Skip to content

44% of installs never finish onboarding, and finishing is the strongest activation signal we have #480

Description

@coaxk

The activation answer that came out of the #202 dive, once the id-churn measurement fault (#473) was fixed and the numbers became readable.

The finding

Genuine installs, latest ping each, last 30 days, n=214:

installs with a real library (100+ files) share
finished onboarding 114 (53.3%) 83 73%
did not finish 95 (44.4%) 13 14%
unknown 5 (2.3%)

Finishing the wizard takes an install from 14% activated to 73% activated. Nothing else in the dataset separates activated from not-activated anywhere near as sharply. Library size, which is what #202 was originally opened about, turned out to be downstream of this.

44% never finish. That is the loss, and unlike the 99.5% figure that started #202, this one is real rather than an artefact.

Two distinct cohorts, do not merge them

The 95 who never finish are the bulk of the loss. 66% of them also report their subgen as unreachable (#479), which is consistent with stopping partway through rather than being a separate failure. For this group the question is where in the wizard they stop.

The 31 who finished and still show <100 files are 27% of finishers and a harder problem. The product reached them, they completed setup, and they still did not activate. A fix aimed at the first group will not touch these.

What is missing before designing anything

onboarding_step is recorded (0-11) and is the obvious next cut: it says where they stop, which is the difference between a wizard that asks too much too early and one that fails at a specific integration. I deliberately have not run it yet, because the step numbering needs checking against the current wizard before the distribution means anything. A step index that shifted when the wizard changed would produce a confident, wrong answer about which screen loses people.

Worth remembering from #202: every earlier conclusion drawn from this table was wrong because the denominator was wrong. The same caution applies to reading step numbers whose meaning may have drifted.

Caveat

Genuine requires positive evidence of a real install, so an install that arrived today on an older client is excluded until its second day. These figures undercount by construction. That was the deliberate trade against a number that was 80x too high.

Related: #202 (the dive), #473 (the measurement fault), #479 (subgen unreachable, overlapping population).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions