Scale winning experiments across channels

Before you scale, check the result. Did it pass the success threshold you wrote down before launch? Did it run long enough? Did anything else change in the same period? Use the rules in Analysing and acting on results and Decide what a win means before the test goes live.

If the result is borderline, re-run it before you build on it. Scaling a false win spreads the error across every channel.

Describe the mechanism in one sentence. Not "version B won" but "a headline that states the outcome with a number gets more clicks than one that describes the product."

The mechanism is what travels. The execution does not. If you cannot state a reason, you have only a result, and you should treat the next tests as exploration.

Take the mechanism and list every channel where the same idea could appear. For the example above:

• The headline of your LinkedIn ads.
• Email subject lines.
• The title of your booking page.
• The opening line of your outbound messages.
• The first slide of your sales deck.

Aim for five to eight places. Cross out any that do not have enough volume to produce a result in a reasonable time.

You cannot test all of them at once. Score each by likely impact, probability that the mechanism carries over, and effort. Put them in your backlog and sort. Use Rank the backlog by expected value, not by enthusiasm for the method.

Channels close to the original, such as the same audience on a different format, usually carry over best. Channels with a different audience or intent carry over least.

Each channel has its own format, length and reader state. A headline that works on a landing page may be too long for an ad, and an email reader is in a different mood from a person scrolling LinkedIn.

Keep the mechanism and rewrite the execution. If you copy the winning asset word for word, you test the asset and not the idea. A failed copy then teaches you nothing.

Give every adaptation its own hypothesis, success threshold and run time, as in Creating strong hypotheses. Run two or three at a time, not eight.

Set a realistic target. Expect the effect to be smaller elsewhere. A planning rule I would use is half of the original lift, and a result at that level counts as a success.

When each test closes, log it: the channel, the adaptation, the result, and whether the mechanism held. Over time this tells you how far a given kind of idea travels. A mechanism that works in four of five channels is a principle. One that works in one place was a local trick. Keep the log as in Compound learnings across experiments.

• Rolling a win out everywhere at once, with no test and no baseline.
• Scaling a result that was never properly validated.
• Copying the asset instead of the idea.
• Expecting the same lift in a different channel.
• Running too many adaptations at once, so none gets enough traffic.
• Forgetting to log the failures.

• You can state, in one sentence, why the original experiment won.
• You have a ranked list of channels with a hypothesis for each, and two or three are running.
• At least one adaptation beats its channel baseline.
• Your log shows which channels the mechanism carried to and which it did not, so the next win scales faster.