Growth is compounding loops and trustworthy experiments — and most of the discipline is refusing to fool yourself before you read the result.

Pure PLG is rare at scale — winners layer PLG + sales-led + marketing-led motions; sequence them, don't pick one. Build growth loops (output re-feeds input — Reforge) over one-shot funnels.

B2B buys at the account level — score Product-Qualified Accounts (usage aggregated per logo), not lone-user PQLs. Ring-fence product-led sales: reps only chase PQAs past a usage threshold, and don't cannibalize accounts that would self-convert.

Progressive disclosure, in-app education, usage-based upgrade triggers, and collaborative features that drive invites.

Gauge demand before building — ship the entry point, count intent, then show “coming soon.” Cheap signal; never leave it live (trust cost).

One Overall Evaluation Criterion (a short-term proxy that predicts long-term value) plus guardrail metrics that must not regress (latency, crashes, unsubs). Choose metrics first — never after peeking.

Before reading any result, check Sample Ratio Mismatch — a split that deviates from design (e.g. 50/50) means a broken test; debug, don't interpret. Any too-good / surprising result is probably an instrumentation bug.

Week-1 lift can be curiosity (fades) or learning (grows) — date-segment and trust the steady state, not the launch spike.

A/A or pre-period balance check + no overlapping confounds (outage, holiday, price change, PR spike, campaign) + pre-commit the decision rule before unblinding (win→X / flat→Y / loss→Z) so you don't rationalize post-hoc.

Segment only on static pre-treatment attributes (country, device, signup cohort); slicing on post-treatment behavior re-introduces selection bias and invents fake wins.

Watch a metric the change should NOT move (an upstream count, a control-surface event) — if it shifts, you have an instrumentation / randomization bug, not a real effect.

A statistically significant tiny lift may not be worth shipping once you price the opportunity cost (eng time, complexity, surface risk) — judge the effect size against the next-best use of the slot, not just the p-value.

Apply a multiple-comparison correction (Bonferroni / Benjamini-Hochberg) across many metrics, variants, or segments; in networked or marketplace products account for interference / SUTVA via cluster or switchback randomization.

In fintech / healthcare experiments, compliance guardrails (consent, adverse-action, clinical-safety) are not optional guardrail metrics — they gate the test.

Skip low-traffic, strategic / irreversible, or un-instrumentable bets — use judgment plus qualitative. State a hypothesis (the causal why), not just a prediction (the what). STEDI names the org's experimentation maturity — a program capability, not a per-test step.