Product Analytics Interpretation

Overview

Product analytics is the backbone of informed product decision-making, but the volume and complexity of data most product teams deal with today has outpaced the capacity of any individual analyst or PM to interpret it manually with speed and confidence. Funnel reports, cohort tables, retention curves, and engagement dashboards each tell a fragment of a larger story — and stitching those fragments into clear, actionable narratives has historically required significant analytical experience and time.

AI changes this equation dramatically. With the right prompting approach, you can take raw exported data from tools like Amplitude, Mixpanel, Heap, or Google Analytics and receive structured interpretations in minutes: drop-off diagnoses, retention pattern summaries, anomaly flags, and recommended hypotheses for further investigation. The result is not that AI replaces analytical thinking — it is that AI compresses the time between data and insight, allowing experienced PMs and BAs to spend their energy on the highest-value activity: deciding what to do next.

This topic covers the full workflow of AI-assisted product analytics interpretation. You will learn how to structure your data inputs, which questions to ask, how to iterate from broad observation to specific hypothesis, and how to produce insight narratives that are ready to share with engineering, design, and business stakeholders. The prompting techniques here are designed for practitioners who already understand product analytics concepts and want to accelerate and deepen their analysis — not replace it.

The skill you will build is systematic: a repeatable process for turning any analytics dashboard export into a structured narrative with observations, patterns, hypotheses, and recommended actions. By the end of this topic, you will have prompt templates and a workflow you can apply to your own product data in your next sprint.

How to Use AI to Analyze Product Usage Data — Funnels, Cohorts, and Retention Curves

Funnel analysis, cohort analysis, and retention curve interpretation are three of the most powerful lenses in a product manager's analytical toolkit. Each surfaces a different dimension of user behavior, and each requires a slightly different approach when working with AI.

Funnel analysis reveals where users drop out of a defined sequence of steps — onboarding, checkout, activation, feature adoption. The raw data from a funnel report typically includes step names, the number of users entering each step, the conversion rate from one step to the next, and the overall funnel conversion rate. When you paste this into an AI prompt, you want it to do more than report numbers back to you — you want it to identify which drop-off points are most significant, reason about likely causes, and propose hypotheses to test.

The key to effective funnel analysis prompting is providing context alongside data. Tell the AI what the funnel represents, what a "healthy" conversion rate looks like for your product category, and what you already know about the user journey. A funnel with a 60% drop-off at step 3 means something very different in a consumer mobile app versus an enterprise SaaS onboarding flow. AI cannot infer context it is not given — your job as the practitioner is to supply it.

Cohort analysis tracks groups of users over time, typically segmented by the week or month they first performed a key action (signup, first purchase, first activation). Cohort tables are notorious for being data-dense and difficult to read at a glance. AI excels at translating a cohort table into a narrative: which cohorts retained better, whether retention is improving or degrading over time, and where the inflection points are. When inputting cohort data, include the cohort definition (what event defines the cohort), the retention metric being tracked, the time period covered, and whether you are looking at 7-day, 14-day, 30-day, or custom intervals.

Retention curve interpretation is about understanding the shape of the curve as much as the numbers. A retention curve that flattens quickly (high early churn, then stable) signals a product that works for a subset of users but fails to engage the majority. A curve that declines steadily indicates a fundamental engagement or value delivery problem. A curve with a sharp drop at a specific interval often signals a specific lifecycle event — such as a free trial expiring or a notification cadence ending. Sharing the curve data with AI and asking it to classify the curve shape and reason about its causes gives you a structured starting hypothesis.

