Benchmark report
v1.0, November 2026
The State of Failed Payment Recovery in 2026: Benchmarks From 407 Subscription Companies
This benchmark is drawn from 407 subscription companies that replaced their billing platform’s own retry model with FlyCode, on Stripe, Shopify, PayPal, Apple Pay and Chargebee, during the six months from 7 March to 6 September 2026. Across the sample, FlyCode recovered 62.1% of failed payment value, against 52.0% for the platforms’ own models on the same accounts before the switch. That is a median lift of 17.6% in recovery rate, and $22.6 million that would not have been recovered at the earlier rate. 37.9% of failed payment value was still unrecovered after 30 days, and this report shows that too.
Most published recovery benchmarks report a recovery rate per dunning attempt. This one measures something different: how much more a dedicated recovery model recovers than the platform’s own model it replaces, whether that is Stripe Billing Smart Retries, a Shopify subscription app’s schedule, PayPal’s retries or Chargebee’s dunning rules. That is the number a finance team can act on.
All figures come from the 407-company dataset for 7 March to 6 September 2026 unless stated otherwise. Industry reference points and earlier FlyCode studies are named where used.
Who wrote this, and how to read it
We sell payment recovery, and our fee is a share of the uplift above each customer’s baseline. We have an interest in these numbers being large, and you should read them with that in mind. Here is what we did about it:
We report the full sample, including the companies where we added little. 32 of the 407 companies measured under 5% lift, and 9 measured no lift distinguishable from zero. They are in every median below.
We measure against the platform, not against zero. FlyCode’s gross recovery rate in this sample is 62.1%. Most of that is recovery the platform’s own model was already delivering, so the figure that counts is the 17.6% improvement over it. It is also the only part we are paid on: our fee applies to the uplift above the baseline and nothing else.
We name the weak point of our own method. FlyCode replaces the platform’s retry model. The two never run on the same payment, so there is no side-by-side control group. The baseline is the same account before the switch, and a before-and-after comparison can absorb other changes. The methodology section says what we did to limit that and what we could not rule out.
We say what did not work. With FlyCode running recovery, 37.9% of failed payment value, $84.9M, was still unrecovered after 30 days. The section on failure causes shows where.
The named customer results near the end are public case studies you can check against the sample medians.
Why this report exists
The recovery benchmarks that circulate in subscription finance answer the wrong question. They report a gross recovery rate, or a rate per dunning attempt, and leave out the thing a CFO actually needs to know: how much of that recovery would have happened anyway.
Stripe Smart Retries, Shopify’s native retry schedule, PayPal’s subscription retries and Chargebee’s dunning rules all recover a meaningful share of failed payments on their own. A vendor quoting a 70% recovery rate on a Stripe Billing account is not telling you how much of that Stripe’s own model would have recovered anyway. The only number that reflects what a recovery platform earns is the difference from that baseline, measured on the same account with the same counting rules.
We also think the industry has been measuring the wrong channel. Dunning emails are visible, so they get benchmarked. Retry timing is invisible, so it does not. In this sample, 71% of recovered value came from when and how a card was retried, with no customer contact at all. A benchmark built around email reply rates describes the smaller lever.
This report does five things differently:
It measures incremental lift, not gross recovery. Every lift figure compares FlyCode’s recovery rate with the rate the platform’s own model achieved on the same account before FlyCode replaced it. Recovery the platform was already delivering is not counted as lift.
It covers 407 companies across five billing stacks and payment rails. Stripe, Shopify subscriptions (Recharge, Skio), PayPal, Apple Pay and Chargebee. Most published benchmarks come from one tool on one processor, often with a sample of around a hundred accounts. This sample is 3.4 times larger than the 119-company sample behind the most-cited one, and spans SaaS, DTC, media and mobile app subscriptions.
It runs for six months, not one. A single month cannot show whether a result holds, whether recovered customers stay, or how long a per-merchant model takes to learn an account. Six months can, and each of those gets its own section.
It looks at things only a retry engine can see. Which days recoveries land on, how many attempts each recovery costs, and what happens to a subscriber after the payment is recovered. An email tool has no view of any of these.
It breaks results down by failure cause, platform, business model and trial status. The aggregate is less useful than the structure underneath it. Insufficient funds does not behave like a generic decline. A failed trial conversion does not behave like a failed renewal. Operators need the breakdown to know where their own gap sits.
The sample: 407 companies
The sample is every FlyCode customer with at least one recovery attempt between 7 March and 6 September 2026, across all five supported billing stacks and payment rails. It is not a curated case-study group. Companies seeing modest results are in the sample alongside the ones with outsized outcomes, and the 14 companies that stopped using FlyCode during the window are included for the months they were live.
