Class 73 gave you a production line for advertising creative: one approved angle, controlled variations, and a review gate before anything spends money. This class answers the next question — once approved creative exists, how do you learn from it without burning budget or fooling yourself?

Volume without a question teaches nothing

New advertisers hear that winners test a lot: ten headlines, twenty images, ten video variants. So they launch ten unrelated concepts at once, watch the dashboard flicker, and declare a winner. A week later the winner stops working and nobody knows why. That was not testing. That was a slot machine with better fonts.

A creative test is a question with a price tag. It says: we believe something specific about our audience, we will change one thing to check it, and we will decide in advance what each outcome means. Volume becomes useful only when every variant serves a decision.

Learn the core vocabulary in plain language:

• Hypothesis: a falsifiable prediction. Not "this ad will do well," but "for solo research builders, a screen demonstration hook will produce more qualified starts than a founder talking-head hook, because the audience trusts visible evidence over claims."
• Control: the current best-approved creative. Everything new is measured against it, not against zero.
• Variable: the single element you change. Hook, proof block, format, first frame, headline, or call to action — pick one per round.
• Sample: the number of relevant people who actually saw the test. One hundred curious clicks from the wrong audience is not a sample; it is noise with a receipt.
• Signal: a meaningful difference tied to the useful action, not just attention.
• Decision rule: what you will do if A wins, if B wins, or if nothing differs. Written before launch.
• Learning agenda: the ordered list of questions you will test over the next month, so each test builds on the last.

One question per round

Suppose Research Desk wants to promote its source-linked sample brief. The team has one approved angle: "every AI claim must carry its citation trail." They could test audience, offer, hook, format, and landing page all at once. Instead they write one question:

That sentence does five jobs. It names the audience (solo research builders). It names the variable (opening hook). It names the control (talking head). It names the metric (qualified brief starts, not views). And it names what stays fixed (offer, audience, destination, spend window). Anyone reading it knows what is being learned.

Contrast that with the common version: "let us test some creatives and see what pops." No hypothesis, no control, three variables changed at once, success defined afterward as whichever number looks biggest. You cannot repeat it, defend it, or build on it.

Hold these constants every round:

FIXED: audience + offer + destination + budget cap + time window
CHANGED: one variable (hook OR proof OR format OR headline OR CTA)
MEASURED: one primary metric tied to value + one guardrail metric

The guardrail matters. If demonstration hooks earn clicks but the visitors bounce on the landing page, you have an attention win and a value failure. Track the click and the completed action together.

Use the local business version. A tailor testing "same-week hemming" holds the offer (hem ready by Friday, fixed price), the audience (10 km radius), and the destination (booking page) constant. The variable is the opening frame: finished garment on a hanger versus tailor measuring at the table. The question: does visible finished proof beat process proof for booking starts? One variable, one metric, one week, fifty euros capped. That is a test a small shop can actually run.

Why ten headlines can still be disciplined

Structured volume is fine. If you have one angle and six hooks, that is six answers to the same question: which opening earns the next step? If you have one hook and four proof types (customer quote, screen capture, before-and-after, analyst endorsement), that is one question about evidence. What fails is ten headlines each making a different promise to a different imagined person with a different destination. That produces movement, not learning.

Build a simple round:

Two variants plus a control is plenty for a first round. Add variants only when you can still give each enough relevant exposure to read. Three underfed variants teach less than two well-fed ones.

Exercise: draft the plan

Create TEST-PLAN.md with these fields, filled in completely:

• Question: one sentence, audience + variable + metric.
• Hypothesis: "We believe [change] will cause [effect] because [reason grounded in customer evidence]."
• Control: asset name, format, current baseline if known.
• Variants: A and optionally B, each labeled with the one variable changed.
• Audience: who, where, and who is explicitly excluded.
• Destination: exact URL or page, with message match confirmed.
• Budget and window: cap and dates, e.g. "120 EUR, 7 days, stop early only if destination breaks."
• Primary metric and guardrail: e.g. "brief starts; guardrail cost per start under X."
• Decision rule: "If A beats control by [threshold] on starts with guardrail intact, A becomes new control. If no meaningful difference, keep control and test proof next. If both fail guardrail, inspect destination before more creative."

Example hypothesis for ByeBuy Classroom: "We believe a lesson-screen demonstration hook will produce more lesson starts than a presenter hook for self-taught builders, because prior comments ask to see the actual workflow before committing time."

Finish line: a TEST-PLAN.md that a teammate could launch without asking what success means, and stop without arguing about what happened.

Verify quickly: hand the plan to someone cold and ask "what changes, what stays fixed, and what do we do if nothing wins?" If they hesitate, the question is still foggy.

Common failure mode: testing the offer and the creative together — new promise, new audience, new page, new hook. Whatever happens, you will not know which change mattered.

Check your understanding

1. What is the difference between a control and a variable, and why hold everything else fixed? 2. Why is "qualified brief starts" a better primary metric than views for Research Desk? 3. What does a decision rule protect you from after the results arrive?

Next

You know how to ask one clean question. Lesson 74.2 teaches you how to read the answer — promising, inconclusive, or warning — without mistaking sample noise, fatigue, or audience mismatch for truth.