The Cohort Retention Analyzer: Find Where Users Drop Off and Why
Why this prompt matters
Diagnosing a retention dip the ad-hoc way — pulling SQL, building a slide, guessing at causes in a meeting — typically takes a product team 2-3 weeks. By the time the root cause surfaces, the cohort that triggered the alarm has often already fully churned, and the next 2-3 cohorts have gone through the same broken experience unnoticed. A structured retention prompt compresses that diagnostic cycle from weeks to minutes, catching a broken onboarding flow or a bad acquisition channel before it silently degrades the next several thousand signups.
What we use it for
You're a growth PM at a 40-person B2B SaaS company. In Monday's leadership meeting, your CEO asks why the April cohort of new signups is retaining 12 percentage points worse than March's cohort at the Week 4 mark, and you have 20 minutes before your next meeting to come back with an answer that isn't 'we're not sure yet.'
Prompt
Act as a senior product analytics consultant who specializes in retention and cohort analysis for SaaS and consumer products. CONTEXT: [PASTE YOUR COHORT RETENTION DATA HERE — a table of cohort start month/week vs. % of users retained at Week 1, Week 4, and Week 12 or equivalent intervals] Product name: [YOUR PRODUCT NAME] Target user segment: [E.G., SMB FINANCE TEAMS, INDIVIDUAL DEVELOPERS, ENTERPRISE IT ADMINS] Known changes during this period: [LIST ANY PRODUCT CHANGES, ONBOARDING REDESIGNS, PRICING CHANGES, OR MARKETING CAMPAIGNS THAT OVERLAP WITH THE COHORT WINDOWS] TASK: Analyze the retention curves across cohorts to identify: 1. Which cohorts retain significantly better or worse than the others, and by how many percentage points 2. The specific week or month where drop-off accelerates most sharply across cohorts 3. Plausible causes tied to the product changes, onboarding friction, seasonality, or marketing source quality provided above 4. A ranked list of hypotheses for the single biggest retention risk, each with a confidence level and a concrete way to validate it CONSTRAINTS: - Do not just describe the data back to me — provide causal hypotheses grounded specifically in the product changes and context I gave you - Rank hypotheses by likely business impact, not alphabetically or by how confident you are - Flag any data quality issues you notice (e.g., sample size too small to trust, non-monotonic retention that suggests a tracking bug, or missing cohorts) - Limit the ranked hypotheses list to a maximum of 5 — do not pad with low-confidence guesses just to fill space - If the data doesn't support a strong conclusion, say so explicitly rather than forcing a narrative OUTPUT FORMAT: 1. Headline Finding — one to two sentences stating the single most important retention insight 2. Cohort Comparison Table — cohorts ranked best-to-worst with retention % at each measured interval 3. Drop-off Point — where retention falls off most sharply, and by how many percentage points 4. Ranked Hypotheses (up to 5) — cause, confidence level (High/Medium/Low), and how to validate it 5. Recommended Next Action — the single most valuable thing to test or fix first, and why it's the highest-leverage move
Result
Headline Finding
The April cohort's Week 4 retention (58%) is 12 points below March's (70%), and nearly the entire gap opens in the first 7 days — this looks like an onboarding regression, not a slow organic decline.
Cohort Comparison Table
| Cohort | Week 1 | Week 4 | Week 12 |
|---|---|---|---|
| February | 81% | 68% | 52% |
| March | 84% | 70% | 54% |
| April | 69% | 58% | — |
April's Week 1 retention alone is 12-15 points below both prior cohorts, and that gap persists rather than narrowing by Week 4 — the damage is happening immediately after signup, not accumulating gradually.
Drop-off Point
The steepest break is Days 1-3 post-signup. February and March cohorts lose about 6-8 points of retention in this window; April loses 19 points. Every cohort loses users in the first three days — April is losing roughly three times as many.