Dimension | Value |
|---|---|
Companies in sample | 407 |
Measurement window | 7 March to 6 September 2026 (six months) |
Billing stacks and payment rails (each company counted once, by primary rail) | Stripe 281, Shopify subscriptions 61, PayPal 28, Apple Pay 16, Chargebee 21 |
Business model | SaaS 203, DTC subscriptions 98, Media and publishing 41, Mobile apps 65 |
All failed payments in the window | 3,614,500 payments, $224.1M |
Recovered within 30 days with FlyCode | 2,244,600 payments, $139.2M (62.1%) |
What the platforms’ own models would have recovered at their pre-switch rate | $116.5M (52.0%) |
Incremental recovery | $22.6M |
Still unrecovered after 30 days | 1,369,900 payments, $84.9M (37.9%) |
What counts as a failed payment. A subscription invoice or order charge that the billing platform marked as failed on its first attempt. Counting is at the invoice level, not the attempt level: an invoice that fails four times and then succeeds is one failed payment and one recovery, and duplicate invoices are removed before any rate is computed. The figures come from billing platform event logs (invoice and charge events), not from dashboard summaries.
What is excluded, and what that does to the numbers. Companies with FlyCode enabled but no failed-payment events in the window; manually invoiced enterprise accounts where card failure is not the payment path; in-app purchases billed by the app stores, which no third party can retry; and each company’s first 30 days live, while the per-merchant model is still training. That last exclusion flatters the result, so here is the figure without it: including the first 30 days, the sample median lift is 16.1% instead of 17.6%.
The headline numbers
For the 407 companies in the sample over the six months from 7 March to 6 September 2026:
Metric | Value |
|---|---|
Recovery rate under the platforms’ own models, same accounts, 12 months before the switch | 52.0% |
Recovery rate with FlyCode (share of failed payment value recovered within 30 days) | 62.1% |
Median lift in recovery rate | 17.6% |
Middle half of companies (25th to 75th percentile) | 10.3% to 25.6% |
Mean lift | 20.1% |
Value-weighted lift (62.1% against 52.0%) | 19.4% |
Median lift including each company’s first 30 days live | 16.1% |
Incremental recovered revenue, six-month window | $22.6M |
Total recovered with FlyCode | $139.2M across 2,244,600 payments |
Recovered without any customer contact (retry only) | 71% of recovered value |
Recovered after a customer-facing step (payment update link, backup card, email or SMS) | 29% of recovered value |
Median incremental recovered revenue as a share of MRR, per month | 0.9% |
Median time from first failure to recovery | 6.1 days |
Median ROI over the window (incremental recovered revenue ÷ FlyCode fee) | 6.8× |
% of companies with 5× ROI or better over the window | 74% |
% of companies with 10× ROI or better over the window | 19% |
% of companies with Stripe as the primary billing stack | 69% |
Reference baseline (not scored). The recovery rate each billing platform’s own model achieved on these accounts in the 12 months before FlyCode replaced it, as a share of failed payment value. Smart Retries is a Stripe Billing feature, so it is listed by the accounts that had it, not as a Stripe-wide default.
Platform’s own model, before the switch | Accounts | Recovery rate before FlyCode |
|---|---|---|
Stripe Billing Smart Retries (enabled) | 228 | 54% |
Stripe fixed retry schedule or no automatic retries | 53 | 41% |
Shopify subscriptions native retry schedule (Recharge, Skio) | 61 | 49% |
PayPal subscriptions built-in retry | 28 | 46% |
Apple Pay (no native retry; processor schedule on the token) | 16 | 43% |
Chargebee dunning rules (merchant-configured) | 21 | 50% |
All platforms, value-weighted | 407 | 52% |
Three things to calibrate before reading further.
On the baseline. FlyCode replaces the platform’s retry model on every account in this sample. The two never work the same payment. The baseline is therefore the account’s own recovery rate under the platform’s model in the 12 months before the switch: Stripe Billing Smart Retries where the merchant had it available and enabled (228 of the 281 Stripe accounts), Stripe’s fixed retry schedule or no automatic retries where not, and the native model on the other platforms. Those rates are in the reference table above, outside the scored figures. Judge the result by the lift, not by the 62.1% gross rate: most of that 62.1% is recovery the platform was already delivering.
On ROI. FlyCode’s fee is charged only on the uplift above the baseline: the revenue recovered beyond what the platform’s own model was recovering in the 12 months before the switch. If recovery does not rise above the baseline, there is no fee. Two things follow. ROI is computed on that same incremental revenue, the $22.6M and not the $139.2M. And the baseline in this report is not only a reporting choice. It is the figure each customer is billed against, so every company in the sample has a financial reason to check it. The median ROI is 6.8×. Because the fee moves with the result, ROI is still a weaker proof than lift, and lift is the number to plan with.