Hands-On Steps

1. Export your funnel report from your analytics tool (Amplitude, Mixpanel, Heap, or similar) as a CSV or copy the table data directly from the dashboard.
2. Identify the context you need to supply: funnel name, the goal the funnel represents, the user segment (e.g., new users, paid users, mobile users), the time period, and any known context about recent product changes.
3. Open your AI tool and paste the funnel data with a structured prompt that includes the context and asks for: step-by-step drop-off interpretation, identification of the most critical drop-off point, likely causes, and hypotheses to test.
4. Review the AI output. Challenge any interpretation that does not match your product knowledge by responding with corrections or additional context.
5. For cohort analysis, export your cohort retention table (rows = cohorts by week/month, columns = time intervals after cohort start). Annotate the export with the cohort definition and retention event.
6. Paste the cohort table into AI with a prompt asking for: identification of retention trends across cohorts, the interval at which the steepest drop typically occurs, whether retention is improving or worsening across recent cohorts, and standout cohorts worth investigating further.
7. For retention curves, export or manually transcribe the retention rate at each time interval (Day 1, Day 7, Day 14, Day 30, Day 60, Day 90). Include the product category and user segment.
8. Prompt AI to classify the curve shape, identify inflection points, compare to typical benchmarks for your product category, and hypothesize causes for each major drop or plateau.
9. Compile the AI interpretations from all three analyses into a single summary and identify the common threads — recurring hypotheses across funnel, cohort, and retention data are the highest-confidence starting points for investigation.

Prompt Examples

Prompt:

I have a user onboarding funnel for a B2B SaaS project management tool. Here is the funnel data for new users in the last 30 days:

Step 1 - Account created: 1,240 users (100%)
Step 2 - First project created: 892 users (72%)
Step 3 - First task added: 601 users (48%)
Step 4 - First team member invited: 203 users (16%)
Step 5 - First message sent: 89 users (7%)

Industry benchmark for this type of tool is approximately 25-35% reaching Step 4 (team invite).

Context: We recently simplified the project creation flow 6 weeks ago. Users are primarily SMB customers, 2-50 employees.

Please:
1. Identify the most critical drop-off point and explain why you selected it
2. For each major drop-off, provide 2-3 likely causes
3. Prioritize which drop-off to investigate first and why
4. Suggest 2 hypotheses we could test to improve the most critical step

Expected output: A structured analysis naming Step 3→4 (task to invite) as the most critical drop-off given the benchmark context, with causes such as insufficient prompting to invite collaborators, value not yet perceived before inviting, or fear of billing implications. Prioritization reasoning and two testable hypotheses with predicted impact.

Prompt:

Here is a monthly cohort retention table for our mobile consumer app (habit tracking). The cohort event is "first habit logged." Retention event is "logged a habit in that month."

Cohort | M0 | M1 | M2 | M3 | M4 | M5
Jan    | 100% | 38% | 24% | 19% | 17% | 16%
Feb    | 100% | 35% | 21% | 17% | 16% | —
Mar    | 100% | 41% | 27% | 22% | —  | —
Apr    | 100% | 44% | 29% | —  | —  | —
May    | 100% | 42% | —  | —  | —  | —

We launched a streak feature in late February. Typical M1 retention for habit apps is 30-40%.

Please:
1. Identify whether there is a trend in retention improvement across cohorts
2. Assess whether the streak feature appears to have had an impact (and how confident you are in that assessment given the data limitations)
3. Identify the most critical retention interval and why
4. Suggest what additional segmentation would be most valuable to run next

Expected output: Analysis noting the upward trend from Jan→Apr cohorts in M1 and M2 retention, a cautious attribution hypothesis about the streak feature (correlation, not causation), identification of M0→M1 as the highest-leverage interval, and recommendations for segmentation by notification opt-in, platform (iOS vs Android), or number of habits logged in M0.

Prompt:

Here is the Day 0 through Day 90 retention curve for our e-commerce mobile app, new user cohort, last 6 months combined:

Day 0: 100%
Day 1: 42%
Day 3: 28%
Day 7: 18%
Day 14: 12%
Day 30: 8%
Day 60: 7%
Day 90: 7%

Product category: e-commerce app, average purchase frequency of 3x per month for retained users.

