On a page or in your code
Mark two versions in your HTML, and a snippet in the head shows one before the page draws. Or read the experiment as a flag in your app or on your server.
Test a change before you keep it.
A/B tests and feature flags on the same script, judged on the goal you pick.
Sample data: a page experiment on the checkout button, decided by the goal “Order paid”. The original reached it for 49 of 1,204 people (4.1 %), variant B for 71 of 1,188 (6.0 %). Result: variant B wins, at least 3 % more often and about 47 % more often, with a 95 % range from +3 % to +110 %.
MIRA FIVE tests one change against the original, split 50/50, and decides once, after the planned end and its window, on a goal you pick. Feature flags roll changes out to a share of people and turn them off again, on the same script.
A page experiment in HTML: two buttons marked with data-mirafive-experiment="checkout-button", one as variant a, one as variant b.
Planned up front, judged once
Pick 2, 3, 4, 6 or 8 weeks and a window to reach the goal. The verdict comes after that, never while counting.
Plan
Pick the goal, 2 to 8 weeks and a window. The planner shows the smallest change it can find.
Count
Half see the original, half see B. A daily check watches the split and warns of harm.
Decide
After the planned end and its window, one verdict: better, worse, no big difference or no clear winner.
Feature flags on the same script
On/off, variants and remote config, in the website or on your server. Roll out to 25 % of people, then everyone, or turn it off with one switch.
Sample data: four feature flags. new-checkout is on for 25 % of people; free-shipping-banner serves everyone and is ready to remove; onboarding-steps serves two variants; trial-days is remote config serving the number 14.
Questions this page answers.
- How do I run an A/B test on my website without a separate tool?
- How long should an A/B test run?
- Can an A/B test be decided by paid orders instead of clicks?
- How do feature flags roll a change out to some users?
- Do A/B tests need consent?
- Can I start an experiment from Claude or Cursor?
Questions about experiments and flags.
Which goal decides an experiment?
One goal or action you pick when you plan it, such as a sign-up or a paid order. Buyers and revenue are shown beside the result.
Why is there no live winner?
Because peeking picks false winners. MIRA FIVE counts to the planned end and its window, then takes one verdict at 95 %, with a range for the size of the change.
Do experiments need consent?
Yes. People are counted once they agree to statistics and experiments, or from your server. Before that a visitor sees the original or a random variant, and nothing is stored.
Are experiments included in Free?
Yes, every plan has experiments and feature flags. On Free an experiment runs at most three weeks, and visitors entering a test never count against your events.
Can Claude start an experiment?
Yes, once you allow it. Creating one needs “May change setup”; starting, ending or switching flags also needs “May turn flags on and off”. Both are off by default.
Keep only the changes that work.
Add the script, mark two versions, pick a goal. The result comes after the planned end.