how to price a test.

every brand runs tests and very few run conclusive ones. the pattern is familiar: a small budget, a short window, three things changed at once, and a conclusion reached from data that could not support it. the channel then gets written off or over-invested in for reasons that had nothing to do with its actual performance.

a conclusive test is not more expensive than an inconclusive one. it is the same money, structured differently.

the cost of getting it wrong is not the test budget — it is the decision that follows. a channel wrongly written off can cost you a year of the cheapest growth available, and a channel wrongly scaled can burn a quarter's budget before the numbers become undeniable. the structure below is cheap insurance against both.

— 01condition one: enough volume to be real.

the practical threshold is roughly thirty conversions. below that, the difference between two variants is mostly noise, and you will confidently draw the wrong conclusion. so work backwards: estimate a cost per conversion for the channel, multiply by thirty, and that is your minimum test budget.

if that number is unaffordable, do not shrink the test — narrow it. one audience, one offer, one geography, so all the volume lands in a single place. a test with thirty conversions in one segment tells you something; the same budget spread over four segments tells you nothing four times.

for high-value, low-volume businesses the arithmetic will not work, and pretending otherwise is a waste. in that case the test is qualitative: run it, then interview the enquiries you got. five conversations with the right buyers is a legitimate result at that scale.

note that conversions here means the thing you care about, not clicks or impressions. a test measured on cost per click can be perfectly clean and completely useless.

and check that the conversion you are counting is being recorded correctly before you start. a surprising share of inconclusive tests are actually measurement failures — duplicate events, a form redirect that never fires, consent-related loss. verify with a live test submission on the day the test launches, not at the end when you are trying to explain the result.

— 02condition two: one variable.

the temptation in a test is to give it the best possible chance — new creative, new audience, new landing page, new offer, all at once. if it works you will not know why, and if it fails you will not know what to fix.

pick the variable with the largest expected effect and hold everything else constant. in paid social that is almost always creative; in search it is usually landing page or match type; in a new channel it is the offer.

this is also why sequential tests beat parallel ones for small brands. run creative for a month, then audience, then page. it is slower and it produces knowledge you can reuse rather than a single ambiguous outcome.

the exception is when you are testing a whole channel rather than a variable within one. launching on a new platform is inherently a bundle — new creative format, new audience behaviour, new auction. accept that, and judge it as a channel decision rather than pretending it is a controlled experiment.

— 03condition three: a defined window.

set the end date before you start, and make it long enough for the platform to leave its learning phase — typically two to four weeks depending on volume. then commit to not touching anything except errors during that window.

the discipline is harder than it sounds. every account manager and every founder has interrupted a test at day nine because the early numbers looked bad, which is exactly when they usually do. resetting the clock repeatedly is how a three-week test becomes a three-month non-answer.

build a mid-point check for errors only: is it delivering, are the links right, is anything disapproved. a mid-point check is not a performance review.

one practical exception on window length: if a channel is clearly failing on delivery — spending almost nothing, or spending at a wildly implausible cost — stop early. that is not impatience, it is a setup problem, and letting it run to the end date wastes both the budget and the calendar slot.

— 04condition four: the decision rule, written first.

before launch, write down what result leads to what action. "if cost per qualified enquiry comes in under x, we scale to y budget. if it lands between x and z, we iterate creative and rerun once. above z, we stop and move the budget to search."

this single paragraph is what separates a test from an expense. it removes the post-hoc negotiation where a disappointing result gets reinterpreted as promising, and it means a stop decision is a plan being followed rather than a failure being admitted.

and record the result somewhere durable, even when it is negative. the accumulated knowledge of what did not work in your category is one of the more valuable assets a marketing function builds, and almost nobody writes it down. a research log costs nothing and stops you retesting the same idea every eighteen months.