On mean vs median. The mean lift (20.1%) is pulled upward by a tail of smaller companies with a very high share of insufficient funds declines. The value-weighted figure (19.4%) sits below it because the largest accounts are closer to the median. Model your own outcome against the median, not the mean.
The distribution of outcomes
Aggregates hide more than they show. Here is how the 407 companies distributed across lift bands in the window:
Lift in recovery rate | Companies | Share |
|---|---|---|
Under 5% | 32 | 8% |
5 to 10% | 65 | 16% |
10 to 20% (median 17.6%) | 146 | 36% |
20 to 30% | 111 | 27% |
30%+ | 53 | 13% |
Three patterns in the distribution:
The bottom band is new accounts and clean accounts. The 32 companies under 5% lift fall into two groups: accounts in their second and third month live, where the per-merchant model is still learning the account’s issuers, and accounts with an unusually low share of insufficient funds declines, where the platform’s default retries already capture most of what is recoverable. Nine of them measured no lift distinguishable from zero. For those nine, the platform default was already enough, and under uplift-only pricing they paid nothing.
The bulk of the sample clusters in the 10 to 30% range, 257 of the 407 companies. This is the range to benchmark against. It reflects what per-merchant retry timing adds on a typical mix of soft and hard declines once the model has learned the account.
The top band is driven by failure mix, not by configuration tricks. Companies above 30% lift share one trait: a high share of insufficient funds declines, often in DTC and consumer subscriptions with monthly billing dates clustered around the first of the month. These are the failures where the processor’s fixed schedule is most wrong and a model trained on issuer behaviour is most right.
If you are benchmarking your own recovery layer against this sample, compare against the median lift of 17.6%, not the top band.
Six months, month by month
A single-month benchmark cannot tell a good month from a durable result. This one covers six, so the question can be asked directly: does the lift hold?
Month of the window | Median lift |
|---|---|
7 March to 6 April | 16.9% |
7 April to 6 May | 17.2% |
7 May to 6 June | 17.8% |
7 June to 6 July | 17.4% |
7 July to 6 August | 18.1% |
7 August to 6 September | 18.0% |
Six-month median | 17.6% |
It does. The monthly median moved within a 1.2 point range and never fell below 16.9%. The slight upward drift is the per-merchant models maturing: 61 companies went live during the window, and each one’s lift rises through its first 90 days. No single month drives the six-month figure, and removing the best month changes the median by less than 0.2 points.
Where failed payments come from: insufficient funds vs everything else
The single most important structural finding in the sample is how concentrated failure causes are. Insufficient funds accounted for 58% of failed subscription payments across the 407 companies, consistent with the 50 to 70% range we see across our broader customer base.
Failure causes across the 407-company sample: share of failures, recovery rate before and with FlyCode, and lift |
|---|
Failure cause | Share of failed payments | Share of failed value | Recovery rate before (platform’s model) | Recovery rate with FlyCode | Lift |
|---|---|---|---|---|---|
Insufficient funds (code 51) | 58% | 54% | 60% | 74.5% | 24.1% |
Generic decline / do not honor (code 05) | 17% | 18% | 50% | 56.5% | 13.0% |
Issuer and risk declines (lost, stolen, fraud flags) | 7% | 9% | 12% | 12.8% | 6.8% |
Card data errors (invalid number, CVC, address) | 6% | 7% | 38% | 42.3% | 11.2% |
Processor and network errors | 5% | 5% | 72% | 85.0% | 18.1% |
Other | 7% | 7% | 45% | 49.2% | 9.4% |
This matters because the failure causes do not respond to the same lever.
Insufficient funds is a timing problem. The customer has the money on a schedule the issuer knows and a fixed retry calendar ignores. Payday patterns, benefit deposit dates and salary cycles differ by issuer, by country and by customer. A model trained per merchant on Visa and Mastercard network data learns those cycles. In our 500+ merchant, $1B volume study we measured 17.6 to 26.2% additional uplift over Stripe baselines on this code alone, and the 407-company sample shows 24.1%. No email fixes an empty account. A well-timed retry does.
Generic declines are a routing and context problem. Do not honor hides several root causes behind one code. Some are soft and recover on retry with a different descriptor, amount or time of day. Some are hard and need a new payment method. Separating the two is the job, and the lift on this code depends on how well a model can tell them apart.
Card data errors and hard issuer declines are customer-action problems. These are where a payment update link, a backup card on file or a customer message earns its place. They are also a minority of failures by value in this sample (16% combined), which is why a recovery strategy built mainly around messaging leaves most of the money on the table.
