We follow one use case on Bixlee, a demo drinks store, from start to finish: we write the hypothesis, define who it is for, decide what counts as success, build the variant, launch the test and read the verdict. With real screenshots of every screen.
This guide shows the full cycle of an experiment in Bixlow: a documented hypothesis, an audience, a conversion goal, a variant built on the live site, an A/B test split in the browser and a Bayesian verdict that says when the result can be trusted.
Bixlee is a demo store (in Spanish) and the traffic in step 6 is simulated: it shows how a result is read, it is not a customer's data.
Bixlee runs a different flavour in every Instagram ad. Every click ends up on the same home page, which opens with a watermelon carousel. The flavour that brought the visitor is several blocks further down, and on a phone hardly anyone scrolls that far.
In Hypotheses we create “Carry the Instagram ad through to the home page” and complete it section by section: the problem, the change, the effect we expect and why, who it affects, how we'll measure it and what we need. Each section is one screen, and questions you can't answer are marked “I don't know” or “Later” instead of being made up.
In Audiences we create “Instagram on mobile” with two conditions that must both hold: the UTM source is instagram and the device is mobile. An audience is a named rule you reuse: the same one serves this test and any personalization later on.
In Goals we create “Add to cart”, of type add to cart, with a click detector on the store's buttons (button[name="add"]). Bixlow detects it on the site without touching the theme's code, and every click is recorded as a conversion for that visitor.
![“Add to cart” goal dialog, Detection section: a click detector on button[name="add"].](/guides/ab-test-walkthrough/en-03-goal.webp)
In Workflows we create “Instagram ad home page” and choose “A/B test”. The editor opens two tabs: Original, which stays as it is, and the variant, which we rename “Flavour from the ad”. In the variant we add an “on page start” trigger that runs an inject action: a block with the flavour's name, the can, the price and the button. Its code reads utm_content, hides the carousel and, on click, adds the store's real product to the cart.
“Launch test” opens a summary with both variants and the split (50% and 50%). We give it a name, pick “Add to cart” as what counts as success and, in the “Who takes part” section, the “Instagram on mobile” audience. Once confirmed, the variant reaches the site.
The experiment screen compares both variants over exposed visitors. In the example, the Original converts at 4.0% and “Flavour from the ad” at 5.3%, with a 98% probability of being the best. Bixlow only declares a winner once each variant has enough sample, the test has run for the minimum time and the probability is above 95%. Until then it says there is not enough evidence yet, and why.
None of this stays locked in this experiment. The audience, the goal and what the hypothesis taught you feed the next idea, and the cycle starts again with more to go on.
No. The agent can build the variant from a sentence. The guide shows the code so you can see what gets injected into the site, but you don't have to write it.
No. Bixlow works on any site where you can add its script: Shopify, WordPress, a custom site or a web app. Bixlee uses a Shopify theme because it is a common case.
The screens are real: they come from a Bixlow project on the demo store. The traffic in step 6 is simulated to show how a result is read.
The usual home page. They don't match the audience, so they never enter the test or count in its results.
That is a result too. You end the experiment keeping the Original, the site stays as it was and what you learned stays in the hypothesis for the next idea.
As long as it takes to meet the sample and duration minimums and reach the required probability. Bixlow tells you at every point what is still missing.
Install the script, describe the idea and Bixlow walks you through the same six steps.
