North Star Metric: An AI Framework for Choosing Your Key Metric

“We’re SaaS, so it’s MRR.” “We’re a marketplace, so it’s GMV.” Most founders pick their North Star Metric by gut — and the team ends up optimizing a number that doesn’t reflect the real value the product delivers to users. MRR grows on the back of annual contracts with high churn. GMV grows through promotional campaigns with negative margins.

Teams focused on one metric that genuinely captures user value find PMF faster than those monitoring 15+ dashboards without clear priority.

This article is a framework for selecting an NSM using AI for data analysis, hypothesis generation, and validation. No abstractions. Just prompts and examples by business model.

What Makes a Metric a North Star

An NSM answers one question: “What single number shows that the product is creating value for both users and the business simultaneously?”

Five criteria for a good NSM:

1. Reflects value for the user. Revenue alone isn’t an NSM. Revenue shows that a user paid, not that they received value. A user could pay, be disappointed, and churn.
2. Measurable. The metric must update at least weekly. “Number of user problems solved” sounds great but can’t be tracked automatically.
3. Leading indicator. NSM predicts future growth. Revenue is lagging: by the time MRR drops, users already left a month ago.
4. Actionable. The team can influence the NSM through product changes. “Market size” is important, but a team of five can’t move that needle.
5. Singular. One metric. Not three. Not “primary and supporting.” One.

NSM by Business Model: A Selection Map

Business model determines the type of value. The type of value determines the NSM category.

SaaS (B2B)

Value: the user solves a work problem faster or better.

Marketplace

Value: the buyer finds what they need, the seller sells.

E-commerce (D2C)

Value: the buyer receives a product they like.

NSM: Number of repeat orders per month or Revenue per customer over 90 days.

Not GMV. GMV grows through acquiring new users, not through value for existing ones. A repeat order proves the first experience was positive.

Subscription Media / Content

Value: the user consumes content regularly.

NSM: Number of days with active content consumption per month (similar to DAU/MAU, but more precise).

Netflix uses “hours viewed per subscriber.” Spotify uses “listening time.” The principle is the same: regular consumption means the content is valuable.

AI Framework for Choosing NSM: Four Steps

Step 1. Gather Product Context

AI won’t choose an NSM for you. But it’ll process the context and surface candidates that you’d miss due to cognitive biases (confirmation bias: “I’m sure our NSM is MRR”).

Prompt for generating NSM candidates:

Product context:
- Product: [name and 2–3 sentence description]
- Business model: [SaaS/marketplace/e-commerce/subscription/other]
- Target user: [who they are and what problem they're solving]
- Current stage: [pre-PMF / post-PMF / scaling]
- Current metrics: [what you're tracking now]
- Core value action: [what the user does when they receive value]

Task: Propose 5 candidates for the North Star Metric.

For each candidate, provide:
1. Metric formulation (precise and measurable)
2. Why it reflects value for the user
3. How often it updates
4. Input metrics (3–4 metrics that drive it)
5. Risks: what can go wrong when optimizing for this metric

Result: five candidates with structured risk analyses. Without AI this takes a week of discussions. With AI you get it in 15 minutes — the team debates finished options instead of generating them from scratch.

Step 2. Filter Through Five Criteria

Each candidate goes through the criteria from the first section. Filtering prompt:

NSM candidates:
[list from step 1]

Rate each candidate on 5 criteria (1–5 points):
1. Value for the user: how well does the metric reflect the moment
   when the user receives benefit?
2. Measurability: can it be tracked automatically, how frequently?
3. Leading indicator: does it predict revenue growth 3–6 months out?
4. Actionability: can a team of [N] people influence this metric
   through product changes?
5. Singularity: does it overlap with other metrics, is it composite?

Format: table with scores, overall ranking, recommendation of top 2
finalists with justification.

Step 3. Validate the Connection Between NSM and Unit Economics

An NSM that doesn’t correlate with business outcomes is just a vanity number. Growing DAU alongside falling LTV means the metric describes activity, not value.