What did not recover. The recovery columns have a mirror image. With FlyCode running recovery, 25.5% of insufficient funds value was still unpaid after 30 days, and for issuer and risk declines the figure was 87.2%. Across all causes, $84.9M of the $224.1M in failed payments stayed unrecovered. Some of that is unrecoverable by design: a card reported lost or stolen should not be retried at all. Some of it is subscribers who had already decided to leave and let the card fail. A better model narrows the gap. It does not close it, and a benchmark that implies otherwise is measuring something else.
The Smart Retries baseline: what Stripe Billing recovers and where it stops
69% of the 407 companies (281) run their primary billing on Stripe. Smart Retries is a Stripe Billing feature, not a Stripe-wide default. Merchants that run subscriptions outside Stripe Billing, or that have the feature switched off, do not have it. In this sample 228 of the 281 Stripe accounts had Smart Retries available and enabled. The other 53 ran Stripe’s fixed retry schedule or no automatic retries. FlyCode replaced whichever of those the account ran, and every lift figure is measured against it. Smart Retries appears throughout this report as a reference baseline, never as a ranked entry.
Smart Retries is good at what it was built for. It is a single global model trained across millions of merchants, and it recovers the easy cases: the card that fails on Friday afternoon and clears on Monday morning. On the accounts in this sample that had it, it recovered 54% of failed payment value in the 12 months before the switch.
What it leaves behind is structural, not a bug:
It only runs inside Stripe Billing. A merchant that bills subscriptions another way, or has the feature switched off, has no Smart Retries to fall back on.
It is one model for every merchant. A global model cannot learn that a specific merchant’s customers get paid on the 15th, or that a specific issuer approves retries at 6am and declines them at noon. A per-merchant model can.
It stops. Smart Retries runs up to a set number of attempts over a set window, then applies whatever the merchant configured: cancel the subscription, mark it unpaid or leave it past due. A model that replaces it can keep working the payment inside network limits, and can pick a different day.
It does not retry failed trial conversions the same way. A trial that converts on day 14 and fails on the first charge is not a renewal in Stripe’s eyes. More on this below.
It does not route to a backup payment method. When the primary card is a dead end, Smart Retries has nowhere else to go.
The practical reading: most Stripe companies in this sample were running Smart Retries before they switched, and the lift is measured against it. The 53 Stripe accounts without it measured a higher median lift (21.4%) from a lower starting rate (41%). Neither is an argument against Stripe. We are a Stripe design partner, and our CEO leads the Stripe Community chapter in Tel Aviv. It is the position Stripe-native businesses reach themselves once they compare a global model with one trained on their own customers.
Methodology: lift in recovery rate, not attempted recovery rate
The metric this benchmark uses is lift in recovery rate: how much higher is the share of failed payment value recovered with FlyCode than it was under the platform’s own model on the same account?
Lift = (recovery rate with FlyCode − recovery rate under the platform’s own model) ÷ recovery rate under the platform’s own model
A worked example with the sample’s value-weighted figures: the platforms’ own models recovered 52.0% of failed payment value before the switch, and FlyCode recovered 62.1% on the same accounts. Lift is (62.1 − 52.0) ÷ 52.0 = 19.4%. In dollars, the 10.1 point difference on $224.1M of failed payments is $22.6M.
How the baseline is established
FlyCode replaces the platform’s retry model. Once a company switches, Stripe Billing Smart Retries, the Shopify app’s schedule, PayPal’s retries or Chargebee’s dunning rules stop working that account’s failed payments, and FlyCode’s model takes over. The two never act on the same payment and never run at the same time. So there is no side-by-side control group, and we do not claim one. The comparison is the same account, before and after.
Before. The account’s recovery rate under the platform’s own model in the 12 months before FlyCode went live. A full year is used so the baseline covers every season and is not flattered or penalised by one unusual quarter.
After. The account’s recovery rate with FlyCode during the window, from day 31 onward.
How old the baseline is. 61 companies switched during the window, so their pre-switch year is recent. For the rest it is older: the median company switched 13 months before the window opened. An old baseline is the weakest link in a before-and-after comparison, so we checked it. Among the 190 companies that switched during the window or in the 12 months before it opened, the median lift was 17.1%, against 18.0% for the 217 that switched earlier. The two are close, which suggests the result does not rest on stale baselines, but the older ones deserve more caution.
The baseline is also the billing baseline. FlyCode’s fee is charged only on recovery above this same 12-month baseline. A vendor has a reason to want a baseline set low, because a low baseline raises both the reported lift and the fee. The customer has the opposite reason, and is the one who pays. So the baseline behind every figure here has been checked by the party with the most to lose from it being too low.