Please:
1. Classify the shape of this retention curve and what it typically indicates
2. Identify the inflection points — where the curve changes rate significantly
3. Diagnose what might be happening at the Day 1 and Day 7 drop-offs specifically
4. Assess whether the plateau at Day 60-90 is healthy or concerning for this product category
5. Suggest the top 3 product or communication interventions to test to improve early retention

Expected output: Classification as a "classic early-churn curve with stable core," inflection points at Day 1 and Day 7, diagnosis of likely causes (Day 1: no purchase intent triggered, Day 7: no re-engagement prompt or value reminder), assessment that a 7% plateau is within range for e-commerce given purchase frequency, and three ranked intervention recommendations such as a Day 2 re-engagement push notification, a first-purchase incentive prompt, and a personalized product discovery onboarding.

Learning Tip: Always include benchmark context in your funnel and retention prompts. AI cannot know what "good" looks like for your specific product category and user segment without you providing it. A 40% M1 retention rate means something very different for a daily habit app versus a quarterly tax filing tool — the same number can be excellent or alarming depending on context. Build a habit of stating the expected or benchmark range every time you prompt for analytics interpretation.

Generating Insight Narratives from Raw Analytics Dashboards with AI

Raw data and dashboards tell you what happened. Insight narratives tell you what it means and what to do about it. The gap between the two is where most product teams lose time — the data exists, but transforming it into a compelling, structured narrative that can drive decisions requires effort that often falls through the cracks in a fast-moving sprint cadence.

AI dramatically compresses this workflow. The process is straightforward: export your dashboard data, provide context about the product and business goals, and prompt AI to generate a structured insight narrative. The output — when you prompt well — is a narrative that follows a consistent format: observation, pattern, hypothesis, and recommended action. This format is powerful because it maps directly to the structure of a good product decision: here is what we see, here is what it suggests, here is what we think is happening, and here is what we propose to do about it.

The dashboard export to AI narrative workflow has one important prerequisite: you must be selective about what you export. Dumping an entire dashboard with dozens of metrics into a single prompt produces shallow, scattered analysis. Instead, identify the 3-5 metrics most relevant to the question you are trying to answer, export or transcribe those, and prompt for deep analysis of that focused set. Breadth is the enemy of depth in AI-assisted analytics.

When generating narratives for sharing with stakeholders, it is important to review and edit AI output before distributing it. AI will produce a structurally sound narrative, but it lacks knowledge of your product's specific history, recent events, or organizational context. Your role is to inject that knowledge — either by providing it in the prompt or by editing the AI output post-generation. The best AI-assisted insight narratives are always a collaboration between the AI's pattern recognition speed and the PM's contextual judgment.

The insight narrative format — observation, pattern, hypothesis, recommended action — is deliberately sequential. The observation is a factual statement about what the data shows. The pattern is the relationship or trend within the data. The hypothesis is the interpretive layer: what explains this pattern? The recommended action is what the team should do based on the hypothesis. Each step builds on the previous, and AI can generate all four layers if you prompt it to.

Hands-On Steps

1. Identify the product question you are answering (e.g., "Why did activation rate drop this week?" or "What explains the difference in retention between mobile and desktop users?").
2. Open your analytics dashboard and identify the 3-5 metrics most relevant to the question. Avoid the temptation to include everything.
3. Export or manually transcribe the data, including time series data where available (week-over-week or month-over-month is more useful than a single point in time).
4. Draft a one-sentence context statement: what product this is, who the users are, and what the business goal is for this metric set.
5. Paste data and context into AI with a prompt requesting a structured insight narrative using the observation-pattern-hypothesis-action format.
6. Review the output. Annotate any AI interpretation that needs correction based on your product knowledge (e.g., "The drop on March 15 was due to a planned infrastructure migration, not a product issue").
7. Revise the narrative by either re-prompting with corrections or manually editing the output.
8. Format the final narrative for the appropriate audience: a short version for Slack/Confluence updates, a detailed version for sprint reviews or steering committee reports.
9. Archive the narrative alongside the raw data in your team's documentation system for future reference and longitudinal comparison.

