Drop your messy marketing performance dataset into Claude. Get back a single-sentence headline, three findings ranked by decision-relevance, and one recommendation per finding. Refuses to declare trends from too few data points. Calls out small-N and seasonal noise explicitly.
You are a skeptical analyst. Data: {{DATA}}. Period: {{PERIOD}}. Goals: {{GOALS}}. Produce a readout with: 1. Headline, one sentence, no hedging 2. Three findings ranked by decision-relevance 3. For each finding: what the data actually shows vs. how it's often misread 4. Two follow-up questions the data CAN'T answer yet 5. One recommendation per finding Refuse to declare a trend from <2 data points. Call out small-N, selection bias, and seasonal noise explicitly.
Sample run against a fictional SaaS company's monthly trial-conversion data. Real first-pass Claude output, not cherry-picked.
**Headline** Trial-to-paid conversion improved from 8.2% → 9.4% in March, but the gain was driven entirely by one channel, and it's not the one we've been investing in. **Three findings (ranked by decision-relevance)** **1. Direct/branded traffic converted at 14.1% (up from 11.3%), and it's not random.** - What the data shows: Direct traffic conversion improved across all 3 weeks of March. Same period, branded search impressions up 23%. - How it's often misread: "More direct = more brand awareness, ride the wave." That's only half right. The lift correlates with the launch of our new pricing page (Mar 11), visitors arriving direct AND landing on /pricing convert at 19%, vs 8% on the homepage. The pricing page is the conversion lever, not brand strength. **2. Paid social conversion declined to 4.2% (from 6.1%).** - What the data shows: Meta CPA up 38%, conversion rate down. ROAS still positive but trending wrong. - How it's often misread: "Creative fatigue, refresh ads." That's a possible explanation, not the only one. The decline correlates with iOS 18.4 (released Mar 4), which expanded App Tracking Transparency to Safari. Could be measurement artifact, not real performance loss. Need a holdout test before declaring a creative problem. **3. Email-driven conversions held steady (12.3%) but volume dropped 18%.** - What the data shows: Per-email conversion stable, but list growth flat for 2 months and unsubscribes up. - How it's often misread: "Email is fine." It's not: the rate is fine, the system is bleeding. We're losing list members faster than we're acquiring them. **Two follow-up questions the data CAN'T answer yet** 1. Of the trial-to-paid conversions in March, how many of them are existing customers from our consulting practice (i.e., warm leads, not new acquisition)? 2. Is the pricing page lift sustained or a novelty bump? Need 60 more days of data. **Recommendations** - Per Finding 1: Audit homepage → pricing page funnel. If 80% of conversions go via pricing, restructure homepage to push there faster. - Per Finding 2: Run a Meta holdout test (10% of geo). Don't refresh creative until we know if the problem is real or measurement. - Per Finding 3: Audit unsubscribe reasons (next 30 days). If "too frequent" >40%, cut send cadence; if "not relevant," cut a segment.
5,000 rows comfortably. Above that, Claude's analysis quality degrades because the context window starts compressing. For larger datasets, filter to the relevant date range first, or chunk by channel.
Yes. Add "if you need Python to compute X, write the code and explain what it produces" and Claude will both write the code and run the analysis. Useful for cohort math or anything requiring date arithmetic.
GA4 surfaces statistical anomalies. This prompt surfaces decision-relevant findings, which is a different (smaller, more useful) subset. GA4 will tell you "Channel X is up 15%." This prompt will tell you "Channel X is up 15% but the lift is one cohort and probably won't repeat."
Mostly. The structure ("headline, three findings, refuse to declare trends from thin data") works for any decision-oriented analysis. The {{GOALS}} input is the load-bearing variable. Without a real goal, the readout drifts into descriptive analytics.