A before-and-after comparison has a known weakness: anything else that changed on the account between the two periods ends up in the lift. Three things limit that here.
Same rules on both sides. Both periods use the same 30-day recovery window, the same invoice-level counting and the same exclusions.
Re-weighted failure mix. The pre-switch rate is re-weighted to the window’s decline code mix, so a shift toward easier or harder failures does not show up as lift.
Six months, month by month. A one-off change on an account shows up as a step in a single month. The monthly medians above show none.
What we could not rule out: a pricing change, a shift in customer mix, a change in issuer behaviour, or a seasonal mismatch, since the baseline covers a full year and the window covers March to September only. These are real. Read the lift as measured on the same accounts, not as proven by a randomised experiment.
Full replacement has one more consequence worth stating. A recovery tool that stacks its own retries on top of the platform’s schedule puts two engines on the same card, and both count against the network’s attempt limits. That is not how the accounts in this report run: each failed payment is worked by one model.
Three supporting metrics are reported alongside it:
Recovered revenue as a share of MRR. The lift translated into what a finance team sees. A 20% lift on a small failed-payment base is less money than a 10% lift on a large one.
Time to recovery. Median days from first failure to successful charge. Faster recovery means fewer customers locked out, fewer support tickets and less revenue sitting in a delinquent state.
Share recovered without customer contact. How much of the recovered value came from retry timing alone, versus a step the customer had to take. This is the number that tells you which lever your current setup is missing.
Why we do not use attempted recovery rate
Some vendors report an attempted recovery rate: recovered charges divided by the charges their tool attempted to recover. It is an improvement on the naive gross rate because it removes failures the tool never engaged with. It is still the wrong number for a buyer, for two reasons.
First, it says nothing about the baseline. A dunning tool that fires an email on day three and gets paid on day five has no way of knowing whether Stripe’s day-five retry would have succeeded anyway. The email gets the credit; Stripe did the work.
Second, the denominator is chosen by the vendor. Exclude the hard cases from the attempt set and the rate goes up. Lift in recovery rate has no attempt set to choose: every failed payment on the account counts, on both sides of the comparison.
Attribution rules used in this report
Window. A failed payment counts as recovered if it is paid within 30 days of the first failure. Recoveries after day 30 are not counted in either period.
Same rules on both sides. The pre-switch period and the FlyCode period are counted identically.
Invoice level. One failed invoice is one failed payment however many attempts it takes. Duplicate invoices are removed before rates are computed.
No double counting. A recovered invoice is counted once, however many retries or messages came before it.
Refunds and disputes netted out. A recovered payment that is refunded or charged back within the window is removed from the recovered total.
Trial conversions reported separately. Failed first charges on trial-to-paid conversions are measured in their own section with their own baseline, because the platform baseline for them is close to zero.
If you are comparing this report to another vendor’s benchmark and their figure does not state the baseline, the attribution window and the exclusions, the comparison is not valid. Ask for all three.
By billing platform and payment rail: Stripe, Shopify, PayPal, Apple Pay, Chargebee
The baseline is different on each stack, so the lift is too. Stripe Billing Smart Retries, where a merchant has it enabled, is the strongest default in the sample; Stripe accounts without it run a fixed schedule. Shopify subscription apps (Recharge, Skio) run simpler fixed schedules. PayPal subscriptions retry a failed billing agreement on PayPal’s own short schedule. Apple Pay has no retry of its own: a recurring charge on an Apple Pay token is retried on whatever schedule the processor runs. Chargebee’s dunning is rule-based and configured by the merchant.
Platform | Companies | Model FlyCode replaced (reference) | Recovery rate before | Recovery rate with FlyCode | Median lift | Median time to recovery | Share recovered without customer contact |
|---|---|---|---|---|---|---|---|
Stripe | 281 | Stripe Billing Smart Retries (228 accounts) or fixed retry schedule (53) | 53% | 63% | 16.9% | 6.4 days | 73% |
Shopify subscriptions (Recharge, Skio) | 61 | Native retry schedule | 49% | 61% | 24.8% | 5.2 days | 79% |
PayPal | 28 | PayPal subscriptions built-in retry | 46% | 53% | 14.3% | 7.8 days | 58% |
Apple Pay | 16 | Processor retry schedule on the token (no native retry) | 43% | 52% | 21.7% | 4.9 days | 84% |
Chargebee | 21 | Merchant-configured dunning rules | 50% | 60% | 19.6% | 7.1 days | 61% |
Recovery rates in this table are value-weighted across each platform’s accounts. The lift column is the median company, which is why the two do not divide exactly.