Prompt for validation through unit economics:

NSM finalist: [metric]
Business model: [description]
Average deal size: [X]
CAC: [Y]
Current LTV: [Z]
30-day retention: [%]

Questions:
1. If the NSM grows by 20%, how does that affect LTV?
   Describe the mechanism.
2. If the NSM grows by 20%, how does that affect CAC?
   (Could reduce it through viral loops, or have no impact)
3. What LTV:CAC ratio do we expect at the target NSM value?
4. Is it possible for the NSM to grow while unit economics worsen?
   Provide a scenario.
5. Propose a formula: NSM → LTV (through what intermediate metrics)

If AI finds a scenario where NSM grows but unit economics deteriorate, the metric has failed validation. For more on calculating unit economics with AI, see the unit economics calculator.

Step 4. Build the Metrics Tree

An NSM without input metrics is a number with nowhere to go. Input metrics show which levers to pull.

Tree structure:

NSM: [key metric]
├── Input 1: [acquisition metric]
│   ├── Tactic: [what to do to grow it]
│   └── Owner: [team/role]
├── Input 2: [activation metric]
│   ├── Tactic: [what to do to grow it]
│   └── Owner: [team/role]
├── Input 3: [retention metric]
│   ├── Tactic: [what to do to grow it]
│   └── Owner: [team/role]
└── Input 4: [monetization metric]
    ├── Tactic: [what to do to grow it]
    └── Owner: [team/role]

Prompt for generating the metrics tree:

NSM: [chosen metric]
Product: [description]
Team: [size, roles]

Build a metrics tree:
1. 4–5 input metrics that directly drive the NSM
2. For each input metric: calculation formula, data source,
   update frequency
3. For each input metric: 2–3 specific growth tactics
4. Dependencies between input metrics (growing one may reduce another)
5. Red lines: at what value of each input metric should you sound the alarm?

Example: Choosing an NSM for an AI-Powered Task Manager

Context: B2B SaaS, teams of 5–20 people, AI automatically prioritizes tasks and suggests decomposition. Subscription at $15/user/month. Pre-PMF stage, 200 active teams.

Candidates after Step 1:

Filtering (Step 2):

WAU is out: it doesn’t reflect value. A user could log in, glance at an empty dashboard, and leave. Time to task completion is out too: too many external factors (task complexity, dependencies on other people).

Finalists: candidates #1 and #4.

Validation (Step 3):

Candidate #1: “Number of tasks completed with AI prioritization.” A 20% increase means teams are finishing more work because of the AI. That raises perceived value, cuts churn, and lifts LTV. The mechanism is clear.

Candidate #4: “Number of accepted AI decompositions.” A 20% increase means AI is generating more useful breakdowns. But accepting a decomposition doesn’t guarantee task completion. A user can accept decompositions and never act on them.

Result: NSM = “Number of tasks completed with AI prioritization per week per team.”

Metrics tree:

NSM: Tasks with AI prioritization completed/week/team
├── Input 1: % of teams with AI prioritization enabled
│   ├── Tactic: onboarding flow with AI activation in first session
│   └── Red line: < 60%
├── Input 2: Number of tasks created per week per team
│   ├── Tactic: integrations with Jira, Linear, GitHub Issues
│   └── Red line: < 10 tasks
├── Input 3: % of tasks where user accepted AI priority
│   ├── Tactic: improve prioritization model, build feedback loop
│   └── Red line: < 40%
└── Input 4: Completion rate (% of created tasks reaching Done)
    ├── Tactic: reminders, weekly digest, streak mechanics
    └── Red line: < 30%

Validating the NSM with Data: Three Checks

Picking an NSM isn’t enough. The metric needs validation against real data. Three checks that will confirm or disprove the choice.

Check 1. Correlation with Retention

Split users into cohorts: high NSM value vs. low. Compare 30-, 60-, and 90-day retention. No correlation means the NSM doesn’t reflect value.

-- Example: NSM correlation with 30-day retention
WITH user_nsm AS (
  SELECT
    team_id,
    COUNT(*) AS tasks_completed_with_ai,
    CASE
      WHEN COUNT(*) >= 10 THEN 'high'
      WHEN COUNT(*) >= 3 THEN 'medium'
      ELSE 'low'
    END AS nsm_cohort
  FROM completed_tasks
  WHERE ai_prioritized = true
    AND completed_at >= CURRENT_DATE - INTERVAL '7 days'
  GROUP BY team_id
)
SELECT
  nsm_cohort,
  COUNT(DISTINCT u.team_id) AS teams,
  AVG(CASE WHEN s.active_30d THEN 1.0 ELSE 0.0 END) AS retention_30d
FROM user_nsm u
JOIN team_subscriptions s ON u.team_id = s.team_id
GROUP BY nsm_cohort
ORDER BY retention_30d DESC;

Expected: the high cohort shows 80%+ retention, the low cohort below 40%. A gap under 10 percentage points means the NSM isn’t doing its job.