Prompt Examples

Prompt:

I need an insight narrative for a weekly product review. Here is the data:

Product: B2C fitness app (subscription, $9.99/month)
Time period: Last 4 weeks (Week 1 through Week 4)
Business goal: Grow monthly active users and reduce churn

Metric data:
- Weekly Active Users: 12,400 / 13,100 / 12,800 / 11,900
- New User Activations (completed first workout): 820 / 890 / 760 / 640
- 7-day retention of new users: 34% / 36% / 31% / 27%
- Subscription churn rate (weekly): 1.2% / 1.1% / 1.4% / 2.1%
- Average workouts per active user per week: 3.1 / 3.2 / 2.9 / 2.6

Context: We launched a new AI-personalized workout feature in Week 2. No major product bugs reported. Seasonality note: this time period is post-New Year's, and some drop-off in engagement is expected by Week 3-4 as resolution-driven signups churn.

Please generate a structured insight narrative using the format:
- Observation: [what the data shows factually]
- Pattern: [the relationship or trend visible]
- Hypothesis: [what explains this pattern]
- Recommended action: [what the team should do]

Generate one narrative for overall health, and one specific narrative for the churn rate spike in Week 4.

Expected output: Two structured narratives. The overall health narrative will observe the declining WAU and engagement trends, pattern-match to a post-resolution churn cycle, hypothesize that cohort mix is shifting toward lower-intent users, and recommend monitoring Week 5 data before intervention. The churn narrative will flag the Week 4 jump as above seasonal norm, hypothesize a potential mismatch between the AI personalization feature and user expectations or a cold-start problem, and recommend a churn cohort analysis segmented by whether users engaged with the new feature.

Prompt:

Convert the following dashboard data into an executive-ready insight narrative (3-4 sentences, no bullet points, suitable for a VP-level email update):

- DAU: down 8% week-over-week
- Session length: up 12% week-over-week
- Feature adoption (new search feature): 23% of DAU engaged with it
- NPS this week: 42 (up from 38 last week)

Product context: SaaS document management tool, enterprise customers. The new search feature launched last Monday.

The narrative should be positive in tone where data supports it, honest about the DAU decline, and connect the data points into a coherent story. End with a clear next step.

Expected output: A polished 3-4 sentence narrative such as: "Early signals from last week's search feature launch are encouraging — 23% of daily active users engaged with the new functionality within its first week, contributing to a 12% increase in session length and a 4-point improvement in NPS to 42. Daily active users were down 8% week-over-week, which we are monitoring closely; this dip aligns with the holiday pattern observed in the same week last year and does not yet indicate a structural change. The team is tracking whether the users engaging with search show differentiated retention in the coming two weeks, which will be the clearest signal of the feature's long-term value."

Learning Tip: Build a reusable prompt template for your weekly insight narratives and store it in your team's shared prompt library. A consistent format means your stakeholders receive narratives in the same structure every week, making it easier to compare across time periods and reducing cognitive load. The best insight narrative templates are ones where you only need to swap in new data each week — the framing, context, and format instructions remain constant.

Using AI to Identify Anomalies, Trends, and Inflection Points in Product Metrics

Anomaly detection is one of the highest-value applications of AI in product analytics. Humans are reasonably good at spotting large anomalies in charts, but we are poor at identifying subtle inflection points, distinguishing signal from seasonal noise, or correlating anomalies across multiple metrics simultaneously. AI can do all three — with speed and consistency that no manual process can match.

An anomaly in product metrics is any data point or period that deviates meaningfully from the established baseline or expected pattern. In product management, anomalies matter because they are often early signals of significant events: a bug that broke a feature, a successful viral moment, a competitor action that changed user behavior, or a policy change that affected a segment. The faster you detect and diagnose anomalies, the faster you can respond — and early response is often the difference between a manageable issue and a crisis.

Trend identification is distinct from anomaly detection. A trend is a sustained directional movement in a metric over time. Distinguishing a genuine trend from random variation (noise) requires either a statistical approach (which AI can assist with) or enough data points over enough time to see a consistent pattern. When prompting AI for trend identification, always provide at least 8-12 data points and ask AI to separate the signal (the underlying trend) from the noise (random variation around the trend).