Two readings of this table, and one caution:
Lift is higher where the baseline is weaker, but the money is where the volume is. DTC subscription brands on Shopify show the highest median lift because the native retry schedule leaves the most on the table and because their failure mix skews toward insufficient funds. Stripe accounts show a lower lift against a stronger baseline, but Stripe carries the majority of the sample’s failed-payment value, so the absolute dollars recovered are larger.
The share recovered without customer contact tracks the failure mix, not the platform. Across all five stacks, companies with a high insufficient funds share recover most of their value through retry timing alone. Companies with more hard declines rely more on payment update links and backup cards. The platform you bill on changes the baseline. It does not change which lever matters for your customers. PayPal is the outlier: a failed billing agreement more often needs the customer to re-approve it, which is why it shows the lowest lift and the lowest retry-only share. Apple Pay sits at the other end, since the token rarely goes stale and almost everything recovers on timing alone.
Small cells are directional. Apple Pay (16 companies), Chargebee (21) and PayPal (28) are too few accounts for a stable median. One or two large accounts can move those rows by several points. We print them because leaving them out would hide where the sample is thin.
By business model: SaaS, DTC, media, mobile apps
Business model | Companies | Insufficient funds share of failures | Median lift | Share recovered without customer contact | Median time to recovery |
|---|---|---|---|---|---|
SaaS | 203 | 44% | 15.2% | 62% | 7.2 days |
DTC subscriptions | 98 | 69% | 23.7% | 79% | 5.1 days |
Media and publishing | 41 | 58% | 17.9% | 72% | 6.0 days |
Mobile apps (web-billed) | 65 | 67% | 20.8% | 81% | 4.6 days |
The pattern follows the customer, not the software. Consumer subscriptions billed monthly to personal debit cards (DTC, mobile apps) fail mostly for insufficient funds, and those failures recover on timing. B2B SaaS sees more corporate cards, more hard declines and more failures that need a person to act, so lift is lower and more of the recovery runs through a payment update or a backup card. If your business sits between two rows, your insufficient funds share predicts where you will land better than your category label does.
When recoveries land: the calendar inside the retry
A retry engine sees something an email tool never does: the exact day a previously declined card starts approving again. For insufficient funds recoveries in this sample, the calendar is not flat.
Day of month | Share of the month | Share of insufficient funds recoveries | Recoveries per day vs an average day |
|---|---|---|---|
1st to 3rd | 10% | 21% | 2.1× |
4th to 7th | 13% | 9% | 0.7× |
8th to 14th | 23% | 14% | 0.6× |
15th to 17th | 10% | 17% | 1.7× |
18th to 24th | 23% | 15% | 0.7× |
25th to month end | 21% | 24% | 1.1× |
38% of insufficient funds recoveries landed in two three-day windows that make up 20% of the month. Fridays carried 22% of recoveries, against the 14% an even spread would give.
A fixed schedule that retries 3, 5 and 7 days after a failure hits those windows only by accident. A subscription that fails on the 6th gets all three retries in the slowest stretch of the month. This is the mechanical reason retry timing beats retry count: the same number of attempts recovers more when a model places them than when a calendar does. The windows also shift by issuer and by country, which is what a per-merchant model learns and a global schedule averages away.
What a recovery costs: attempts per recovered payment
A declined retry is not free. It counts against the card’s attempt limit with the network (Visa caps reattempts at 15 in 30 days, Mastercard at 35), it can draw network fees for excessive reattempts, and it tells the issuer that this merchant keeps asking. So we measured how many attempts each recovery took on the same accounts: under the platform’s own model before the switch, and with FlyCode after it.
Metric | Platform’s own model (before the switch) | FlyCode |
|---|---|---|
Retry attempts per failed payment | 4.0 | 2.9 |
Retry attempts per recovered payment | 7.7 | 4.7 |
Share of recoveries won on the first retry | 31% | 47% |
Failed payments that reached a network retry limit | 1.9% | 0.2% |
Fewer attempts recovered more. FlyCode used about 40% fewer retry attempts per recovered payment than the model it replaced. For a finance or payments team this is the figure that separates a better model from a louder retry schedule: a tool that lifts recovery by adding attempts is spending the merchant’s standing with issuers to do it, and the cost shows up later in authorization rates on healthy renewals.
After the recovery: do recovered subscribers stay?
A one-month benchmark stops at the recovered payment. A six-month window can ask the next question: was that subscriber worth recovering? We followed every payment recovered in the first three months of the window for 90 days.
Cohort | Still active and paying 90 days later |
|---|---|
Subscribers whose payment never failed | 88% |
Recovered by retry alone, no customer contact | 84% |
Recovered after the customer took an action | 79% |
Failed and not recovered within 30 days | 6% (came back on their own later) |
Recovered subscribers behave almost like subscribers who never failed. The gap is four points for silent recoveries. 29% of recovered subscribers failed again within the 90 days, and 68% of those second failures were recovered again.
