Skip to content
allanninal.dev/build
Collaborate

Part 2 of 7 · Win-back campaigner series ~5 min read

How lapsed gets defined

The definition of lapsed determines everything downstream, and the default definition — a fixed number of days for everybody — is wrong for most of the list in one direction or the other.

Key takeaways

  • Use the customer’s own median interval for that product, not a global number.
  • Two and a half times their gap is a defensible line and is a judgement, not a law.
  • Seasonal products need the season, not the interval.
  • Fewer than three orders means no interval; those customers are handled separately or not at all.
  • A customer who bought once is not a lapsed customer, and treating them as one is a category error.

One threshold, two failures

How a fixed ninety-day lapse rule fails different buying frequenciesA stacked bar chart with three bars measured in days. Two series: their own buying interval in green, and the extra days a ninety-day rule makes you wait in red. The weekly buyer has an interval of seven days and eighty-three extra days of waiting. The monthly buyer has an interval of thirty-two days and fifty-eight extra days. The annual buyer has an interval of three hundred and sixty-five days and no extra waiting, because ninety days is well inside their normal gap. A note says the annual buyer is flagged at ninety days and has not lapsed at all, so the rule is wrong in both directions.0200400600800~90Weekly buyer~90Monthly buyer~365Annual buyerTheir own buying interval, daysExtra days a 90-day rule makes you waitThe annual buyer is flagged at 90 days and has not lapsed at all. Wrong both ways.
Fig 1. Why a single threshold cannot work. The weekly buyer is left for nearly three months past the point they clearly stopped; the annual buyer is chased a quarter into a normal gap.

The interval that matters is theirs

Take the median gap between that customer’s orders of that product. Median rather than mean, because one long holiday gap distorts an average and does not distort a median. Per product rather than overall, because somebody who buys a consumable monthly and furniture every four years has two intervals and averaging them describes neither.

Then lapsed is roughly two and a half times that median with no order. For the weekly buyer that is about eighteen days; for the annual buyer it is two and a half years. Both of those feel right to anybody who knows the customer, which is the test.

Where the multiplier comes from

It is a judgement, and it is worth saying so rather than dressing it up. Below about two times, you are contacting people who are merely a bit late and the message reads as pestering. Above about three, the relationship has cooled enough that a reminder no longer works and you need the harder conversation.

Two and a half sits in the middle and should be adjustable per product category, with the value visible in the configuration rather than buried in code. Somebody will want to change it, and the argument about what it should be is a better argument to have explicitly.

The customers with no interval

How a customer is assessed as lapsed against their own intervalA vertical chain of five steps entered by a box labelled A customer to assess, with an order history. Step one asks whether there are three or more orders of this product; if not it exits to No interval exists, and does not infer one. Step two asks whether the product is seasonal, judged by product rather than by customer; if so it exits to Use the season, lapsed after the next season passes. Step three computes the median gap rather than the mean. Step four asks whether they are past two and a half times it with no order; if not it exits to Not yet, check again next week. Step five marks them lapsed and hands them to suppression. A note says one-time and two-time buyers are a different problem with a different message.AWS ACCOUNTA customer to assesswith an order historyThree or more orders?of this productNo interval existsdo not infer onenoIs it seasonal?by product, not customerUse the seasonlapsed after next seasonyesMedian gapnot the meanPast 2.5x it?with no orderNot yetcheck again next weeknoLapsedhand to suppressionOne-time and two-time buyers are a different problem with a different message.
Fig 2. How lapse is decided for one customer. The first exit is the largest group in most businesses and the one that should not be run through this system at all.
  • Compute
  • App integration
  • Security & identity
  • Management
  • Analytics
  • People

One-time buyers are not lapsed

In most businesses they are the largest group by a wide margin, and calling them lapsed is a category error with practical consequences: the message that works for a two-year customer who stopped is completely wrong for somebody who bought once eighteen months ago and may not remember doing it.

The honest answer is that they are a different campaign, with a different message, and quite possibly not worth running at all. Excluding them keeps the win-back list small and its response rate high, which is what makes the whole thing worth doing.

Seasonal products

Somebody who buys garden furniture every May has an interval of a year, but they are not lapsed in February; they are waiting for spring. Seasonal products need the season as the unit: lapsed means the season came and went without an order.

Seasonality is a property of the product, set once by somebody who knows the business, not something to be inferred from thin per-customer data. Two orders in two consecutive Mays is not enough evidence to conclude anything, and an inference engine confident enough to try will also be confident about noise.

Next: whether to contact them at all.

All posts