Inflection points are moments where the rate of change shifts — where a declining metric starts recovering, or where a growing metric plateaus or reverses. Inflection points are critically important in product management because they often correspond to specific product, market, or organizational events. Identifying them retrospectively helps you understand what drove them; monitoring for them prospectively helps you act before a negative trend accelerates.

When providing time series data to AI for anomaly and trend detection, structure your input consistently: one metric per prompt (or clearly separated sections for multiple metrics), dates or time periods labeled clearly, and a note about any known events (product launches, marketing campaigns, outages, seasonal events) that could explain variation. This context is not optional — without it, AI will hypothesize causes for anomalies that may already be explained by known events, wasting your investigation time.

Hands-On Steps

1. Gather your time series metric data — ideally 12 weeks or 12 months of data for robust trend identification. Export from your analytics tool or BI platform.
2. Annotate the raw data with any known events (feature launches, outages, campaigns, seasonality) in a separate column or as notes in your prompt.
3. Paste the data into AI with a primary ask: identify anomalies (data points that deviate from the pattern), identify the trend direction and confidence, and identify any inflection points in the period.
4. For each anomaly AI identifies, ask a follow-up prompt: "For the anomaly identified in [week/month], propose three possible causes — one product-related, one external/market-related, and one data quality or measurement issue."
5. For trend analysis, ask AI to separate the trend from seasonal or cyclical patterns: "Is the trend you identified potentially explained by seasonal patterns? What would we need to confirm this?"
6. Create an anomaly log: document each identified anomaly with the date, metric affected, severity (% deviation from baseline), and AI-generated hypotheses. Share with your engineering and data teams for investigation.
7. Run the same anomaly detection prompt on correlated metrics to check for consistency: if DAU dropped in Week 8, did session length, new activations, and revenue also shift in Week 8? Correlated anomalies across metrics increase confidence that a real event occurred.
8. Schedule a recurring anomaly detection review — weekly or bi-weekly — using a standing prompt template that you run on fresh data each time.

Prompt Examples

Prompt:

Here are 12 months of Daily Active User (DAU) data for our mobile productivity app. Please analyze this data and identify anomalies, trends, and inflection points.

Month | DAU (average daily that month)
Jan   | 45,200
Feb   | 46,800
Mar   | 49,100
Apr   | 51,400
May   | 53,200
Jun   | 52,800
Jul   | 48,300
Aug   | 47,900
Sep   | 54,100
Oct   | 58,700
Nov   | 61,200
Dec   | 57,400

Known events:
- May: launched push notification reminders feature
- July: major competitor launched similar product
- September: launched integration with popular calendar app
- December: expected seasonal decline (holiday period)

Please:
1. Identify any anomalies (months that deviate significantly from the expected trend)
2. Describe the overall trend direction and characterize it (linear growth, plateau, acceleration, etc.)
3. Identify inflection points — months where the rate of change shifted
4. Distinguish which variations are likely explained by the known events vs. require further investigation
5. Based on the trend, project what January should look like if the current trajectory holds

Expected output: Identification of July as a significant negative anomaly (likely competitor-related based on timing), September as a positive inflection point (calendar integration), the overall trend as growth with a mid-year disruption, December as within expected seasonal range, and a January projection with a confidence range based on the preceding trend.

Prompt:

I need to distinguish signal from noise in this weekly conversion rate data. Please identify whether the recent changes represent a real trend or random variation.

Week | Signup-to-Activation Conversion Rate
W1   | 34.2%
W2   | 35.8%
W3   | 33.1%
W4   | 36.4%
W5   | 35.2%
W6   | 34.7%
W7   | 31.8%
W8   | 30.2%
W9   | 29.4%
W10  | 31.1%
W11  | 28.7%
W12  | 27.9%

Context: No major product changes in W1-W6. In W7 we changed the onboarding flow by adding a 3-step profile setup before activation. No other known changes.

