How to build your first belief graph

Every growth team has a belief graph. Most of them just live inside three people's heads.

Ask your head of growth why the onboarding flow is three steps instead of five. You'll get a story about a test from 18 months ago, a half-remembered confidence number, and a shrug about whether the finding still holds. That story is a belief. The belief graph is what happens when you write all of those stories down, link them together, and update them when reality shifts.

This guide walks you through starting a belief graph from scratch — what to seed it with, how to link entries, and the team rituals that keep it honest instead of letting it calcify into another abandoned Notion wiki.

Why beliefs, not a Notion doc

Teams try to solve this with documentation, and it always rots. A belief is different from a doc in three ways that matter:

• Beliefs are structured. Subject, claim, confidence, evidence, audience — not prose. You can query them, rank them, auto-stale them, feed them to an agent.
• Beliefs are linked. Every belief points to the experiments that created it, the beliefs it depends on, and the beliefs that would conflict with it. A doc is a dead end; a belief is a node.
• Beliefs are updatable by the system, not just by a human editor. When an experiment ships a winner, the source belief's confidence moves automatically. A doc only changes when someone remembers to change it.

The structure of a belief entry

Every entry has the same shape, regardless of domain. The shape is what lets the graph reason over itself.

{
  "id": "bel_7c3e21",
  "subject": "Activation — First-value moment",
  "claim": "Users who hit their first 'aha' event within 10 minutes of signup retain at 3x the rate of users who take longer",
  "confidence": 0.82,
  "audience": "free_trial_signup AND plan = 'starter'",
  "evidence": [
    { "type": "experiment", "ref": "exp_4a91", "weight": 0.6 },
    { "type": "cohort",     "ref": "coh_2d07", "weight": 0.3 },
    { "type": "intuition",  "weight": 0.1, "author": "sarah" }
  ],
  "links": [
    { "relation": "DEPENDS",  "target": "bel_9f01" },
    { "relation": "CONFLICTS","target": "bel_2ab4" }
  ],
  "owner": "sarah@acme.co",
  "reviewed_at": "2026-03-05"
}

Confidence is a number from 0 to 1, not a color. Apex doesn't let you store "high confidence" as a string, because the whole point is that the number should shift when evidence arrives. A belief at 0.82 today can be at 0.54 tomorrow if three experiments land against it.

Start with 5 core beliefs

Resist the urge to dump everything you think you know into the graph on day one. The graph gets better as beliefs collide, not as beliefs accumulate. Start with five load-bearing claims — the ones that would most change what you do if they turned out to be wrong.

B2B SaaS starter set

• ICP — "Our best customers are Series-B-to-C companies with a technical founder and 50–500 employees."
• Activation — "A customer who invites a second teammate in their first session retains at 2x the rate of solo-trialists."
• Pricing — "Moving from seat-based to usage-based expands net revenue retention without hurting new-logo conversion."
• Channel — "SEO for comparison keywords drives 3x the LTV of paid search for the same acquisition cost."
• Retention — "Users who don't hit their second value moment within 30 days churn within 90 days, regardless of plan."

E-commerce starter set

• AOV — "Bundling a complementary SKU at checkout increases AOV by 12% without hurting conversion."
• LTV — "Customers acquired via Meta retargeting have 40% higher 12-month LTV than cold-prospecting customers."
• Margin — "Free shipping above $75 increases order count more than it reduces per-order margin."
• Content — "Product video on the PDP converts 20% better than static imagery for first-time visitors."
• Loyalty — "Customers enrolled in the rewards program have 3.2x order frequency vs non-enrolled."

Marketplace starter set

• Supply — "Liquidity follows supply density per zip code, not total platform supply."
• Trust — "Verified-provider badges lift booking conversion by 25% for new-to-category buyers."
• Take rate — "A 15% take rate captures the most revenue without shifting supplier behavior to off-platform."
• Matching — "Response time under 10 minutes is worth more than a 20% price discount."
• Retention — "A second successful booking in the first 30 days is the activation gate — not the first."

Link beliefs with relations

A belief that stands alone is a hot take. A belief that's linked to other beliefs is a model. Apex supports three relation types, and you should learn all three before you seed more than ten beliefs.

• IF / THEN — "IF activation happens in session one (bel_a), THEN 30-day retention is above 70% (bel_b)." Used when one belief implies another conditionally.
• DEPENDS — "The pricing belief depends on the ICP belief." If the ICP belief's confidence drops, the graph surfaces the pricing belief for re-review.
• CONFLICTS — "This belief contradicts a prior belief that is still in the graph." Both stay until an experiment resolves one. Conflicts are healthy; resolving them is where learning happens.

Use experiments to update beliefs

The magic happens when an experiment ships. In Apex, every experiment is born attached to a belief — you can't run a rootless test. When the experiment resolves, the belief updates:

• A won experiment increases confidence toward 1, weighted by sample size and effect size.
• A lost experiment decreases confidence and is attached as counter-evidence.
• A null result (inconclusive within the power budget) leaves confidence untouched but flags the belief as "tested, unresolved" so it doesn't get re-tested without a better instrument.
• Any DEPENDS belief is automatically flagged for review, because a parent just moved.

This is the closed loop. An experiment doesn't just produce a winner — it propagates a delta through your graph. Over time the graph learns about your business faster than any individual on the team can.

Team rituals that keep it alive

The graph only stays healthy if the team agrees on three rituals. Skip these and the graph will still capture experiments, but it won't capture your team's intuition — which is most of the value.

Weekly belief review (20 minutes, Monday)

Open the Intelligence tab's Stale Beliefs view. Any belief not reviewed in 90 days shows up. Walk through them: archive, re-affirm, or queue an experiment. This replaces the quarterly strategy offsite with a compounding weekly cadence.

Tag outdated beliefs instead of deleting

When a belief becomes wrong, don't delete it. Mark it archived and add the superseding belief as a CONFLICTS link. This way a future teammate can see the history and not re-run an experiment your team already ran in 2024.

Prune low-confidence legacy entries quarterly

Any belief that's been at confidence < 0.3 for two quarters without an experiment attached gets archived. It's not serving the team; it's just noise. Apex surfaces these in the Quarterly Prune view so you don't have to audit manually.

Tools for adding beliefs

Beliefs can come from anywhere a thought lives — the dashboard, the SDK, or your IDE agent. Meet your team where they already work.

• UI — Intelligence tab → New Belief. Good for deliberate entries with full evidence context.
• SDK — apex.beliefs.create({...}) from a Node script or a migration. Good for seeding a graph from an existing research doc.
• MCP — "@apex add belief: our best customers are B-to-C SaaS with a technical founder" from inside Cursor or Claude. Good for capturing the thought the moment you have it, without leaving the editor.

import { apex } from "@/lib/apex-sdk";

await apex.beliefs.create({
  subject: "Activation — Team invite",
  claim: "Inviting a second teammate in session one lifts 30-day retention by 2x",
  confidence: 0.7,
  audience: "free_trial_signup",
  evidence: [
    { type: "cohort", ref: "coh_2d07", weight: 0.6 },
    { type: "intuition", weight: 0.4, author: "sarah" },
  ],
});

90 days from now

Seed five beliefs today. Review them weekly. Attach every experiment you run to the belief it's testing. Let the system do the propagation work.

That's the point. A company's growth intelligence shouldn't be bottlenecked by whether the person holding it stays, remembers, or explains it well in a one-on-one. Write it down, link it, and let the loop compound.