Product Analytics for PMs (Mixpanel, Amplitude, SQL): A Working Guide (2026)

Funnel diagnosis: a worked example

Setup: a sign-up funnel with five steps (Landing → Email Entered → Verification Sent → Code Entered → Onboarded). Day-1 data shows 71% step-1 conversion, 64% step-2, 88% step-3, 41% step-4, 79% step-5. Step-4 (Code Entered) is the bottleneck. The diagnostic workflow:

1. Build the funnel chart. Funnels > new funnel > add events in sequence; set conversion window (24 hours); group by user property (device type, geo, traffic source).
2. Group by traffic source. Identify which source has the worst step-4 conversion. Common finding: paid social → 22% step-4 vs. organic → 51% step-4. Likely cause: low-quality traffic; investigate before redesigning.
3. Build a behavioral cohort. 'Users who entered email but did not enter code within 1 hour.' Save the cohort.
4. Drill into the cohort. Use the cohort as a filter in the User Lookup view. Inspect 10 user sessions; look for friction patterns — for example, users who opened email on mobile and didn't have email-app authentication set up.
5. Hypothesize and test. Run a 50/50 A/B test: control vs. magic-link-instead-of-code. Pre-register MDE; run for 2 weeks; ship if step-4 conversion lifts >5pp with statistical significance.

The judgment work isn't the chart; it's choosing which segment to drill into and which hypothesis to test first. Senior PMs who lead with the high-traffic-source finding before redesigning save weeks of work.¹

SQL primer for PMs

Production-fluent PMs at senior+ can write a basic cohort-retention SQL query without analyst help. Worked example: 7-day retention by signup-week cohort.

WITH cohorts AS (
  SELECT
    user_id,
    DATE_TRUNC('week', signup_date) AS cohort_week
  FROM users
  WHERE signup_date BETWEEN '2026-01-01' AND '2026-03-31'
),
active_day7 AS (
  SELECT DISTINCT user_id
  FROM events
  WHERE event_name = 'session_start'
    AND DATE(event_time) BETWEEN signup_date + INTERVAL '7 days'
      AND signup_date + INTERVAL '8 days'
)
SELECT
  c.cohort_week,
  COUNT(DISTINCT c.user_id) AS cohort_size,
  COUNT(DISTINCT a.user_id) AS retained_d7,
  ROUND(100.0 * COUNT(DISTINCT a.user_id) / COUNT(DISTINCT c.user_id), 1) AS retention_pct
FROM cohorts c
LEFT JOIN active_day7 a ON c.user_id = a.user_id
GROUP BY c.cohort_week
ORDER BY c.cohort_week;

The shape works for any SQL warehouse (BigQuery, Snowflake, Redshift). The PM-leverage version: parametrize the day cutoff, add a traffic-source segment, save as a regular query for monthly check-ins.

Common analytical mistakes at senior+

• Reading conversion rates without confidence intervals. '21% lift' on a sample of 800 users is noise. Senior PMs always cite confidence intervals or p-values; junior PMs cite the point estimate.
• Instrumentation trust. Funnel results are only as good as the events firing reliably. The first question on any unexpected funnel shift: did the instrumentation change recently? Many 'wins' are instrumentation drift.
• Sample-ratio mismatch (SRM). When the A/B variant assignment isn't 50/50 in the data, the experiment is invalidated. Always check SRM before quoting results.
• Novelty effects. A new feature lifts metrics in week 1, decays by week 4. Run experiments long enough to surface this; don't ship on week-1 wins.
• Multiple-comparisons risk. Running 10 sub-population analyses on the same experiment guarantees a 'win' by chance. Pre-register the primary metric; treat sub-population analyses as hypothesis-generating, not confirmatory.

Tools by company tier

AI-augmented analysis: where it helps and where it doesn't

Both Amplitude (Ask Amplitude) and Mixpanel (Spark) shipped natural-language query features through 2024–2025. Production patterns:

• Helpful. 'What's our day-7 retention by traffic source for users who signed up in the last 30 days?' — both tools generate the chart in seconds.
• Helpful with verification. 'Build a cohort of users who completed onboarding but didn't return within 14 days, grouped by acquisition source.' Always verify the cohort definition before trusting it.
• Limited. 'Why is conversion dropping?' — output is hypothesis-generation, not conclusive analysis. Treat as a brainstorming aid; validate with a real query.

The judgment work (which segments to slice, which sub-populations to investigate, what to do with the result) is still PM craft. AI accelerates the chart-authoring step.²