Choosing the First Test: Big, Plausible, Reversible

Choose a first test that is big, plausible and reversible. Big means the versions differ in something a visitor can feel: the offer, the page structure, the lead magnet, the presentation of price. Plausible means there is a reason, from recordings, complaints or step rates, to believe this change addresses a real hesitation. Reversible means you can return to the original in minutes if the new version harms trading, without unpicking a rebuild. Aim the test at a page that collects conversions, write the hypothesis before you start, and run it for whole weeks. A modest, honest first result, even an inconclusive one read correctly, is worth more than a dramatic claim nobody should believe.

At low traffic the detectable effect has to be large, so the change has to be large. Reordering a product page so price, delivery and the buy action appear together is big. Rewriting a headline to state a different promise is big. Testing a bundle against a single product is big. Moving a button, swapping a shade, or rewording a link is not big, however tidy it feels. This is not a claim that small changes never matter; it is a claim about what your counts can resolve, covered in the pillar guide. Spend your few readable tests where the swing lives, and let small polish ship on judgement without a test costume.

A plausible test answers a named hesitation. Visitors ask support the same delivery question, recordings show people scrolling for returns information, the funnel drops at the step where total cost first appears. Each is a thread worth pulling. A change copied from a competitor's homepage or a conference talk may still work, but it starts with no local evidence, so a flat result teaches nothing about your visitors. Write the evidence into the hypothesis itself, following writing a hypothesis you can lose: because we observed this, we change that, and we expect this step to move. Plausibility is what makes even a loss useful.

Reversibility protects both revenue and nerve. Before starting, confirm the original version is saved, the switch back is one action, and no other system depends on the new version: no print run quoting it, no ad campaign built entirely on it, no integration that breaks on revert. Offer tests need extra care, since an offer shown to visitors must be honoured for those visitors even after a revert. If a change cannot be undone cleanly, it is a launch, not a test, and it should be judged and logged as one. The first test in particular should be cheap to abandon, because its second job is teaching you the routine under low stakes.

Aim where conversions already happen. For most stores that is the product template, a busy category or landing page, or the lead capture page that feeds the list. Avoid the homepage as a first test unless it carries a single clear job, because homepage visitors arrive with mixed intent and the result will be hard to interpret even if it moves. Avoid the checkout itself on hosted platforms, where your control is limited and the cost of an error is highest. The right page is the one where a named hesitation, a large change and a monthly count of conversions meet. If no page qualifies yet, the first project is not a test; it is fixing and observing until one does.

Decide the primary measure in advance: the conversion the page exists for, not a softer measure chosen later because it moved. Decide the run in whole weeks and the minimum conversions you want per version. Check the variant on a real phone, check that both versions are recorded by analytics, and check that no popup, campaign or scheduled change targets the same page during the run. Then leave it alone. The discipline of not touching a running test feels passive and is in fact the work. Every mid-test adjustment restarts the clock on trust, and a first test fiddled with teaches the worst possible habit: that the rules bend when the graph is exciting.

When the planned period ends, read the result in the order you pre-committed: the primary measure, then the guard measures such as order value or enquiry quality, then the notes on trading conditions. Keep, revert, or declare it inconclusive and say why. File the full result using the test archive, including the screenshot and the exact dates. Do not let a first inconclusive result end the programme; at low traffic it is a normal outcome that still confirms the gate, the logging and the reading all work. The programme's value compounds from routine honesty, not from a first dramatic win.