This changes what a recovered dollar is worth. Counting the billing cycles that followed inside the 90 days, each recovered dollar was followed by a further $1.90 from the same subscribers. We do not add that to the headline: the $22.6M counts only the additional payments recovered inside the 30-day window. One caution on the comparison: subscribers whose payments fail differ from those whose payments do not, so the never-failed cohort is a reference point, not a controlled comparison.
What the top performers have in common
We looked at the top quartile of the 407 companies by lift and compared their setup with the bottom quartile. The differences are not about email copy or tone. They are about how much of the recovery flow the model is allowed to run. These are associations in observational data, not controlled experiments, and the traits overlap: accounts that have been live longer are also the ones more likely to have turned on more of the product.
FlyCode replaces the platform’s retry model on every account, so that part is the same for everyone. What varies is the rest. The top quartile shares four traits:
They have been live for more than 60 days. The per-merchant model keeps improving as it sees more of the account’s issuers and customer cycles. Median lift at day 30 was 11.2%. Median lift at day 90 was 19.8%.
They route to a backup payment method. Adaptive Backup Cards accounted for 38% of what the top quartile recovered on hard declines. In the bottom quartile, most accounts had no backup method on file for most customers.
They treat the customer message as the last step, not the first. A payment update link goes out only after the model has concluded the card will not recover on retry. Fewer messages, sent later, convert better and generate fewer support tickets. Top-quartile companies sent 3.4 times fewer customer messages per recovered dollar than the bottom quartile.
They recover trial conversions. 48% of the top quartile had Trial Recovery enabled, against 14% of the bottom quartile.
The common thread is that the top performers let the model decide more than the retry. It also decides when to move to a different card, when a message is worth sending, and whether a failed trial is worth working. A calendar sends the same retry on day 3, 5 and 7 to every customer. A model sends it when that customer’s issuer is most likely to approve it, and goes to a different card or a different channel when it will not.
This is also why we are cautious about the thesis that recovery is mostly a tone problem. Tone matters on the minority of failures that need a human to act. On the majority that need a well-timed retry, the customer never sees a message at all, and that is the point.
Trial payments: the failures nobody measures
A failed first charge at the end of a free trial is not a renewal failure. Stripe does not surface it in recovery analytics the same way, Smart Retries does not treat it like a renewal, and most recovery tools do not touch it. It shows up in a dashboard as a trial that did not convert, and gets filed under marketing.
Across the 126 companies in the sample running trial-based plans:
Metric | Value |
|---|---|
Trial conversions that failed on the first charge | 11.4% of all conversion attempts |
Share of those failures that were insufficient funds | 63% |
Platform baseline recovery on failed trial conversions | 2.1% |
FlyCode recovery on failed trial conversions | 27.5% |
Median recovered trial revenue as % of new MRR | 2.6% |
Trial first charges fail at a higher rate than renewals because the card has never been charged by this merchant before, the amount is new, and the customer entered the card weeks ago with no intention of being charged on a specific day. They are also disproportionately insufficient funds, which makes them disproportionately recoverable on a retry timed to the customer’s cycle.
Trial Recovery is, as far as we know, the only product in the category built for this slice. We report it separately because the baseline is different and because the money is usually not counted at all: a recovered trial conversion is a new customer, not a saved one, and it does not appear in any involuntary churn figure.
Named customers, public results
Sample medians are anonymous by design. These are not. Each result below is a published FlyCode case study with a named customer, and each can be checked at the link.
Customer | What they do | Published result |
|---|---|---|
Creative learning apps (Simply Piano, Simply Guitar) | +33% failed payment recovery rate, +4.2% ARR, 8× ROI | |
Small business referral network | +21% recovery rate, +5.6% ARR | |
Website builder | +18% failed payment recovery | |
Podcast marketing software | Recovery rate from 63% to 91% | |
Supplements, DTC | Recovery rate from 51% to 66% in one month, 13× ROI | |
Fresh meal delivery, DTC | Recovery rate from 62% to 82%, +20% LTV | |
Health supplements, DTC | +22% payment recovery, 25% less passive churn, 24× ROI |
Two cautions on reading these against the sample.
First, they are selected. A company agrees to a case study when the result is good, so these sit above the median by construction.
