How to run growth experiments: the 4-part anatomy — hypothesis, test, measure, learn — plus a real write-up that took signup completion from 11% to 19%.
A growth experiment has four parts: hypothesis, test, measure, learn. You state a specific, testable belief about how one change will move one metric, run the change in isolation for long enough to see a real result, measure it against a threshold you wrote down in advance, and record what it taught you. Run one per week. Fifty structured experiments a year is the difference between growing and guessing.
Most founders run zero structured experiments. They make random changes and hope. This post gives you the full anatomy, a real experiment write-up you can copy, the weekly rhythm, and the five ways experiments fail.
Before you run experiments, you need the dashboard
You can't experiment on metrics you can't see. The prerequisite is a single dashboard with five metrics — acquisition, activation, retention, revenue, referral — that you review every Monday morning. That dashboard is a post of its own: the 5 SaaS metrics that actually matter. Once the dashboard exists, experiments are how you stop watching the numbers and start moving them.
The four parts of a growth experiment
Hypothesis. A specific, testable belief about how a specific change will move a specific metric. "If I change the headline on my landing page from X to Y, signup conversion will improve by at least twenty percent." Not "I think people would like a new headline." Specific. Testable. It names the metric you're moving and the magnitude you expect.
Test. A clean way to run the change for long enough to see a result. For a high-traffic page, an A/B test that splits visitors fifty-fifty. For a low-traffic page, a sequential test: old version for two weeks, new version for two weeks. The key is statistical isolation — you change one thing at a time so you can attribute the result to the change.
Measure. The metric movement, or the lack of it, over the test period. Was the lift you predicted real? Larger or smaller than expected? In the direction you predicted, or the opposite?
Learn. This is the step most founders skip. What did the result teach you about your customers? If conversion improved, why — what does that say about what buyers actually care about? If it didn't, what does that say about the hypothesis you were operating from? The learning is more valuable than the result, because the learning compounds across every future experiment.
A real growth experiment write-up (copy this format)
Here's what a complete experiment looks like in practice — experiment fourteen from a real backlog. Adapt the format; keep the discipline.
| Field | Experiment #14: Reduce signup friction |
|---|---|
| Hypothesis | Moving from a three-step signup (email → password → company info) to a one-step signup (email only) will improve signup completion by at least 30% |
| Why we believe it | Twelve of fourteen recent churn emails mentioned "couldn't be bothered to fill in everything"; three named the password requirement. Current completion (landing to verified email) is 11%; the B2B SaaS benchmark is closer to 18% |
| The change | Remove the password field, use magic-link auth, move company info to a "complete your profile" prompt on first session |
| What we measure | Signup completion rate. Secondary: activation rate from verified email to first core action, to make sure the funnel didn't break further down |
| Sample and length | All visitors for two weeks — around 1,200 visitors |
| Success threshold | At least 25% improvement (hypothesized 30%; 25% counts as a real signal) |
| Result | Signup completion went from 11% to 19%. Activation held steady at 64%. Net new paying customers up about a third versus the previous two weeks |
| Learning | Friction at signup costs more than friction post-signup. The profile prompt was tolerated and didn't increase churn. Hunt for other friction to remove with the same playbook |
That's a complete experiment in about three hundred words. Total time to design, run, and write up: maybe two hours spread over two weeks of clock time. And notice where the value is — not in the 19%, but in the sentence "friction at signup costs more than friction post-signup." That learning applies to four future experiments and shapes how you design every onboarding flow you'll ever build.
The rhythm: one experiment per week
Some weeks the experiment is small — changing a headline. Some weeks it's bigger — rebuilding the onboarding flow. The cadence matters more than the size, because the compound effect of running fifty experiments a year is enormous, and your competitors are running zero.
Realistically, a solo founder can run maybe twenty to thirty real experiments per year once you account for the ones that need building time. That's fine. Twenty structured experiments beat two hundred random changes, because random changes teach you nothing you can reuse.
Five ways growth experiments fail
You decide the result before the data comes in. You believed in the hypothesis — that's why you ran it. Two weeks later the result is ambiguous: maybe a 12% improvement, maybe noise. Your brain wants to call it a win. The discipline is writing down the success threshold before the experiment runs. If the result lands below it, the hypothesis is wrong even if the direction was right. Trust the threshold.
You run too many in parallel. Impatience makes you launch four experiments in the same week. When the metrics shift, you can't attribute the shift to any single change. One experiment at a time, with enough space between them to read each result cleanly.
You stop early on a good signal. "Signups are already up forty percent in three days — rolling it out everywhere." Three days of data is noise. The variance in any short window is huge, and the early good signal disappears almost as often as it persists. Wait the full two weeks.
You ignore failed experiments. A failed experiment is more valuable than a successful one, because the surprise is bigger. You hypothesized 30% and got flat — the gap between your mental model and reality just got named. Write it down. The founders who learn fastest treat failed experiments as the most useful data, not the embarrassing data to hide.
You test things that don't matter. Button colors, headline word counts, font sizes — easy to run, feels productive, rarely moves anything. Spend your experiment budget on pricing, onboarding, and signup conversion: the metrics that move revenue. The button color can wait.
This weekly loop — dashboard on Monday, one experiment always running, learnings written down — is the operating system the Solopreneur Accelerator installs in every founder. It's boring. It works. That's the point.
Frequently asked questions
How many growth experiments should I run?
One per week as the target cadence. In practice a solo founder can run twenty to thirty real experiments a year, because bigger ones need building time. That's still enormous — most founders run zero structured experiments and make random changes instead.
What makes a good growth experiment hypothesis?
It names the change, the metric, and the expected magnitude: "if we move to one-step signup, completion improves by at least 30%." If your hypothesis doesn't include a number and a metric, it's an opinion, not an experiment.
How long should a growth experiment run?
Two weeks for most solo-founder traffic levels — either an A/B split or two weeks of the old version followed by two weeks of the new. Never call a result on three days of data; short-window variance produces false winners constantly.
What should I do when an experiment fails?
Write up the failure with the same care as a win. A flat result against a 30% hypothesis means your model of the customer was wrong in a specific, named way — that correction improves every experiment after it. Then go back to your five metrics and pick the next test.