Acquisition & Channel Monitor · Wisdom notebook
Ask the follow-up.
Agentic conversational BI · one thread across the channel mix — what broke, what works, and what the blended numbers hide.
The dealer channel is routed. Did anything else in the mix drop out of its own conversion band this month?
RKScored every acquisition channel's stage conversion against its own seasonal band day by day rather than monthly, then, for anything outside, split the same slice by device, plan, region and campaign to separate a market move from a fault.
One, and it lasted three days. The app's order-submit rate fell to 30% on 15 October against a 78% expectation, sat at 48% either side of it, and was back to normal by the 18th. The reason it is a fault and not a market move is the device split — 80% of it is one operating system:
App channel · order-submit rate, day by day
Orders reaching provisioning fell 60% on the same day and activations for the slice ran 65% below baseline — the downstream shape of a submit step that failed silently. A guest whose submit does nothing assumes the app is thinking, and leaves.
And the other way — which cohort is converting and staying best, and is any of it repeatable?
RKRan the same channel × market × plan scan for cohorts above their band rather than below it, then held each one to day ninety so the activation number and the retention number could be read together instead of separately.
The metro premium cohort, and it is ahead on both halves at once — which is the part that makes it worth copying rather than just worth reporting:
Metro premium cohort · November
88% of it is one plan and 72% one customer segment, led by the West. The 84% vintage concentration is the caveat worth carrying: this is one quarter's intake performing, so it is a pattern to test in the adjacent markets rather than a rule to plan on yet.
Put the whole mix side by side — which channel is failing on which measure?
RKCollapsed the month's scans into one grid: funnel conversion against each channel's own band, 90-day early-churn of the cohort it delivered, and what that cohort billed in its first ninety days.
No two of them fail the same way, which is why a single channel-performance number would have caught none of them. Three different failures and one model, on three different measures:
The channel mix · funnel, retention, revenue
A blank cell means the scan did not flag that channel on that measure, not that the number is good. Read across: the app broke for three days, the dealer channel lost to a competing offer, the comparison site bought subscribers who left, and one channel is ahead on both conversion and what the cohort bills. Only the third of those is invisible on the day it happens.