Running a Real A/B Test on a Lovable Landing Page (Without Breaking It)
Lovable rebuilds your page from a prompt. Here is how to test two versions properly: separate pages, one metric, and discipline about when to stop.
AI page builders make A/B testing dangerously easy in one way and annoyingly hard in another: regenerating a variant takes minutes, but regenerating also changes everything, which destroys the test. The workflow that works treats variants as separate pages that ship separately.
Build version A, freeze it, and deploy it to your main URL. Then build version B as a distinct page, changing only what your hypothesis names. 'A longer hero headline converts better' is a test; 'let the AI make a better page' is a slot machine. My rule: the two pages differ in exactly one element, and I can write down what it is.
The split mechanism is boring and should stay boring: a simple 50/50 redirect on arrival, or your analytics tool's experiment feature if it has one. Do not hand-roll assignment logic with a chatbot's help unless you enjoy discovering that returning visitors saw a different version each visit, which quietly ruins the data.
Decide the metric and the finish line before launch: one conversion event, a minimum sample, and a calendar date. The most common AI-builder testing failure is not a bad test, it is an abandoned one, because regenerating the page mid-test is always one prompt away. Freeze both pages, write the stop date on the calendar, and touch nothing.
With those disciplines, the actual result reporting takes one prompt to your analytics export. My first honest test: the AI's 'improved' hero lost by a little, which was the most valuable outcome possible, because it taught me the tool optimizes for style by default and conversion only if I measure it.