Skip to content

Part 5 of 7 · Abandoned form recoverer series ~5 min read

How recovery is honestly measured

This category reports conversion rates that are wrong by a large factor, and the reason is simple enough that it is worth stating plainly: a substantial number of people who abandon a form come back and finish it without any prompting at all, and the naive measurement credits every one of them to the message.

Key takeaways

  • Many abandoners return on their own. A naive rate credits the message with all of them.
  • A holdout — a share who are deliberately not messaged — is the only honest measurement.
  • Ten per cent is enough, and it costs a small number of recoveries to know the truth.
  • An honest uplift is usually modest, and modest is still worth having.
  • The complaint and opt-out rate belongs in the same report as the uplift.

The people who come back anyway

How a holdout group makes recovery measurement honestA vertical chain of five steps entered by a box labelled One hundred abandonments that passed the basis check. Step one messages ninety and holds back ten as a random, fixed holdout. Step two asks how many of the messaged returned, say eighteen of ninety. Step three asks how many of the holdout returned, say one of ten. Step four computes the naive rate at twenty per cent, which is wrong. Step five computes the real uplift as twenty per cent minus ten per cent, giving ten percentage points. A note says half the naive number was people who were coming back regardless.AWS ACCOUNT100 abandonmentsbasis check passedMessage 90hold back 10Holdoutrandom, and fixedMessaged: how many returned?say 18 of 90Holdout: how many returned?say 1 of 10Naive rate20% -- and wrongReal uplift20% minus 10% = 10 pointsHalf the naive number was people who were coming back regardless.
Fig 1. Why a holdout is the only honest measurement. The gap between the naive rate and the uplift is entirely people who needed no prompting.
  • App integration
  • Management
  • Analytics
  • Front-end & mobile

Why a holdout rather than a before-and-after

The obvious alternative — compare the recovery rate before the system existed with after — is confounded by everything else that changed. Seasonality, a site redesign, a different mix of traffic. A holdout running concurrently is the only comparison where the two groups differ in one thing.

Ten per cent is enough at typical volumes and it costs a small number of recoveries a month. It is worth paying: without it, nobody can say whether the system is doing anything, and a system nobody can evaluate tends to acquire a second message and then a third.

Keeping the holdout honest

The holdout has to be assigned randomly and permanently at the moment of abandonment, not chosen afterwards, and the assignment has to be recorded before the outcome is known. Assigning it retrospectively, or excluding holdout cases that look unusual, produces a number that is worse than not measuring.

It also has to survive somebody wanting to switch it off. The pressure to message the holdout arrives about four months in, framed as “we know it works now”, and the correct answer is that knowing it still works next year requires it to keep running.

What an honest number looks like

One quarter of form abandonment recovery measured honestlyA horizontal row of five boxes. Abandonments: three hundred and ten this quarter. Basis passed: one hundred and four. Messaged: ninety-four. Uplift: nine percentage points. Extra enquiries: about eight. A note says eight enquiries a quarter is modest, real, and worth the two hours it took to build.ONE QUARTER, HONESTLYAbandonments310 this quarterBasis passed104Messaged94Uplift9 percentage pointsExtra enquiriesabout 8Eight enquiries a quarter. Modest, real, and worth the two hours it took to build.
Fig 2. A quarter measured honestly. The number is smaller than the category usually reports and it is one somebody could reproduce.
  • App integration
  • Machine learning
  • Security & identity
  • Analytics
  • Front-end & mobile

Eight extra enquiries a quarter is a genuinely modest result and it is worth stating as such rather than dressing it up. For a business where an enquiry is worth a few hundred pounds, eight is a good return on a small system; for one where an enquiry is worth twenty, it probably is not worth building at all, and knowing that is useful.

The drop from three hundred and ten abandonments to one hundred and four that pass the basis check is the other number worth looking at, and it is the system behaving correctly. Two thirds of abandonments are people who typed an address somewhere without context, and not contacting them is the point.

The number next to the uplift

The complaint and opt-out rate belongs in the same table, always, because the uplift on its own invites the obvious optimisation and the two numbers together do not.

Reported together, every quarter

  • Uplift: 9 percentage points against the holdout.
  • Opt-outs: 3 of 94 messaged asked not to be followed up again.
  • Complaints: 0 spam reports, 1 reply asking how we got their address.
  • That last one matters even at a count of one. It means somebody did not recognise the context, which is a signal about the form rather than about them.
  • If opt-outs exceed about 5%, the basis check is too loose and should be tightened rather than the wording softened.

That last rule is the one worth writing down somewhere durable. When a recovery message generates complaints, the instinct is to rewrite the message, and the message is almost never the problem. The problem is that it was sent to somebody who did not have enough context to recognise why, which is a basis question.

Next: what all of this costs to run.

All posts