Retention & Loyalty Early-Warning · Wisdom notebook
Ask the follow-up.
Agentic conversational BI · one thread from the frequency signal to the cohorts losing guests, and the one keeping them.
Before the seller cohort — is anything moving in the loyalty base itself this month?
RKScored visit frequency, average basket and repeat rate for every loyalty tier against each tier's own seasonal baseline, then looked for tiers where the measures disagree — because a tier where frequency and basket fall together is a spend story, and a tier where only one of them falls is a churn story.
One tier, and the measures disagree in the way that matters. High-value members visited 11% less often than their seasonal baseline in March while the average basket held flat — they are not spending less when they come, they are coming less often. That is the earliest of the churn tells, and it is reachable:
High-value loyalty tier · March, against its own seasonal baseline
Frequency falling with basket flat is disengagement, not a budget cut — the members still buy the same shop, they just skip weeks. A tier that is still transacting is a tier a win-back can still reach; the same signal read after the basket falls is a report, not a warning.
And on the marketplace side — which seller cohorts are outside their own return-rate band, not the marketplace average?
RKCompared each cohort's marketplace return rate with the baseline for its own tenure and sub-category rather than with the marketplace average, because a post-holiday new-seller cohort and a mid-October apparel cohort do not share a normal.
Two, and only one of them is a seller problem. The new Q3 cohort in Small Appliances is at 31% against its own ~12% baseline. Marketplace apparel is at 13% against its own ~9% — a much smaller gap, and it arrived in the same week the trend-launch apparel stopped selling:
Marketplace return rate, each cohort against its own baseline
Read against the marketplace average both would look mildly elevated. Read against their own, one is nearly three times its normal and needs a cohort review; the other is four points over and is the return-side echo of a merchandising miss that has already been routed.
Is anything in the book holding guests rather than losing them — and does it look like the opposite of these?
RKRan the same guest-experience grid over the cohorts moving in the other direction, on the measures a guest actually feels: what comes back, whether the promise was met, and what the guest says afterwards.
Yes — and it is the exact inverse. Quiet-luxury owned-brand apparel returns about 30% less than the rest of the apparel book while selling 42% ahead of plan at close to full price. Guests are choosing it deliberately and keeping it:
Three cohorts, on what a guest actually feels
The two failing cohorts fail on the promise first and on sentiment second; the one that works never has to make a promise it might miss. Markdown on the quiet-luxury cohort runs about 28% under the rest of apparel, so none of the retention is being bought.