Reference
How the analysis works, where the benchmarks come from, what FeeLeak refuses to calculate, and who operates it.
How the analysis works, where the benchmarks come from, what FeeLeak refuses to calculate, and who operates it.
This page exists because the person who signs an audit is the one who has to defend it. If you are checking a FeeLeak report before putting your name on it — or challenging one someone sent you — this is the document to read.
Every card transaction carries a cost the card networks set: interchange, paid to the card-issuing bank, and assessments, paid to Visa, Mastercard, Discover or American Express. Those are identical no matter which processor handles the account. Nobody negotiates them.
Everything above that figure is the processor's own compensation — the markup. That is the only negotiable layer, and separating it from the network cost is the entire job of a merchant processing audit.
FeeLeak reads a statement, splits those two layers, compares the markup against what comparable merchants pay, and classifies every remaining line.
Extraction and classification are deliberately handled by different parts of the system, because they fail in different ways.
A language model reads the statement and returns facts: each line's description, amount, transaction count, volume, rate and per-item charge. Models are reliable at reading values off a page. They are less consistent at judgment calls — is this line interchange, or is it markup? — and that inconsistency moves between runs.
So FeeLeak makes that judgment in its own code, with a fixed classifier, not in the model.
Processors rarely print the word "interchange" beside an interchange charge; they print the card brand's
program name — CPS/Retail, Merit III, PD2 Signature,
EIRF, Regulated-Issuer. The classifier matches those program names directly.
Markup is then always computed as the remainder — total fees minus pass-through — rather than taken as a separately extracted figure that could drift against the total.
Why this matters for reproducibility. The same statement analyzed twice returns the same figures, because the figures no longer depend on the part of the process that varies. Extraction runs at zero temperature, and the reconciliation rules are hard requirements: line items must sum to total fees, no dollar counted twice, section subtotals are not line items.
When extracted line items do not reconcile to within 10% of stated total fees, FeeLeak falls back to the model's category totals and says so on the report, naming the coverage percentage, so you know that particular result is less reproducible than usual.
Statements that bundle interchange produce no savings figure at all.
Interchange plus assessments is never below roughly 1.2% of volume. The networks set that floor. A statement that appears to have almost no pass-through cost has not found a merchant with free interchange — it has bundled or tiered pricing that never itemizes it.
Without that separation, the processor's markup cannot be isolated, and any savings figure would be guesswork. So below the 1.2% floor FeeLeak suppresses the overpayment figure entirely, suppresses the renegotiation letter, and tells you to request an interchange-itemized statement instead.
This was found by testing against real statements. An earlier version reported a merchant as overpaying $42,811 a year when their actual all-in rate was 1.77% — a good rate. An audit that produces an impressive wrong number is worth less than no audit.
The comparison basis is a market benchmark for the merchant's volume band (micro, small, mid, large) and transaction profile (card-present, mixed, card-not-present), expressed as a fair processor markup in basis points, per-item, and monthly fee.
It is never a rate FeeLeak or any affiliate is offering. A proposal dressed as an audit is a proposal, and the moment a sell rate appears in the comparison the document stops being independent.
Current status, stated plainly: the benchmark tables are industry seed estimates, not yet derived from observed statement data. Every report discloses this in its own methodology section and labels the comparison directional. As the corpus of parsed statements grows, the benchmarks will be replaced with observed aggregates and reports will state the sample size behind each one.
Benchmarks are also weak at very low volume, where a flat monthly fee dominates the effective rate. Below roughly $5,000 a month, compare dollars rather than percentages.
Every analysis carries an extraction confidence score. Below 85%, the report is marked unverified and the app tells you to check the line items against the source statement before releasing it to a client. That flag is on the first page of the PDF, not buried.
Confidence drops on statements with overlapping sections, poor scan quality, unusual formats, or figures that will not reconcile. When the extraction cannot balance, FeeLeak lowers confidence and names the gap in dollars rather than guessing.
The default recommendation in every report is to renegotiate with the merchant's current processor before considering replacement. That is not diplomacy — it is what the evidence supports:
Each report includes a ready-to-send letter to the incumbent setting out the specific requests. Replacement is the reasonable next step only when the incumbent declines to move.
FeeLeak is operated by Primary Payments LLC, which also places merchant processing as an independent agent.
FeeLeak's analysis compares a statement against market benchmarks, never against a rate FeeLeak or any affiliate is offering, and its default recommendation is renegotiation with the merchant's current processor rather than replacement. Where a referral to a processing agent is offered, it is optional, disclosed, and made only after renegotiation has been attempted.
Stating this plainly is the point. The tool was built by someone who reads these statements
professionally — that domain knowledge is why it can tell CPS/Retail from a discount fee at
all. The obligation that comes with it is disclosure, and design choices that hold up under scrutiny
rather than assurances that don't.
If you are evaluating a FeeLeak report and something doesn't add up — or you think the method is wrong somewhere — email wade@vantyx.io. Corrections to this page get made in public.