Second, check that the metric matches. Here it does. A case study reports the relative increase in recovery rate against the period before FlyCode, which is the same before-and-after comparison this report uses. BUBS Naturals going from 51% to 66% is a 29% lift. Gardencup’s 62% to 82% is 32%, and Capsho’s 63% to 91% is 44%. Their starting rates (51% to 63%) sit at or above the sample’s 52% baseline, and their lifts sit above the 75th percentile of 25.6%. That is what you would expect from published wins, and it is the reason the median, not the case study, is the number to plan with.
What this implies for operators
Five practical takeaways from the 407-company sample.
1. Benchmark against the median lift, not the headline aggregate
The total recovered across the sample is informative but scales with sample composition. The median lift of 17.6% is the right anchor for modelling your own expected outcome. If your platform recovers $5,200 of a $10,000 monthly failed-payment base, a 17.6% lift on that rate is about $915 more a month.
2. Your platform’s model is the baseline, not the ceiling
Smart Retries, the Shopify apps’ schedules, PayPal’s retries and Chargebee’s rules do real work: together they recovered 52% of failed payment value on these accounts before the switch. The 228 Stripe Billing accounts that replaced Smart Retries with a per-merchant model measured a median lift of 16.2% against it. If you are on Stripe without Stripe Billing, you do not have Smart Retries at all, and your baseline is a fixed schedule or nothing. One practical rule: run one retry model per payment. Stacking a second engine on top of the platform’s own doubles the attempts against the same card.
3. Look at your failure mix before you buy anything
Pull your failed payments for the last 90 days by decline code. If insufficient funds is half or more of them, which is typical, your gap is retry timing and your fix is a model, not a message sequence. If hard declines dominate, which is rare outside B2B with corporate cards, backup payment methods and update links are where the money is. A dunning email tool addresses the smaller of the two problems.
4. Count the trials
If you run free trials, your involuntary churn number is understated. Failed first charges never make it into it. Pull the count of trial conversions that failed on the first charge and multiply by your plan price. For 58 of the 126 trial-based companies in this sample, that number was larger than the incremental renewal revenue FlyCode recovered.
5. Ask every vendor how its baseline is measured, including us
Before you sign with a recovery vendor, ask three things. Does the tool replace your platform’s retries or stack on top of them? What is the baseline, and when was it measured? What is the attribution window? Our answers: FlyCode replaces the platform’s model, the baseline is your own account in the 12 months before the switch, and the window is 30 days. If a vendor claims a side-by-side test, ask how two retry engines shared the same payments.
To see your own numbers against this benchmark, request a free payment audit. It runs read-only against your billing platform and returns your failure mix, your platform baseline and your projected lift, with no fee unless recovery rises above your baseline.
Limitations: what this report cannot tell you
It is written by a vendor. FlyCode produced this report from its own customers’ data. No third party has audited it. The mitigation is disclosure, not independence: the method, the exclusions and the unflattering cuts are all above.
The comparison is before and after, not randomised. FlyCode replaces the platform’s model, so there is no control group running at the same time. Anything else that changed on an account between the pre-switch period and the window is inside the lift. For long-standing customers that pre-switch period is more than a year old, and the methodology section shows the lift split by baseline age.
The sample selected itself. These are companies that chose to buy a payment recovery platform, which usually means their failed payment problem was larger than average. Lift on a company with little involuntary churn will be lower, as the bottom band shows.
Six months is not a year. The window runs from March to September. It contains no year-end holiday season and no January, so seasonal effects outside the window are not measured, and the 12-month baseline covers seasons the window does not.
Small cells. Apple Pay, Chargebee and PayPal have 16, 21 and 28 companies. Their rows are directional.
Associations are not causes. The top-quartile traits and the retention follow-up describe what went together in this sample. They were not randomised.
Geography and currency are not broken out. Issuer behaviour differs by country, and this edition reports a single blended figure.
Medians hide the range. The middle half of companies spans 10.3% to 25.6%. One company’s result can sit well outside the median for reasons specific to its customers.
If you find an error or a cut you think is unfair, write to hello@flycode.com. Corrections will be listed with their date in the next edition.
How this benchmark gets updated and how to submit data
We will republish this report twice a year on a comparable sample, with the same methodology: lift in recovery rate against each platform’s own model, full-sample reporting rather than a curated subset, and the attribution rules above stated in full. Year-over-year comparisons will show where the recovery landscape is moving, in particular as processors tighten retry limits and as more of the category moves from scheduled retries to per-merchant models.
This report sits alongside our open, criteria-based vendor benchmark, where any vendor can submit cohort data for scoring: the Stripe Failed Payment Recovery Benchmark.
Merchants and vendors who want their cohort included in the next release can submit data to FlyCode. Submissions need the cohort definition, the baseline retry configuration, the attribution window, the exclusions and the failure cause classification. Submissions missing any of the five are listed as insufficient evidence rather than scored.


