Agency retention intelligence packet

Retention Intelligence Packet

Operating question: What should the agency tell this client about retention risk and next actions?

Run this as an evidence-first operating play. Gather the evidence below, separate exact / estimated / partial / unavailable data, apply the decision rules and vetoes, and never invent missing numbers. Be adversarial toward platform self-reported figures — trust commerce truth.

Evidence to gather

• Commerce (Shopify/Woo/BigCommerce/etc.) — orders with customer identifiers, ≥12 months. Derive: repeat-purchase rate (share of customers with ≥2 orders) and its trend, % of revenue from returning customers, AOV new vs. returning, and cohort retention — for each acquisition month, the % of that cohort that ordered again by day 30 / 60 / 90.
• Klaviyo — per-flow performance for the core lifecycle (Welcome, Browse/Site Abandon, Abandoned Cart, Post-Purchase / Cross-sell, Winback): live status, open/click/placed-order rate, attributed revenue. Plus list growth (net new subscribers minus suppressions), deliverability (open-rate trend, spam-complaint rate, unsub rate, sending-domain health), and email's share of total revenue.
• Attentive (or SMS platform) — subscriber growth, opt-out rate, journey/automation performance, and SMS share of total revenue, where connected.
• Targets / context — the client's own retention goal or LTV/payback target if one exists; otherwise the category benchmark you'll judge against (state which).

Optional, if available:

• Promotion calendar — a sitewide sale spikes repeat-rate and returning-customer revenue, then borrows from the next window. Cohorts straddling a Black Friday read hot for the wrong reason.
• Recent flow/template changes — a flow paused or rebuilt mid-window explains a "decay" that is really an edit.
• Acquisition-mix shift — a surge of discount-led or marketplace first orders lowers repeat-rate without anything being broken; it's a different customer.
• Subscription / replenishment products — change the entire baseline; a 55% repeat-rate is unremarkable for coffee, excellent for mattresses.
• Known deliverability incidents — a domain reputation dip or a Klaviyo dedicated-IP warmup in progress.

How to decide — in order

1. Gate on data sufficiency first. Before reading any trend, check sample and history. Cohorts need ≥12 months of orders and ≥~50 customers per cohort to be stable; a flow needs ≥~200 recipients in the window to judge its rate; a list segment under ~50 profiles is anecdote, not signal. Anything that fails the gate is FIX (data/access gap) and is reported as "can't see this yet," never guessed.
2. Establish the baseline before the trend. Is this client's absolute repeat-purchase rate healthy for its category (most non-replenishment DTC lands ~20–30% repeat; replenishment 40%+)? A flat 18% repeat-rate is a bigger story than a 2-point wobble on a healthy 35%.
3. Read the trend against a clean window. Compare the trailing 90 days to the prior 90 (and year-over-year if seasonal). Strip promo-distorted weeks before calling a move real. A change inside roughly ±3 points / ±10% relative on a normal-sample metric is noise — WATCH, don't raise.
4. Localize the leak — acquisition, first-repeat, or loyal. Decompose: is returning-customer revenue share falling because fewer customers come back (a first-to-second-purchase problem → Post-Purchase flow + winback), or because the same returners spend less (an offer/merch problem)? Cohort decay tells you when customers drop; flow health tells you whether the safety net is even deployed.
5. Check the safety net is actually on. Map each core flow: live or off, sending or stalled, converting or not. A dark Post-Purchase or Winback flow is the single most common, most fixable retention leak — and it's an agency win, not a client failing.
6. Gate on deliverability before blaming content. If open-rate has stepped down, spam complaints are above ~0.3%, or the sending domain is unauthenticated, that is the retention story, not "the emails need a refresh." Deliverability problems masquerade as content problems.
7. Translate to client-safe, then triage to raise vs. monitor. Convert each surviving signal into plain, non-alarming language with the number, source, window, and confidence behind it. Raise only the two or three that clear the Decision Rules; everything else is monitored quietly. End with one recommended next play.

The prompt to run

You are my retention analyst preparing a client-meeting intelligence packet for an
ecommerce agency account manager. I run retention for this client across their store,
Klaviyo (email), and Attentive (SMS).

GOAL: tell me what to say to THIS client about retention health — which signals are real
enough to raise in the meeting, which to monitor quietly, and the single next play to
recommend. The output must be client-safe: plain, non-alarming, no internal jargon.

I will paste: commerce repeat-purchase rate + trend, % revenue from returning customers,
30/60/90-day cohort retention, core Klaviyo flow performance + status, list growth,
deliverability (open trend / spam rate / unsub rate / domain auth), email & SMS revenue
share, Attentive subscriber growth + opt-out rate, the client's category, and any
promo/flow-change context. Some data will be missing.

