Research I run myself, and experiments instead of arguments

Never test a random idea. A hypothesis starts from a leak in the funnel data — the page a new visitor lands on, the activation step, the screen where a free user decides whether to pay — or from what people say in interviews, and the two disagree often enough to be worth running both. On a bank-embedded fintech the interviews showed a motivation problem no screen could fix, so the answer was a different product, not a better form.

The flow worth a moderated session is the one that fails in front of other people. Liveness instructions for a biometric gate — which have to work at a glance, in a hurry, for someone who has never seen the system — were usability-tested before they shipped. An enterprise HR platform went from personas and journey maps through wireframes and prototypes to usability tests of three roles on one data model. The question is never whether they liked it; it is whether they got through without being told how.

Test variations along one controlled dimension at a time. Holding block structure constant and changing styling alone is how the one decisive result was produced — and why it transfers to the next page instead of being a lucky accident.

Run experiments long enough, across new and returning traffic, and segment by device from the first day. Use Bayesian probability of outperforming rather than a read taken too early — on a large funnel an early read produces a number that is confidently wrong, and one wrong call ends a team’s appetite for testing.

I will not claim an audit I cannot show, so the evidence is the site you are reading. One focus rule for every interactive element, visible in both themes — it replaced a password field that hid its outline behind a border too faint to count as one. A skip link and a main landmark on every page. Scrollable results tables reachable from the keyboard, because the columns past the edge are the point of the table. Reduced motion honoured in one rule, added after a review found the motion page promising it and the stylesheet not delivering it. Contrast measured per theme, not assumed: the accent sits at 5.38:1 in light and 8.60:1 in dark; the one it replaced measured 3.71:1 as a link and 4.10:1 under white button text, so it failed twice. And one adjacent habit from early on: self-service terminals in Arabic and Urdu taught me that localisation is a layout problem before it is a translation problem.

Convert the delta into currency, conservatively — the floor, not the ceiling, because a number nobody can argue down is worth more than one that invites argument — and put it in front of the people who decide. When the organisation does not ship the winner, put a number on that too. The strongest result of the programme on the eight-figure funnel was never rolled out; the forgone-revenue figure is the case I should have made beforehand rather than documented afterwards.