Retention loops

retention-loops

When to use

• D30 retention is flat or declining and the team cannot name a single product loop that pulls the user back — retention is treated as marketing's problem, not the product's.
• A new feature shipped but did not move retention — there is no closed loop between trigger, action, and reward, so the feature is a destination, not a habit.
• The product depends on a network effect that has not been instrumented as a loop — invites, content, or data are produced but the loop that pulls the next user back is unwritten.

Do NOT use to fix days 0–30 onboarding friction (route to onboarding-design), classify churn causes (route to churn-prevention), or design human-led account-expansion plays (route to expansion-playbook).

Cognition cluster

• Mental model 14 — Meadows leverage points. A retention loop is a feedback structure: the leverage sits in the loop's gain (how strong the reward is) and delay (how long until the reward lands), not in the surface UI. Pick the leverage point — gain or delay — over surface polish. See docs/contracts/mental-models.md § 14.
• Mental model 8 — Compounding. A loop with even small gain per cycle compounds across cohorts; a one-time activation bump does not. Verify which loops compound before investing cycles into them. See mental-models.md § 8.
• Mental model 18 — Pull vs. push. A trigger the user pulls (intrinsic need surfaced by the product) compounds; a trigger the vendor pushes (marketing notification firing) decays the channel and trains the user to mute. See mental-models.md § 18.
• Context-spine — product + customer-segment + funnel-stage. Read the product slot for which capability can carry a loop (a loop is only as strong as the action it routes through), the customer-segment slot for which segments have the latent need the loop addresses, and the funnel-stage slot for where the loop sits relative to activation and paid. See context-spine.

Procedure

Step 0: Inspect — name the current loops, if any

Inspect the product. For each suspected loop, write the closed form: "<trigger> → <action> → <reward> → <trigger again>." If the loop cannot be written closed, it is not a loop; it is a funnel ending. Inspect whether the reward arrives quickly enough to reinforce the action — verify the delay against the segment's attention cycle.

Step 1: Classify each loop as single-user vs network

1. Single-user loop — trigger and reward both originate from the same user (a daily-summary email triggered by yesterday's activity).
2. Network loop — trigger or reward involves another user (a teammate's comment, a partner's reply, a customer's reaction).

Network loops compound harder but require minimum-viable-network density; below density they look broken. Classify before investing.

Step 2: Audit the gain and the delay per loop

For each loop:

1. Gain per cycle — what observable utility does the user receive (information, social affirmation, time saved, reduced error)? Gain measured as the user's revealed willingness to repeat the action.
2. Delay — time from trigger to reward. A delay longer than the segment's attention window kills the loop regardless of gain.
3. Decay — does the loop weaken when the user already has the reward? Most product loops decay; design the next loop before the first decays.

Step 3: Pick the binding loop and isolate it

Of the loops named, pick the one whose gain × frequency × eligible segment-size is largest. Verify the loop is intrinsic-pull, not vendor-push: confirm the trigger originates from a user action or state, not from a marketing schedule. A push-trigger labelled as a loop will burn the channel.

Step 4: Design the missing step, not the missing UI

If the binding loop is broken, the broken step is almost always: trigger missing, action too far from trigger, reward delayed, or no path back to next trigger. Design the missing step, not a UI tweak. UI tweaks polish a loop that already closes; they do not close an open one.

Step 5: Hand back

Hand the loop inventory, the binding-loop selection with gain / delay / decay, and the step-level redesign to the implementing team and to activation-design — activation is the loop's first cycle, and the activation event must complete the first cycle of the binding loop. Retention work without a named loop is rearranging notifications.

Related Skills

WHEN to use this

• Designing or auditing product-led retention loops.
• Selecting the binding loop and redesigning its missing step.

WHEN NOT to use this

• Days 0–30 onboarding milestones — route to onboarding-design.
• Cause-classification of churn events — route to churn-prevention.
• Human-led expansion plays — route to expansion-playbook.
• Activation-event selection (first cycle of the binding loop) — route to activation-design.

When the agent should load this

• "Why don't users come back?"
• "Design a habit loop for feature X."
• "Is this loop single-user or network?"
• "Welcher Loop tr\u00e4gt eigentlich unsere Retention?"

Output

1. loop-inventory.md — every named loop in closed form: trigger → action → reward → next trigger, with single-user vs network tag.
2. gain-delay-audit.md — per-loop gain · delay · decay · eligible-segment size · revealed repeat-rate.
3. binding-loop-redesign.md — selected loop, the broken step, and the redesign in step terms (not UI terms).

Gotcha

• A loop whose reward arrives outside the segment's attention window will look broken even when gain is high; delay kills loops more often than gain does.
• A network loop below minimum-viable density behaves like an open funnel; instrumenting it and designing it before density is theatre.
• "Notifications fire daily" is not a loop; it is a push schedule. A loop needs a closed return path from reward to next trigger that the user — not the vendor — closes.

Do NOT

• Do NOT invest in surface UI on a loop that does not close; the loop closes by adding a step, not polishing one.
• Do NOT instrument network loops as single-user loops; the metric will look broken until the network reaches density.
• Do NOT design more than one binding loop at a time; concurrent loop changes destroy the signal.

Runnable example

Mid-market collaboration tool, D30 retention 41 %, two suspected loops named.

• Loop inventory — (L1) user receives daily summary → opens product → reviews changes → leaves a comment → teammate notified (network). (L2) user creates a doc → bookmark surfaces in nav → user reopens (single-user).
• Gain–delay audit — L1 gain medium, delay 24 h (within attention window), decay low (network refreshes); L2 gain low, delay 0, decay high (bookmark stale within a week).
• Binding loop — L1 selected (gain × frequency × segment-size dominates). Broken step: "teammate notified" fires but does not route teammate back to the originating doc — the loop opens.
• Redesign — add teammate-return path: notification deep-links into the doc at the commented passage; verify with cohort A/B at 4-week horizon. Predicted: D30 +6 pp ± 3 pp.
• Hand-off — loop inventory + redesign → eng team; activation event redefinition (one comment + one teammate notified) handed to activation-design.