Iteration

"This got 4% engagement" is a result. "This underperformed because the angle didn't match the segment's actual priorities" is a learning. Only the second one tells the next hypothesis anything useful — a result on its own is just a number waiting to be interpreted, or worse, ignored.

A mistake repeats not because nobody noticed it the first time, but because nobody wrote down why it happened in a form the next decision could actually read. Without that write-up, every cycle starts from the same blind spot as the one before it.

Compounding rarely looks dramatic cycle to cycle — it looks like slightly sharper angles, slightly fewer wasted tests, slightly faster convergence on what works. The effect is only obvious in hindsight, comparing cycle twenty to cycle one.

An angle that stops getting used without anyone recording why leaves the door open to trying it again in six months, for the same reasons, with the same result. Retiring an angle on the record — what it was, why it didn't hold up — is what actually closes that door.

Changing everything at once after one result makes it impossible to tell which change actually mattered. Changing nothing wastes the learning entirely. The useful middle ground adjusts the one variable the learning actually spoke to, and holds the rest steady enough to isolate the next comparison.

Iteration builds on what the last cycle learned. Starting over discards it — sometimes because the premise really was wrong, more often because starting fresh feels more productive than sitting with an uncomfortable result. Telling the two apart honestly is worth doing before making the call.

A test's winner isn't just a result — it's the new thing every future variant has to beat. Promoting it quietly without resetting the baseline means the next test is being measured against something that's already outdated.

Testing slows down not because it stops working, but because it stops feeling urgent once something is performing well enough. That's exactly when the cost shows up — a system that stops testing while "good enough" is still true stops finding what would have been better.

A learning usually updates the next hypothesis, but its more durable value is updating the standing picture of the segment itself — what a learning reveals about what an audience actually cares about belongs back in research, not just in the next test's brief.

A year in, the individual wins matter less than the shape of the system: research that's stayed current, a strategy that's been sharpened by dozens of small corrections, and a body of learnings specific enough that a new hypothesis rarely starts from a genuine guess anymore.