Please:
1. Assess whether the data in W1-W6 represents random variation around a stable baseline, or shows a trend
2. Assess whether the drop starting in W7 represents a statistically meaningful shift from the W1-W6 baseline
3. Calculate the approximate magnitude of the change (how much has conversion dropped?)
4. State your confidence level and explain what additional data would increase confidence
5. Recommend whether the product team should act on this now or gather more data

Expected output: Statistical framing showing W1-W6 as a stable baseline (~34.9% average with normal variance), W7-W12 as a clear downward shift (~5-6 percentage point drop from baseline), assessment of this as a meaningful, non-random change almost certainly linked to the W7 onboarding change, and a recommendation to act now given the sustained nature of the decline and clear causal candidate.

Learning Tip: When running anomaly detection, always ask AI to propose a "data quality" hypothesis alongside the product and market hypotheses. Many apparent anomalies in product metrics are actually tracking bugs, instrumentation changes, or data pipeline issues. Before escalating an anomaly to leadership or launching an investigation, spend 15 minutes asking your data team whether the measurement itself changed. AI cannot know this — you can. A well-run anomaly investigation always rules out measurement error first.

How to Ask the Right Analytical Questions — Prompting AI for Deeper Investigation

The quality of your analytical output from AI is entirely determined by the quality of your analytical questions. This is the meta-skill of AI-assisted analytics: before you open a prompt, you need to know what question you are actually trying to answer. Vague questions produce vague analysis. Specific, well-framed questions produce actionable insights.

The most common failure mode in AI-assisted analytics is asking descriptive questions when you need diagnostic ones. "What does my retention data look like?" is a descriptive question — AI will summarize the data back to you. "Why is my Day 30 retention 40% lower for users who signed up via paid acquisition compared to organic?" is a diagnostic question — AI will attempt to explain a specific, observed difference. Diagnostic questions require you to have already done enough exploratory analysis to know what you are trying to explain.

The iterative analysis technique is how you move from broad to specific efficiently. Start with a broad question about a metric or cohort, receive the AI's initial interpretation, then drill down into the most interesting or surprising element of that interpretation. This conversation-style approach to analytics — each prompt building on the last — is far more productive than attempting to get all insights from a single prompt. Think of AI as an analytical collaborator in a back-and-forth dialogue, not a vending machine you insert a question into and receive a complete answer from.

The "so what?" technique is perhaps the single most valuable analytical habit you can build with AI. After any analytical finding, prompt AI explicitly: "So what? What does this mean for our product decisions?" This forces AI to connect the metric to a decision, the observation to an action. Analytics that does not connect to decisions is expensive documentation. The "so what?" prompt ensures every piece of analysis you do terminates in a product recommendation, even a provisional one.

Segment-first thinking is another powerful technique for deeper analytical questions. Rather than analyzing aggregate metrics, ask AI to generate hypotheses about which user segments might show different patterns, then structure your analysis to test those hypotheses. "Which user segments are most likely to explain the difference between our overall retention rate and our best-case retention rate?" is a question that opens up a rich analytical investigation with a clear goal.

Hands-On Steps

1. Before opening an AI prompt, write down the specific product decision you are trying to inform. If you cannot state a decision, you are doing exploratory analysis — which is fine, but name it as such and use broader, generative prompts.
2. Start with a broad prompt to orient the analysis: "Here is [metric/data]. Give me your top 3 observations and 2 questions you think are most worth investigating further based on this data."
3. Review AI's suggested follow-up questions. Choose the one most relevant to your product context and pursue it with a focused follow-up prompt.
4. Apply the "so what?" technique after each substantive finding: add "And what does this finding imply for product decisions or priorities?" to the end of your prompt, or as a follow-up.
5. Use segment decomposition prompts to drive deeper investigation: "We see [aggregate finding]. Which user segments should we analyze separately to understand whether this finding holds uniformly, or whether specific segments are driving the aggregate?"
6. When you reach a hypothesis, prompt AI to generate a falsification test: "We hypothesize that [X]. What data would we need to see to confirm this? What data would tell us we are wrong?"
7. Document the analytical thread — the sequence of prompts and findings — in a shared discovery log. This creates a record of your reasoning process, not just your conclusions, which is valuable when revisiting decisions.
8. Close every analytical session with a "decision readiness" prompt: "Based on everything we have analyzed, are we ready to make a decision about [specific decision]? If not, what is the single most important thing we still need to understand?"

