Lever 2: treat every email as creative

We treat one email as one ad. A single piece of copy can't tell you whether the channel works any more than one paid ad can.

We test wide, expect most versions to lose, and use those losses to find the message the market actually understands. I think of the process like an accordion:

• Expand: Run many offers, angles, and messages.
• Find heat: Spot the combinations that produce the event you chose.
• Compress: Put more volume behind the winners.
• Keep exploring: Reserve part of the program for the next idea.

Once we find message-market fit, we use a 70 / 20 / 10 split:

• 70% of volume goes to proven winners.
• 20% goes to better or adjacent versions of what already works.
• 10% goes to ideas we haven't tried before.

That allocation belongs inside each account. Don't blend results across campaigns and call the shared winner universal. A message can print cash for one market and fail completely in another.

What wide testing looked like for Fyxer

Over 90 days, we sent 12 million emails across 2,749 variants.

At the start, the average campaign needed 7,317 emails for one positive reply. The worst needed 25,000. Later, the program averaged one signup per 520 emails.

In the video, I called that a 45x improvement. Keep the raw event beside each ratio because the early event was a positive reply and the later event was a signup. The direction is clear, but the denominators aren't identical.

The outcome was $4.3 million in annual pipeline for Fyxer. I estimate that if we had stopped early, the best campaign would have produced closer to $400,000 a year.

There’s a deeper breakdown here: How We Built a $32.2M ARR Cold Email Machine for Fyxer AI

The short version of how we test

• Hold the audience and offer constant when testing copy.
• Hold the copy and offer constant when testing audience.
• Name every creative so it can be traced back to the account and through clicks, visits, conversions, and revenue.
• Keep the control live until the challenger has enough downstream events to compare.
• Record failed angles. A failed test is useful only if we can avoid repeating it by accident.
• Don't call a winner because one person replied. Decide the evidence threshold before the send. Basically, get statistically significant data.

Across our data, fancier personalization usually performs worse. Some personalized lines still work, but I make them earn their place with conversion data like any other creative element.