Returning-customer rate depends on the history you keep

A returning-customer rate needs more than a reporting month. It also needs a rule for how far back to look. If the reporting month stays the same but earlier purchases disappear from the export, the accounts buying this month have not changed. What the report can recognize about their history has.

We held that monthly group fixed and compared exports containing 90, 180 or 365 days before the month began. Each export also contained the whole reporting month. An account counted as observed returning if it had a qualifying purchase before the month, or at least two distinct purchase occasions within it. ^[2]

That second route matters. Someone can make their first observed purchase and another purchase later in the same month. Removing earlier history does not remove that within-month repeat. Our comparison therefore does more than label everybody absent from a short previous-period list as new.

Across January through November 2011, using 90 rather than 365 days of pre-month history reduced the observed-returning share by 11.6 to 25.9 percentage points. The median monthly difference was 19.3 points. The table keeps all eleven months visible, including the smaller January difference.

The direction is built into the calculation: adding earlier purchases can reveal more repeat evidence, never less. The useful result is how much the label changed in this retailer’s records, and how consistently that change appeared across the fixed months. These figures describe a historical reporting choice, not a change in customer behavior.

January gives a concrete example from the actual data. There were 735 customer accounts with a qualifying purchase that month. With 90 days of earlier history, 580 counted as observed returning: 78.9% of those accounts. With 365 days, the count was 665, or 90.5%. Restoring the older records changed 85 labels, a difference of 11.6 percentage points.

Now look at the smaller group a marketer might examine next. The short-history report left 155 accounts with no observed repeat. Of those, 85 had earlier qualifying purchase evidence in the 365-day comparison. That is 54.8% of the short-history no-repeat group, not 54.8% of all January buyers.

Both percentages answer a useful question. The 11.6-point difference describes the whole monthly report. The 54.8% describes how much older evidence sits inside the group left without a repeat label. Confusing their denominators can make the same change sound much larger or smaller.

This distinction persisted beyond January. In each of the eleven months, more than half of the 90-day no-repeat group had earlier evidence within 365 days. The monthly shares ranged from 51.3% to 67.7%. Their earlier qualifying purchases were outside the shorter history but already present in the retailer’s records.

For a campaign brief, the immediate question is therefore what “new” is meant to promise. A first-order welcome message assumes something stronger than “no repeat found in this export.” The analysis does not tell us which message would work better. It shows why the audience description needs to be settled before that creative decision.

Using 180 days before the month brought the report closer to the 365-day comparison. The remaining monthly difference was 3.9 to 11.7 percentage points, with a median of 7.9. Among accounts still without an observed repeat under the 180-day rule, 22.9% to 52.3% had earlier evidence under the 365-day rule.

We also checked whether the result depended on how a purchase occasion was defined. One alternative included all positive numeric-invoice activity, including lines outside the product-code scope. Another counted distinct purchase days, so two purchases on one day could not create a within-month repeat. The history differences remained under both definitions. ^[4]

Neither 180 nor 365 days emerges as the right setting for every business. Extending 365 days to all available earlier records still changed the primary monthly share by 0.5 to 4.1 points. The archive began in December 2009, so even its longest view could not establish anyone’s lifetime purchase history.

The extra history available beyond a year also grew as 2011 progressed. That changing opportunity prevents us from reading the later-month differences as evidence of a seasonal effect. The practical comparison is between clearly named windows, not between a supposedly incomplete report and perfect customer knowledge.

A shorter history can be intentional. A team may want to distinguish accounts that bought recently from accounts whose last visible purchase was farther back. A 90-day recent-buyer segment can serve that purpose. It becomes misleading only if its description promises a different fact, such as a customer’s first purchase with the business.

Start with the decision the report will support. A monthly trading summary may need a stable returning-account definition so that changes over time are interpretable. A re-engagement brief may deliberately focus on a defined gap since a purchase. A welcome brief may need evidence about the first purchase. Those jobs should not inherit one vague “new versus returning” label.

Ask the report owner for the earliest purchase date retained, the reporting cutoff and the exact repeat rule. If an agency or creator receives an audience description, include those details in plain language. “No repeat observed in the retained 90-day history or this month” is less convenient than “new customers,” but it does not promise knowledge the export lacks.

Keep the same monthly accounts in the comparison. Replacing a three-month customer file with a year of customers would change both the history and the audience denominator. Our audit restored earlier evidence for the same accounts buying in the reporting month; it did not compare two differently sized customer lists.

Finally, distinguish a data change from a performance change. If a new export restores older orders, the reported returning share can rise even when no campaign has influenced a purchase. Mark that definition change in the report before attributing the movement to a promotion. The study measures this reporting mechanism, not the effectiveness of any message or channel.

