SaaS Growth Hacking Playbook 2026: Experiments That Move CAC
Growth hacking started as a way to run fast, cheap experiments across the funnel. In 2026 the useful part is the discipline: a written hypothesis, a fair test and a decision rule. This playbook gives you the backlog, the scoring method and a weekly cadence.
TL;DR
- Growth work is a loop of hypothesis, experiment, result and decision, run weekly on one funnel metric at a time.
- Score ideas with ICE (impact, confidence, ease) and re-score after each result.
- Judge winners by payback and retention, not by the size of a traffic spike.
- Pre-register the sample size and the stop rule, or small samples will fool you.
- AI can generate variants quickly, but it does not replace judgment about your customers.
What growth hacking means now
Strip away the folklore and growth hacking is applied experimentation. You identify the step in your funnel that limits growth, propose a change, test it fairly, and keep what works. The word “hack” misleads because durable gains rarely come from tricks.
The funnel is simple enough to fit on one page: acquisition, activation, retention, revenue, referral. Most teams over-invest in acquisition because it is visible, then discover that the leak is activation or retention.
Find the constraint first
Before you brainstorm, draw your funnel with real numbers. For a self-serve product it might look like this:
| Stage | Metric | Last month |
|---|---|---|
| Acquisition | Visitors to signup | 3.1% |
| Activation | Signups reaching first value | 22% |
| Retention | Week-4 active share | 31% |
| Revenue | Trial to paid | 9% |
| Referral | Invites per active user | 0.2 |
The weakest stage relative to its potential is where you work. Here, activation at 22% means three out of four signups never see value. Fixing that beats buying more traffic.
Build and score the backlog
Collect ideas from support tickets, session recordings, customer interviews and competitors. Write each as a hypothesis:
We believe that [change] for [segment] will improve [metric] by [amount], because [evidence].
Then score each with ICE: impact, confidence and ease, each from 1 to 10, averaged. It is crude on purpose. The point is to compare ideas and to update scores as you learn. Where evidence is thin, lower the confidence.
Design experiments that can fail
- Pick one primary metric and one or two guardrails, for example conversion with refund rate as a guardrail.
- Calculate the sample you need before launching, based on baseline conversion and the smallest lift that would matter.
- Randomise where you can. Before and after comparisons mix in seasonality.
- Write the decision rule: ship, iterate or kill, and at what result.
- Do not peek and stop. Stopping when you see a winner inflates false positives.
For low-traffic products where a proper A/B test would take months, use qualitative checks: five customer sessions, a fake-door test, or a manual concierge version of the feature.
A worked example
A team hypothesises that a checklist in the empty dashboard will lift activation. Baseline activation is 22% and the smallest lift worth shipping is two points. They calculate that they need a few thousand signups per variant, so they run the test for three weeks. Result: 24.1% versus 22.0%, within the guardrails. They ship it and log the learning: guided first actions beat blank states for this segment. Next hypothesis: pre-filled sample data.
Examples of experiment types
- Acquisition: comparison pages, landing page message tests, SEO content.
- Activation: checklists, templates, shorter signup, a welcome call for larger accounts. See onboarding and time-to-value.
- Retention: reminder emails triggered by behaviour, integrations that deepen use.
- Revenue: packaging and paywall placement, informed by pricing work.
- Referral: share prompts after wins, covered in referral loops.
A weekly cadence
- Monday: review last week’s results and decide ship, iterate or kill.
- Tuesday: choose the next experiment from the scored backlog and write the spec.
- Wednesday to Friday: build, launch and check instrumentation.
- Friday: update the learning log: hypothesis, result, decision, what you would try next.
The learning log is the compounding asset. Six months of entries tell you what your customers respond to.
Using AI sensibly
Language models are good at drafting copy variants, summarising feedback and proposing hypotheses. They are bad at knowing which of your customers will actually change behaviour. Use them to widen the backlog, then test with real traffic. Keep a human responsible for claims and compliance.
Pitfalls
- Vanity wins. A spike from a launch that does not retain is not growth. See the Product Hunt playbook.
- Changing several things at once, so you cannot say what worked.
- Declaring victory on tiny samples.
- Ignoring payback. A channel that works but costs too much per customer fails the unit economics.
An experiment spec template
Use one page per experiment so results are comparable:
- Hypothesis: we believe that X for segment Y will change metric Z by N, because of evidence E.
- Primary metric and guardrails.
- Sample size and duration, with the calculation attached.
- Variants, with screenshots.
- Decision rule: ship if the primary metric rises by at least N and guardrails hold. Otherwise iterate or kill.
- Result and learning, written within two days of finishing.
Your next step
Draw your funnel with this month’s real numbers, circle the weakest stage and write three hypotheses for it. Score them with ICE and run the top one next week. The log you start now becomes your most valuable growth asset.
Bottom line
Growth hacking in 2026 is a weekly habit of finding the funnel constraint, writing a falsifiable hypothesis, testing it fairly and logging the result. Judge outcomes by retention and payback, protect yourself from small-sample noise, and let AI widen your ideas without choosing for you.
Frequently Asked Questions
Who coined the term growth hacking?
Sean Ellis used it in 2010 to describe people whose true north is growth. The practice that matters today is systematic experimentation, not tricks.
How many experiments should a small team run?
Quality beats volume. One to three well-designed experiments a week on the stage that limits growth is plenty for a small team. More tests only help if you can analyse them properly.
What sample size do I need?
It depends on baseline conversion and the lift you want to detect. As a rule, tiny samples produce noisy winners. Use a sample-size calculator before you start and do not stop early because the numbers look good.
Where should a startup begin?
At the funnel stage with the biggest drop-off between people who showed interest and people who got value. That is usually activation, which is why onboarding experiments often beat acquisition experiments.