Prompt Examples

Prompt:

I have this high-level finding from our analytics: Our mobile app's 30-day retention is 22%, but our power users (defined as users who log in 5+ days per week) have 85% 30-day retention.

I want to understand this gap and what it implies for our product strategy.

Please:
1. Generate 3 analytical hypotheses that might explain this gap (beyond the tautological "power users use the product more")
2. For each hypothesis, suggest the specific data or analysis we would need to test it
3. Apply the "so what?" framework: if each hypothesis turned out to be true, what would it imply for product strategy?
4. Suggest the single most important question we should answer next to move from observation to strategy

Expected output: Three hypotheses such as "power users discovered a specific feature set that the average user hasn't" (test via feature adoption analysis by retention segment), "power users have a fundamentally different use case or job-to-be-done" (test via user interviews and behavioral segmentation), and "the onboarding experience fails to route users toward the behaviors that correlate with power use" (test via funnel analysis comparing Day 1 behavior of power vs. non-power users). Each hypothesis with a strategic implication and a recommended next question.

Prompt:

I'm about to do a deep-dive analytical session on our checkout conversion rate, which has been flat for 3 months despite a 15% increase in traffic.

Before I start pulling data, help me structure this investigation.

Please:
1. Generate the 5 most important analytical questions I should answer in this session, ordered from most foundational to most specific
2. For each question, tell me what data I need to gather and what a "surprising" vs. "expected" answer would look like
3. Identify potential analytical traps I should avoid (e.g., confounding variables, segment mix shifts, attribution issues)
4. Suggest a hypothesis I should have going into this session so I am testing something specific rather than fishing

Context: E-commerce platform, B2C, selling home goods. Traffic increase came from a paid social campaign targeting a new demographic (25-34 age group, vs. our traditional 35-50 core). Mobile traffic has gone from 45% to 62% of total traffic in this period.

Expected output: Five ordered questions starting with "Is the conversion rate flat for all traffic or is a new low-converting segment diluting an improving rate for existing segments?" (most foundational) and drilling down to mobile conversion analysis, new vs. returning user behavior, traffic source segmentation, and checkout step-level analysis. Analytical traps would include the segment mix shift from the paid campaign as the likely primary confound. The suggested going-in hypothesis would be that the overall flat conversion rate masks improving conversion for the existing demographic and significantly lower conversion for the new 25-34 paid segment that is less familiar with the brand.

Learning Tip: Develop a personal library of "analytical question starters" — opening prompts you use to begin any new analytical investigation. Good starters include: "Before I analyze this data, help me define what a meaningful finding would look like," "What are the 3 most likely explanations for [observation], and what would prove each one right or wrong?" and "What data would change my current assumption about [topic], and how would I get it?" These starters consistently produce more actionable outputs than diving straight into data analysis.

Key Takeaways

• AI compresses the time between data and insight, but the quality of that insight depends entirely on the context and specificity you provide in your prompts.
• Funnel, cohort, and retention analysis each require different prompt structures; always include benchmarks, user segment context, and known product events.
• The insight narrative format — observation, pattern, hypothesis, recommended action — gives AI a clear structure to follow and gives stakeholders a consistent, scannable output format.
• Anomaly detection prompts should always include three hypothesis types: product-related, market/external, and data quality — rule out measurement error before escalating.
• The "so what?" technique ensures every analytical finding connects to a product decision rather than remaining as interesting-but-inert information.
• Iterative, conversational prompting produces deeper analysis than single-prompt, all-at-once requests; start broad, then drill into what is most interesting or surprising.
• Always document the analytical thread — the sequence of prompts and findings — not just the final conclusion, so your reasoning process is auditable and repeatable.