An A/B test is an experiment in which site traffic is split randomly between two versions (A and B) differing in one element, to determine on data which performs the goal better.
The mechanics: half the visitors see the current version, half the changed one (a different headline, layout, shipping price). Once a sufficient sample is collected, statistics decide whether the difference in conversion is real or random.
Rules of a valid test: one variable at a time (otherwise you don't know what worked), a sample and duration fixed in advance (peeking at results and stopping "while winning" is fooling yourself), full weekly cycles (weekends differ from weekdays).
Honest take: with low traffic, A/B tests take months. Usability testing and proven patterns work better then – return to A/B when traffic grows.