The improvement loop: experiment, don't redesign

When a founder decides to lift conversion, the instinct is a redesign, a bold new page that fixes everything at once. That instinct is wrong, and the data on what actually happens in optimisation programmes shows why.

Expect compounding small wins, not a miracle

Convert.com's analysis of completed A/B tests is sobering in the most useful way: 60% of tests deliver under a 20% lift, and 84% come in under 50%. The dramatic, everything-changes win is the rare exception, not the plan. The reliable path to a materially better conversion rate is not one heroic redesign, it is a steady cadence of modest wins that compound. A 10% lift here, a 15% lift there, shipped one after another, beats waiting on a single bet that statistically probably underdelivers anyway. Set your expectations to compounding, and you will run the programme that actually works instead of the one that feels exciting.

This also defuses the redesign trap. A full redesign changes a dozen variables at once, so when conversion moves you cannot tell which change did it, and when it drops you cannot tell what to undo. You have spent weeks and learned nothing transferable. A sequence of isolated changes teaches you something every time, and the knowledge compounds alongside the lifts.

Prioritise by traffic and gap, not by opinion

The question of what to test next should never be settled by whoever has the strongest opinion in the room, even when that room is one founder. Prioritise by traffic volume multiplied by the size of the current conversion gap. The highest-traffic page with the worst relative performance is where a given percentage lift returns the most absolute leads, so that is where you test first. This is the per-source table doing its second job: it already told you which page and channel leak most, so it also tells you where an experiment pays off most. Opinion-led testing scatters effort across pages that cannot move the total; volume-and-gap prioritisation concentrates it where the arithmetic rewards you.

The lean cadence, and when not to A/B test

The cadence is deliberately simple: one test at a time, on your highest-priority page, ship the winner, repeat. One test at a time keeps the signal clean and the operator sane, which matters enormously when the operator is also doing everything else.

There is an important honest caveat for solo operators, and ignoring it wastes months. A/B testing needs traffic to reach statistical significance, and most lean B2B sites do not have it. Running an underpowered test produces a result that looks like a winner but is noise, and acting on noise is worse than not testing at all. When you lack the volume for significance, do not fake it, switch methods. Lean on qualitative research, watch session recordings, read the words customers use, and above all fix message-match, the single highest-leverage non-statistical move you have. For most lean founders, getting the ad-to-page-to-form promise into one continuous story moves the needle faster and more reliably than any A/B test their traffic could ever validate. Test when you have the volume; research and match when you do not; never pretend a starved test is a real one.