A short handoff can make the distinction usable without turning a campaign discussion into a database review. Ask the person maintaining the order export to complete the fields below. They describe the reporting evidence; message testing and permission to contact customers remain separate decisions.

Purpose: [monthly performance report, welcome message or re-engagement brief]. State the decision the label will inform. If the objective is recent buying, name the recent period explicitly.

Reporting month and audience: [start date], [end date], [accounts with a qualifying purchase in that month]. Keep that account list fixed when comparing history windows. Note whether an account can represent a business rather than an individual shopper.

Available history: [earliest retained date], [days before month start], [whether the complete reporting month is included]. Record an export limit as a limit, rather than describing the first visible order as the first order ever.

Purchase rule: [which transactions count], [how equal timestamps and same-day orders are handled], [how missing identifiers, conflicting invoices and cancellations are treated]. Ask for the actual rule used in the report, not an assumed definition.

Comparison: [returning accounts under each history], [unchanged monthly account count], [accounts whose labels differ]. Add the share of the short-history no-repeat group with older evidence. That last denominator helps the team understand the specific audience it is about to describe.

Decision and follow-up: [retain the recent-buyer definition, restore available history or rename the label], [owner], [next reporting date]. Keep the chosen rule consistent in future reports, and annotate a change when the data source or definition changes.

In the January example, the comparison field would say “580 versus 665 observed-returning accounts out of 735; 85 changed labels.” The audience note would add “85 of 155 accounts without a repeat under the shorter rule had older evidence.” Neither line calls those accounts first-time customers or predicts how they will respond to a campaign.

Before comparing a dashboard with an export, check the definitions behind both. Shopify’s customer-report documentation distinguishes first-time and returning customers and recognizes repeats within the first purchase period. Amazon has described extending purchase-history availability for new-to-brand measurement, especially where buying cycles are longer. Their reporting definitions differ from this monthly export audit. ^{[3] [7]}

Ask which earlier purchases the tool can see, whether it includes additional purchases within the reporting month, and whether the account denominator matches the export. Different definitions can produce different answers without either calculation being broken. A shared metric name is not enough to establish comparability.

The eleven-month comparison shows why this check can matter. It keeps the monthly accounts fixed, makes the purchase rule explicit and preserves the smaller as well as larger differences. Another business can use that distinction when reviewing its report, without borrowing this retailer’s percentages as a performance target.

We used the two-sheet Online Retail II workbook published by UCI, covering December 1, 2009 through December 9, 2011. January through November 2011 were fixed as the complete reporting months with a full 365 days of earlier source coverage. December 2011 was excluded because the source ends partway through it. ^[1]

The workbook contains 1,067,371 rows. Across the whole source, 243,007 rows had a missing or invalid customer identifier. We did not assign an identity from neighboring rows. Separately, 82 numeric invoices contained conflicting timestamps and were excluded in full; no timestamp was guessed. These are overlapping source-quality counts at different levels, not successive deductions from a monthly customer total. ^[5]

A qualifying line required a numeric invoice identifier, a valid customer identifier and date, and positive quantity and price. The primary product-code rule used five digits with optional letters, included two documented product-code exceptions, and excluded two documented carriage or packing charges. Descriptions were reviewed before the customer-history calculations. The public code records the exact rules. ^[5]

Each account’s eligible invoices at an identical timestamp became one observed occasion. This prevents two invoice numbers at the same recorded time from inventing an earlier purchase. Repeated lines across or within the source sheets cannot create additional occasions. Cancellation-prefixed invoices and nonpositive lines do not create purchases; we did not link refunds to original orders or claim that positive invoices were ultimately retained.

For each month, the denominator was the accounts with at least one eligible occasion during that month. An account was observed returning if it had an earlier occasion in the retained pre-month history, or at least two strictly ordered occasions within the month. History began exactly 90, 180 or 365 days before month start and included that boundary. The following month’s first instant was excluded from the reporting period. ^[2]

We screened invoice integrity using the whole supplied archive before making the monthly comparisons. Within that fixed eligible source, later purchases cannot affect an earlier month’s label. This is a retrospective audit, not a simulation of what a live system would have known before later conflicting source records arrived.

The download contains all 165 rate rows, 99 comparison rows and the fixed summary and sensitivity results. The comparisons are descriptive arithmetic for this archive, without confidence intervals or a claim that repeated account-month observations are independent. A separate implementation reproduced every registered aggregate from the original workbook. No customer identifiers or account histories are included in the public outputs. ^[6]