Ranked Hypotheses
1. Onboarding flow change shipped March 28 (Confidence: High). The new 5-step guided setup replaced a 2-step quick-start exactly 3 days before the April cohort's signup window opens. Early-days drop-off cliffs immediately after an onboarding redesign are one of the most common and well-documented causes of this exact pattern. Validate by: pulling funnel completion rates for the new 5-step flow — if completion rate is below 60%, this is very likely the primary cause.
2. Paid channel mix shift (Confidence: Medium). Marketing notes indicate an increased spend shift toward a new paid social channel starting April 2. Lower-intent traffic from a new acquisition channel commonly produces exactly this kind of front-loaded drop-off. Validate by: segmenting April's Week 1 retention by acquisition source — if the new channel's cohort retains dramatically worse than organic/existing channels, this compounds with hypothesis 1 rather than replacing it.
3. Pricing page A/B test bleed-through (Confidence: Low). A pricing experiment launched April 10 for a different user segment; possible but unlikely cross-contamination into this cohort given the targeting rules described. Validate by: confirming experiment targeting logs excluded new signups during this window.
Data Quality Note
April's Week 12 figure is missing because the cohort hasn't reached that age yet — do not treat this as a data gap requiring investigation.
Recommended Next Action
Pull the new 5-step onboarding funnel's step-by-step completion data before your next meeting. If there's a specific step with an unusually high abandonment rate, you'll likely have a concrete, fixable culprit — not just a correlation — within the hour.
Retention dashboards are everywhere in product analytics tools, but a dashboard only shows you that a number moved — it never tells you why. Most teams respond to a bad retention curve by pulling the data into a spreadsheet, staring at it, and speculating out loud in a meeting. That process is slow, unstructured, and heavily biased toward whichever theory the most senior person in the room happens to favor first.
This prompt is designed to replace that guessing exercise with a structured causal analysis, using the same reasoning steps an experienced growth analyst would actually walk through — but compressed into a single request.
Why the prompt is built this way
The Role instruction — "senior product analytics consultant" — matters more than it looks. Without it, general-purpose models tend to default to describing the data ("Cohort A retained better than Cohort B") rather than diagnosing it. Framing the model as a consultant who has to justify recommendations to a client pushes it toward causal reasoning instead of restating numbers you already have in front of you.
The Context section deliberately asks for more than just the retention numbers. Including known product changes, onboarding redesigns, and marketing campaigns during the same window is what turns this from a data-description exercise into an actual diagnostic one — the model can only propose grounded hypotheses if it has something to ground them in. A retention prompt fed numbers alone will produce generic explanations ("users may be losing interest") that are true of every SaaS product and useless for taking action.
The Constraints block does three specific jobs. First, it forces ranking by business impact rather than alphabetical or confidence order, which matters because the highest-confidence hypothesis is not always the highest-impact one to fix. Second, it caps the hypothesis list at five — without this, models tend to over-generate low-value guesses to seem thorough, which buries the one or two ideas actually worth acting on. Third, the explicit instruction to flag data quality issues catches a failure mode specific to retention data: non-monotonic curves (where week 4 retention is higher than week 1, which is mathematically impossible for a proper retention curve) almost always indicate a tracking bug rather than a real behavioral pattern, and an analyst who doesn't check for this will build a whole narrative on broken data.
What good output looks like
A useful run of this prompt should produce something you could paste directly into a Slack message before a leadership meeting: one headline finding stated as a claim (not a description), a small comparison table, the specific point where the curve breaks, and — critically — a next action rather than just more analysis. If the model's output reads like a longer version of the input data, the prompt has failed and probably needs a stronger constraint on avoiding restatement.
Where this pattern extends beyond retention
The same Role + grounded Context + ranked-hypothesis Constraints structure works for any "why did this metric move" problem — churn spikes, conversion rate drops, support ticket volume surges. The specific fields change, but the shape — force causal reasoning, ground it in known events, cap the hypothesis count, flag data quality — is reusable any time you're handing a model a time series and asking "why."