Arccos Edge · product marketing ·
Built on users and 54 days of behaviour — including the seven weeks Snowplow dropped, recovered from Statsig. Rates are measured over a trailing 28-day window ().
Where should marketing spend attention? Bubble size is the addressable population; colour is the four-week trend in reach rate. The prize quadrant is low reach, high stickiness — proven value that not enough people have found.
Vertical axis is Sticky28 — of people whose first action was ≥28 days ago, the share who came back and acted in at least two separate weeks. Surfaces with fewer than 100 mature first-actors are not plotted. Exact values for every bubble are in the Opportunity table below.
Is the problem distribution or persuasion? Blue is daily users reached, orange is users who acted, both 7-day rolling. Falling reach with steady conversion is a navigation problem; steady reach with falling conversion is a positioning problem.
Full history, 4 Jun – 27 Jul. Player Home is a lower bound before 23 Jul: Statsig captures ~78% of its entry impressions. The bias is constant, so the shape holds.
Exactly where each surface loses people. A fat grey band is a discovery problem that email can fix. A fat blue band means they saw it and passed — that is copy, CTA or latency, and more awareness will not help.
Which surfaces earn a second visit? Weekly retention after a user's first deliberate action. High conversion with low retention means a good demo attached to a weak ongoing job — promote it at the right moment, not as a habit.
W1–W4 = share of the same mature cohort (first act ≥28 days before the window end) who acted again in each subsequent week — one denominator for all four points, so the decline is behaviour, not cohort mix. n is that cohort. Cohorts under 100 are suppressed rather than shown noisily.
Where would better discovery create the most additional actors? Components are shown so the ranking can be audited rather than trusted blindly.
Score = 100 · addressable.35 · conversion.35 · headroom.20 · decline.10, each a percentile rank, geometrically combined so raw volume cannot dominate. Conversion is empirical-Bayes shrunk. Potential actors asks: if marketing reached 10% of who is left, how many would act?
Who exists to talk to today. Audiences are priority-exclusive top to bottom — a user counts in the first journey that claims them, so the sizes can be summed and Braze cannot double-send. Audience lists live in Snowflake for Braze to pull — this page carries counts only.