What Your Payment Data Already Knows About Your Business

Most businesses look at their payment data exactly once a month, as a single number: revenue. That number is the least informative thing in the file. The same records answer whether customers come back, what your revenue actually depends on, which customers are price-sensitive, and whether this month's growth came from acquiring more people or from the ones you have spending more — questions usually settled by intuition or by paying someone to research them badly. The obstacle isn't access. It's that the standard way of summarizing transactions, monthly totals and averages, conceals nearly every finding worth having.

Cohort retention

The single most valuable analysis available from transaction records, and it takes one pivot table.

Group customers by the month they first purchased. Track what share of each group transacts in each subsequent month.

First purchase	Month 1	Month 3	Month 6	Month 12
January	100%	41%	29%	22%
April	100%	38%	26%	19%
July	100%	33%	21%	—
October	100%	29%	—	—

Read down a column. Month-3 retention fell from 41% to 29% across the year — and total revenue may have risen the whole time, because more customers were being acquired. The business is growing and its economics are deteriorating simultaneously, which no monthly revenue chart shows.

What causes it, usually:

• A channel change. New acquisition sources bring different customers, and the selection effect is frequently larger than anything about the product.
• Discounting, which acquires price-driven customers who don't return at full price.
• A product or service change affecting the second purchase rather than the first.
• Genuine market saturation in the best segment.

The reason this analysis is worth more than its effort: it's a leading indicator. Deteriorating retention shows up in cohorts months before it shows up in revenue, because acquisition masks it — and by the time revenue reflects it, the accumulated cost is substantial. This is the same cohort logic our benchmarking analysis applies to collections, for the same reason: cohorts separate a change in performance from a change in mix, and blended figures never can.

The distribution, not the average

Average order value is among the most-quoted and least-useful business metrics, for a specific reason: most transaction distributions aren't symmetrical, so the average sits where relatively few transactions actually occur.

Plot the transactions in bands instead:

Band	Share of transactions	Share of revenue
Under $50	34%	7%
$50–$150	41%	26%
$150–$400	11%	18%
$400–$1,200	9%	27%
Over $1,200	5%	22%

The average here lands somewhere around $180 — in the $150–$400 band, which is the sparsest part of the distribution. The business doesn't have one customer type; it has at least two, and the average describes neither.

What the shape enables:

• Different pricing for different clusters, since the price-change arithmetic in our pricing guide should be run per segment rather than blended.
• Rail selection by band. Card fees have a percentage component, so large transactions cost disproportionately more — and the flat-fee alternatives in our bank payments guide pay for themselves in the top band specifically.
• Fraud screening by band, since the cost tradeoff in our fraud analysis differs entirely between a $40 order and a $1,400 one.
• Knowing what a small-transaction customer is worth, since the 34% generating 7% of revenue may still be valuable if they convert upward — which the cohort table answers.

Concentration

What your revenue actually depends on, which most owners estimate badly in the optimistic direction.

Calculate:

• Share of revenue from your top customer, top five, and top ten.
• Share from your largest channel.
• Share from your largest product or service line.
• Share from your largest geography where relevant.

Then ask the only question that matters: what happens if the largest one leaves?

Concentration isn't inherently bad — serving fewer, larger relationships is frequently more efficient. What it changes:

• How much cash buffer you need, since losing a concentrated relationship is a step change rather than a drift.
• How a lender views you, per the assessment in our valuation guide — concentration reduces borrowing capacity and can trigger covenant attention.
• What a buyer will pay, since concentration is among the largest valuation discounts.
• Your negotiating position on price and terms with the concentrated party, who knows their weight.

The subtler point, and the one worth taking from our correlation analysis: a customer base that looks diversified by count can be concentrated by exposure. Two hundred customers in one industry, or one region, or dependent on one upstream supplier, move together. Counting customers understates concentration whenever they share a common factor — and payment data frequently reveals the sharing through timing correlation that a customer list doesn't show.

Where growth came from

Revenue growth has exactly three sources and they behave completely differently:

Growth = New customers + Existing customers spending more + Price

Decompose a period:

Source	Contribution	What it implies
New customers	+$210,000	Costs acquisition spend and working capital
Existing spending more	+$84,000	Cheapest growth available
Price	+$41,000	Flows almost entirely to profit
Lost customers	−$126,000	The number usually missing from the story
Net	+$209,000

The fourth row is the one that changes decisions. This business acquired $210,000 of new revenue and lost $126,000 of existing revenue — so 60% of its acquisition effort went to replacing customers it already had. A retention improvement worth a fraction of the acquisition spend would produce more.

And the cash implication ties to our growth analysis: new-customer growth consumes working capital and price growth doesn't. A business growing through the third row is growing without a cash requirement; one growing through the first row is funding it.

Timing and seasonality

The quickest analyses, and each maps to an operational decision:

• Revenue by day of week and hour — staffing and opening hours.
• Revenue by week of month, which frequently tracks pay cycles. A business whose customers cluster around pay dates is exposed to the income timing in our volatility analysis, and can move promotions accordingly.
• Monthly seasonality across several years, which drives the financing timing in our seasonal guide.
• Time between first and second purchase, which tells you when a follow-up would actually land — most businesses guess this and guess late.
• Payment method by band and channel, which feeds the acceptance cost work in our effective rate guide.

The second-purchase interval is the underrated one. A business that knows its median repeat interval is 34 days can time outreach to day 30, rather than to whatever cadence the marketing calendar happens to use.

What the data can't tell you

Before the limits, one more analysis worth the effort.

Who is price-sensitive

You can't run a controlled pricing experiment on your history, but transaction records reveal sensitivity through behaviour you've already observed:

• Response to past price changes. If you've adjusted prices, compare cohorts either side. Which segments continued at the same frequency and which reduced?
• Discount dependence. Customers whose purchases cluster around promotional periods are telling you their reservation price. A customer who only ever buys on discount is a customer whose full price you've never established.
• Basket composition. Customers buying only your lowest-margin lines behave differently from those buying across the range.
• Frequency versus value. A customer buying often at low value and one buying rarely at high value have different sensitivities and should not be treated as one segment.

Why this matters practically: the break-even volume loss in our pricing analysis should be computed per segment rather than blended. A business raising prices uniformly is applying one answer to populations with different tolerances — and the segment most likely to leave is usually the one contributing least, which means the blended calculation understates how much room exists.

The related finding worth checking: whether your discount-dependent customers ever convert to full price. The cohort table answers this directly — track customers acquired during promotional periods separately and see whether their subsequent purchases occur at full price. If they don't, promotional acquisition is buying revenue rather than customers, and the acquisition cost should be measured against a single transaction rather than a lifetime.

Four honest limits, because over-reading transaction data produces confident wrong answers.

It shows what happened, not why. Retention fell — because of the channel, the product, the price, or the market? The data ranks hypotheses; it doesn't test them. Every finding here is a hypothesis, and confirming it requires a change and a comparison.

It's censored. You see people who bought. You don't see people who considered and didn't — the same structural blindness our reject inference analysis describes in lending. A business whose checkout is losing people sees nothing at all in its transaction file.

Identity is imperfect. Repeat customers using different cards or emails look like new customers, which biases retention downward and inflates acquisition counts.

Correlation isn't causation, and small samples aren't findings. A cohort of 40 customers moving four points is noise. Check that the effect is larger than the variation between similar periods before acting on it.

Making it a habit

1. Export twelve months of transactions with customer identifier, date, amount, method, and channel.
2. Build the cohort table first. Highest value, and it establishes the baseline everything else is read against.
3. Plot the distribution, in bands.
4. Compute concentration, top one, five, and ten.
5. Decompose growth, including lost customers.
6. Write down what you expect before you look, which turns each analysis into a test of your intuition rather than a confirmation of it.
7. Re-run quarterly, and compare against your own prior figures rather than against benchmarks.

Step six is the discipline that makes this worth doing. An owner who predicts 45% month-3 retention and finds 29% has learned something specific about their own judgment, which is more durable than the individual number.

Frequently asked questions

---

This guide is educational and does not constitute financial or business advice. Worked figures are stylized illustrations. Analysis of customer transaction data should be conducted consistently with applicable privacy obligations and with the terms under which the data was collected.