TL;DR
Paste your cohort grid plus three things the model can’t see — cohort sizes, dated events, and the exact question leadership asked — and ask for exactly 3 sentences: early-window direction, long-tail shape, and one outlier. Force specific number ranges, flag anything inside cohort-to-cohort noise, and report direction across the trailing 6 cohorts rather than first-vs-last. The prompt below does this. It runs equally well on ChatGPT (GPT-5.5), Claude (Sonnet 4.6 / Opus 4.7), or Gemini 3.1 Pro — none of them see your raw data, so accurate transcription matters more than which model you pick.
The task
Mixpanel (or Amplitude, or your warehouse) is showing the retention cohort table: 12 weekly cohorts down, 12 weeks across, color-coded red to dark green. Your CEO walks past your screen and says “so, is retention getting better?” You have 90 seconds before they ask again. The chart genuinely looks “kind of better in some columns and weirdly bad in row 4,” and you do not want to give the wrong directional answer to a question that defines next quarter’s roadmap. You want a 3-sentence readout that says what is actually happening, plus one outlier worth investigating before someone builds a strategy on it.
Where AI helps, and where it does not
AI is good at pattern recognition across the rows and columns you describe: direction of week-1 retention across cohorts, whether the curve flattens at the same week or shifts, which cohort is genuinely an outlier versus statistical noise. It is also good at producing the disciplined 3-sentence format that survives an exec’s glance, where most analysts default to a 10-bullet wall.
What AI cannot do: see your raw data, so you must transcribe the table accurately, ideally with weeks in the column header. It cannot diagnose root causes either; correlation with a feature ship date or pricing change is a hypothesis, not a finding. And it cannot tell you whether a 3-point W1 lift is statistically significant. Small cohorts are noisy: in a 50-user cohort, 5 extra returning users move retention by 10 points; the same 5 users move a 1,000-user cohort by 0.5 points. The model will confidently call a 24% → 27% swing “improvement” when it might be coin flips.
A specific failure mode: AI defaults to reporting “improvement” whenever the most recent cohort beats the oldest cohort, ignoring the noisy middle. Tell it explicitly: “report direction across the trailing 6 cohorts, not first vs. last; and call out cohort-to-cohort noise where it exists.”
The 2026 wrinkle: AI tourists and the M3 rebase
If your sign-up flow got more frictionless in the last year, your early cohorts are inflated by what a16z calls AI tourists: people who sign up out of curiosity, poke around once, and never come back. They make healthy products look weaker because they bloat the denominator. The fix a16z popularized is to rebase the curve from M3 instead of M0 (or W4 instead of W0 for weekly cohorts): once tourist churn has cleared, the survivors made a real decision to stay, and M12/M3 is a cleaner early signal of long-term retention than M12/M0. Tell the model which baseline you want; the default (everything over the W0 acquisition number) is the number that’s quietly lying to you.
What to feed the AI
- The cohort table values (paste as a markdown grid, or describe row by row; the model is fine with either)
- The cohort axis: weekly, monthly, or by signup source (this changes what counts as a real cohort)
- Which metric the cells represent (D1, W1, M1, or “any-action retention”; the model needs to know)
- The cohort sizes: a 47% number on a 30-user cohort is not the same signal as on a 3,000-user cohort
- Dated events that might explain shifts: feature ships, pricing changes, ad campaigns, launches, infra changes
- The question leadership actually asked, in their words (“is retention getting better” vs. “why is the AppSumo cohort still alive” are completely different readouts)
- Any seasonality you already know about (holiday, end-of-quarter, school year)
- The baseline you want: standard (over W0/M0) or rebased (over W4/M3) to strip tourist churn
- The acceptable confidence level: “directional is fine” vs. “I need to put this in the board deck”
Copy-ready prompt
Read this retention cohort table and write a 3-sentence readout for [leadership / PM team / board].
Cohort axis: [weekly / monthly / by signup source]
Metric in cells: [D1 / W1 / M1 / any-action retention]
Baseline: [standard, over W0/M0 / rebased, over W4/M3 to strip tourist churn]
Cohort sizes: [paste or describe]
Table values (cohorts as rows, weeks as columns):
[paste markdown grid]
Dated events (with dates): [paste]
Known seasonality: [paste or "none"]
The exact question leadership asked: "[quote]"
Return exactly 3 sentences:
1) Direction of early-window retention (W1 or M1) across the trailing 6 cohorts. Name the specific number range, not "improving." Flag if cohort-to-cohort noise is high relative to the trend.
2) Long-tail shape: does the curve flatten, and at which week? Compare the W1-to-W8 (or M1-to-M6) gap across cohorts: widening or narrowing?
3) The single most interesting outlier cohort: name the cohort, the number, and the most likely event explanation. Be explicit that this is hypothesis, not finding.
End with one line: "Next chart I would pull: [specific chart]" — the chart that would test the hypothesis from sentence 3.
Do not call a < 3-point swing on a < 200-user cohort "improvement"; flag it as within noise.Shorter variant: single-line Slack answer
Cohort table: [paste]. Leadership asked: "[quote]". Write a one-line answer with the specific number and direction, no caveats. Then a second line with the one cohort I should actually look at this week.Which model, and what it costs
All three frontier models read a pasted grid well; the readout quality difference is small next to transcription accuracy. As of June 2026:
| Model | Best for this task | Context (in-app) | Plan to use |
|---|---|---|---|
| GPT-5.5 (ChatGPT) | Fast 3-sentence readouts, Slack one-liners | ~320 pages on Plus; full 1M only on $200 Pro | Plus $20/mo |
| Claude Sonnet 4.6 / Opus 4.7 | Careful noise/significance reasoning, big grids | 1M tokens standard | Pro $20/mo (Opus on Max) |
| Gemini 3.1 Pro | Tables already living in Google Sheets/Workspace | 1M tokens | Google AI Pro $19.99/mo |
For a single 12 × 12 grid, any free tier handles it. The paid tiers matter only if you paste many cohorts at once or attach the raw export. Whatever you pick, the model is reading your transcription, not your dashboard, so the upstream accuracy is where the readout lives or dies.
