band to find the real variance. Prove each change against a concurrent control arm, read it per segment, and price it net of fees.