PRE-FLIGHT: First list which required inputs I provided vs. missing. If this client's
commerce + Klaviyo/Attentive retention data over a sufficient window (>=12 months of
orders for cohorts, enough recipients/profiles to clear sample) is missing, STOP and
return only (a) what's missing and (b) how to get it — never estimate it or alarm the
client; flag it as blind.

RULES:
- Data-sufficiency gate FIRST. Cohorts need >=12 months and ~50+ customers each; flows need
  ~200+ recipients; segments under ~50 profiles are anecdote. Anything below sample is FIX
  (a "can't see this yet" gap) reported plainly, NEVER guessed or smoothed over.
- Judge the absolute baseline against the stated category benchmark BEFORE the trend.
- Compare trailing 90d vs prior 90d on a promo-clean window. A move within ~+/-3 points or
  ~10% relative on a normal sample is noise -> WATCH, do not raise it with the client.
- Localize the leak: fewer returners (first-to-second-purchase / Post-Purchase / Winback)
  vs. same returners spending less (offer/merch). Use cohort decay for WHEN, flow health
  for WHETHER the safety net is deployed.
- Check deliverability before calling anything a content problem.
- Every row must carry: a number, its source, the time window, and a confidence level.
- For each raised signal, write a client-safe line: plain English, no internal metric names
  exposed raw, no alarm. Where access is missing, say so explicitly instead of guessing.

RETURN:
1. A one-paragraph client-facing summary (the AM could read it aloud).
2. A status table using exactly this header row:
   | Signal | Reading | Evidence | Confidence | Client-safe line | Next action |
   Use "—" for any cell you cannot fill. Do not add or drop columns, and do not replace
   the table with prose.
3. The 2-3 signals worth raising in the meeting, and why each cleared the bar.
4. What's being monitored quietly (and what would promote it to a "raise").
5. The single recommended next play, with owner and timing.
6. What evidence is blocked/low-sample and what you'd need to upgrade it.

Decision rules

• Raise it (FIX / REFRESH / KILL as fits) — only when the signal clears sample (passes the data-sufficiency gate) and the move is material: a baseline meaningfully below category, or a trend beyond ~±3 points / ~10% relative on a clean, promo-free window, and you can attach a number, source, time window, and confidence level to it. If you can't defend all four, it doesn't go in the meeting.
• FIX — a core flow is off/stalled, deliverability has degraded (open-rate step-down, spam complaints >~0.3%, unauthenticated domain), or the commerce↔Klaviyo sync is broken. These are concrete, agency-ownable, and usually the highest-leverage thing to raise.
• REFRESH — a live flow or program is decaying but still credible: Post-Purchase converting below benchmark, a Winback that's gone stale, returning-customer AOV slipping while traffic holds.
• WATCH — directional or early: a 1–2 point move inside noise, a young cohort, a segment near the sample floor, or a window polluted by a promo. Monitor quietly; do not raise.
• KEEP — repeat-rate and returning-revenue share inside the category band, core flows live and converting, deliverability clean. Say so plainly — a confident "retention is healthy, here's the proof" is a deliverable.
• Every recommendation must include a number, source, time window, and confidence level.

Vetoes — stop if any apply

• Never alarm a client on weak or low-sample data. A scary-looking move on a 40-customer cohort or a 150-recipient flow is not a finding — it's a data gap. Under sample, the only honest move is "we need more weeks."
• Don't expose raw internal metrics without framing. "Your day-60 cohort decay is 71%" is not a client sentence. Translate every number into plain, non-alarming language; keep the raw figure in the evidence column, not the talk track.
• Where data access is missing, say so explicitly — never paper over it with a guess. "We can't yet see SMS attribution because Attentive isn't connected" beats an invented number every time.
• Don't call a promo-distorted window a trend. A repeat-rate spike or dip that straddles a sale is borrowed demand, not retention movement.
• Don't attribute causality to a single platform metric. A Klaviyo open-rate dip can be Apple MPP inflation unwinding, not disengagement. Corroborate before you conclude.
• Don't blame content when deliverability is the cause — and vice versa.
• Don't recommend any write (pausing/launching a flow, sending a campaign, editing segments) without an explicit human approval step.

Output

A client action packet: a one-paragraph client-facing summary, then a status table, then the 2–3 to raise and the recommended next play.

Status table columns:

Signal	Reading	Evidence	Confidence	Client-safe line	Next action
Post-Purchase flow	Off	Klaviyo flow status, 90d	High	"There's an easy win we'd like to switch on for repeat buyers."	FIX — launch flow (Retention Lead, this sprint)

---

Run this on your live store data in one prompt — start a free 14-day trial of ShopMCP (no credit card). It connects your store, ads, analytics, email/SMS, and finance and runs this play read-only by default, applying changes only on your approval.