Sample output
A useful 3-sentence readout: “Week-1 retention has lifted from 24-27% to 30-33% across the last 6 cohorts, directional improvement with cohort-to-cohort noise of about 3 points. All cohorts flatten by week 4 and the W1-to-W8 gap has not narrowed; what we’ve fixed is the cold start, not the long-term hold. The Aug 15 cohort shows 47% W1 on 412 users, which lines up with the AppSumo deal, and these users are still over-represented in week-8 active users, suggesting deal-driven users are surviving longer than expected rather than overall retention shifting.”
A useful “next chart” line: “Next chart I would pull: per-cohort week-4 to week-8 retention only, to confirm the long-tail is genuinely flat across the recent improvement and not lagging.”
A useful Slack one-liner: “W1 retention is up ~6 points across the last 6 cohorts (24-27 → 30-33), but the long-tail looks the same; we fixed onboarding, not stickiness. Aug 15 cohort is the one to look at this week.”
How to refine
- Force specific numbers: “Replace any phrase like ‘recent cohorts have improved’ or ‘long-tail looks stable’ with the actual number range, e.g., ‘W1 moved from 24-27% to 30-33% across the last 6 cohorts.’ If you can’t be specific, the data doesn’t support the claim.”
- Make the long-tail check explicit: “Compute the W1-to-W8 gap (or M1-to-M6) per cohort. Tell me whether the gap widens or narrows over the trailing 6 cohorts; that is the long-tail finding.”
- Flag noise honestly: “If the trend swing is smaller than typical cohort-to-cohort noise, say so. ‘Directional improvement within noise’ is a valid finding; ‘improvement’ alone overclaims.”
- Tie outlier to event, not feature credit: “Name the outlier cohort with both its number and its likely event correlation. State explicitly that this is hypothesis. Do not claim feature X caused the lift unless we shipped only one thing in that window.”
- Match the readout to the audience: “For board, 3 sentences and one chart. For PMs, add the per-cohort table and the next-chart line. For the data team, return the markdown table with my interpretation underneath, not in place of, the data.”
Common mistakes
- Reading week-1 movement and stopping: the long-tail tells a different story and is what matters for LTV
- Comparing the oldest cohort to the newest: the noise in the middle is often the actual signal; trailing 6 is the standard window
- Ignoring outlier cohorts: they usually carry the most signal about what is actually driving acquisition mix
- Confusing cohort size with cohort quality: a 47% W1 on 30 users is not better than a 30% W1 on 3,000 users; small-n noise dominates
- Forgetting to rebase: if frictionless sign-up inflated your recent cohorts with tourists, the W0 baseline understates real retention; check M12/M3 too
- Cherry-picking the prettiest cohort for the deck: if leadership notices the surrounding cohorts later, you lose credibility on every readout that follows
- Claiming feature credit without isolation: if you shipped 3 things the same week as a retention lift, “associated with” beats “caused by”
- Calling a 2-point swing “improvement”: anything inside typical cohort-to-cohort noise is a coin flip, not a finding
- Pasting only the table without events or cohort sizes: the model pattern-matches the numbers but misses the actual story
FAQ
- Should I share the full cohort table with leadership?: For a technical audience (PMs, eng leads, data team), yes. For execs and the board, share the 3-sentence readout plus one visual (usually a small-multiples chart of the curve shapes, or one heat-map of the trailing 6 cohorts only). The full grid trains them to skim.
- How many cohorts before trends are real?: For weekly cohorts, 6+ before you trust direction, 12+ before you trust long-tail shape changes. For monthly, 3+ for direction, 6+ for long-tail. Below those thresholds, report as “directional, within noise.” Each cohort also wants hundreds of users minimum before percentages are stable.
- Why does the W1-to-W8 gap matter more than W1?: W1 reflects onboarding and first-day value. The W1-to-W8 ratio reflects whether the product creates a habit. You can move W1 with a better welcome email; moving W1-to-W8 requires actual product fit.
- What is the M3 rebase and do I need it?: Instead of measuring retention against M0 (everyone who signed up), measure against M3 (everyone still active after tourist churn clears). a16z showed M12/M3 is a cleaner early predictor of long-term retention for AI-native products with frictionless onboarding. If your sign-up flow got easier in the last year, run both baselines and compare.
- The model says retention improved but my gut says no. What’s going on?: Usually the model compared first vs. last cohort and missed the noisy middle. Add: “compare the trailing 6 cohorts against the prior 6 and call out variance, not just direction.”
- What if I have multiple acquisition sources mixed in?: Split the cohort by source before asking for a readout. A mixed-source table hides whatever is really happening; the underlying sources usually move in opposite directions and cancel.
Related
- AI funnel analysis readout
- AI chart takeaway
- Excel / Spreadsheet Analysis Prompts
- AI KPI narrative
- AI dashboard takeaway
Tags: #AI writing #Data analysis #Workflow #Retention #Cohort