Measure the funnel first
Before any change, instrument each step and find where people leave. Conversion work applied to the wrong step improves nothing regardless of how good the change is.
Segment it — the drop-off is frequently concentrated in one device class, one traffic source, or one country, which points directly at the cause.
Find out why, not just where
Analytics show the step; they do not show the reason. Session recordings, support contacts, and a handful of watched sessions usually reveal it within an hour.
Common causes are unglamorous: a validation error that does not say what is wrong, a required field people cannot answer, a slow step people assume has failed, a cost revealed later than expected.
Fix the obvious problems before testing
Broken things do not need an experiment. If mobile users cannot complete a step, fix it; running a test to confirm that a defect reduces conversion wastes weeks.
Reserve testing for genuine uncertainty between reasonable alternatives.
Test properly or not at all
State the hypothesis and the metric before starting. Run until the planned sample is reached rather than stopping when the result looks favourable, which is the most common way tests produce false conclusions.
Change one thing at a time, or you will know the result without knowing the cause.
Reduce what you ask for
Every field, every step, and every decision is an opportunity to leave. The highest-return conversion work is usually removal: fields that could be inferred, steps that could be merged, accounts that could be created after the transaction rather than before.
Ask what each element earns. Several will have been added for a reason that has expired.
Accept inconvenient results
A test that shows the redesign performing worse has saved you from shipping it. Reinterpreting the data until it agrees with the plan removes the entire value of having run it.
Record what did not work as carefully as what did; it is the more useful half of the record.