Check 2. Correlation with Revenue

Prompt for AI analysis of the NSM–revenue relationship:

Data (CSV or description):
- High-NSM cohort: [retention, ARPU, LTV]
- Medium-NSM cohort: [retention, ARPU, LTV]
- Low-NSM cohort: [retention, ARPU, LTV]

Questions:
1. What is the correlation coefficient between NSM and LTV?
2. Is the difference between cohorts statistically significant?
   (given a sample of N teams)
3. What is the minimum NSM value that corresponds to positive
   unit economics (LTV > 3× CAC)?
4. Is this threshold the "magic number" for the activation metric?

That threshold question connects directly to activation analysis — finding the single user action that predicts retention is its own method, covered in the magic number playbook. For more on monitoring metrics in production, see LLM observability with Langfuse.

Check 3. Sensitivity to Product Changes

The NSM should respond to product changes. Run an A/B test on one input metric. If that input grew but the NSM didn’t move, the connection between them is weaker than assumed.

Prompt for planning a validation experiment:

NSM: [metric]
Input metric to test: [metric]
Hypothesis: [if we change X, NSM will grow by Y%]
Current sample size: [number of users/teams]

Calculate:
1. Minimum sample size for statistical significance
   (p < 0.05, power 0.8)
2. Minimum experiment duration
3. Expected effect size
4. Risks: what could distort the result (novelty effect,
   seasonality, selection bias)

Anti-Patterns: Five NSM Selection Mistakes

1. Vanity metric as NSM. Registrations, downloads, page views. These grow through marketing, not through the product. 10,000 registrations with a 2% activation rate is 200 real users.

2. Revenue as NSM at the pre-PMF stage. Early-stage revenue depends on the founder’s sales hustle, not product value. The founder closed 10 contracts through personal connections. Revenue grew. PMF didn’t happen.

3. Composite metric. “Engagement score = 0.3 × DAU + 0.4 × actions + 0.3 × session_length.” Nobody on the team knows what to pull. The coefficients are arbitrary. Growth in one component hides decline in another.

4. Metric without control. “Number of users who recommended the product to friends.” How do you measure that? A quarterly survey? NPS? If a metric can’t update weekly automatically, it doesn’t qualify.

5. Copying someone else’s NSM. Spotify uses “listening time.” That doesn’t mean a 3-person podcast platform should copy it. Spotify’s optimizing engagement for an ad model. A subscription podcast platform needs different signals entirely.

When to Change Your NSM

An NSM isn’t forever. Three situations where it’s time to revisit:

Stage change. Pre-PMF: the NSM should focus on activation and the “aha moment.” Post-PMF: it shifts toward retention and expansion. At scale: efficiency metrics like revenue per employee or gross margin start to matter more.

Business model change. Moving from freemium to enterprise changes what “value” means. “Number of active free users” becomes irrelevant when 80% of revenue comes from 5 enterprise clients.

NSM has hit a ceiling. The metric stopped growing not because the product is struggling, but because it’s reached a natural limit. A 95% completion rate won’t hit 100%. Find a new growth vector — and a new NSM to go with it.

Checklist: From Selection to Implementation

1. Describe the product context using the Step 1 template
2. Generate 5 NSM candidates through AI
3. Filter by 5 criteria (scoring table)
4. Validate the finalist through unit economics
5. Build a metrics tree (4–5 input metrics)
6. Check NSM correlation with retention on real data
7. Run an A/B test on one input metric
8. Set up a weekly report: NSM + input metrics + trend
9. Revisit the NSM when stage, model, or ceiling changes

The core point: an NSM is a focus tool, not a hunt for the “right number.” When 10 people look at one metric and understand how every product decision moves it, decisions get faster and clearer. AI speeds up the selection and validation. But the call stays with the team.

Need help choosing and validating your North Star Metric? I help startups build AI products and automate processes — belov.works.