Lesson 74.1 taught you to ask one clean question with a control, one variable, and a decision rule. Now the results arrive — a small table of impressions, clicks, starts, and costs — and the temptation is to declare a universal winner by lunchtime. This lesson teaches the calmer skill: sorting signal from noise and choosing the next sensible move.

Three verdicts, not one winner

Every test ends in one of three places:

• Promising: the variant beats the control on the primary value metric, the guardrail holds, and the pattern is plausible enough to confirm with a follow-up round. Promising means "promote to new control and keep watching," not "scale forever."
• Inconclusive: no meaningful difference, or a difference too small and ragged to trust. This is normal. It means the variable did not matter as tested, or the sample was too thin to tell.
• Warning: attention looks fine but value collapses, costs breach the guardrail, or everything fails. Something structural — destination, audience fit, promise truth — needs inspection before more creative.

Beginners treat all three as "winner and loser." That is how a lucky Tuesday becomes a permanent strategy. Your TEST-PLAN.md decision rule from 74.1 exists precisely to prevent that rewrite of history.

Read this hypothetical Research Desk round:

Variant A looks promising: more clicks and nearly double the starts at lower cost per start. Variant B is inconclusive-to-weak: slightly more attention than control, no value gain. But before promoting A, check the four traps below. A result without those checks is a rumor with numbers.

Sample, fatigue, mismatch, and the story you want to be true

Sample size in plain language. Sample is not impressions; it is relevant exposures plus completed actions. Sixteen starts versus nine is directionally interesting, not proof. With small counts, two extra conversions swing the story. Call A promising, run it as the new control for a second week, and confirm before scaling spend. If your test produced only two or three total actions, the honest verdict is inconclusive — collect more evidence.

Fatigue. Audiences tire of the same creative. A control that ran for six weeks may lose simply because it is familiar, not because the new idea is fundamentally better. Note launch dates and frequency (average views per person). If frequency is high and performance slides for all variants together, you are watching wear-out, not a creative contest.

Audience mismatch. The wrong audience produces confident-looking garbage. A broad interest audience may click a clever hook and never start a research brief, while a narrow analyst audience clicks less but converts. If clicks rise and starts do not, suspect mismatch or a destination gap before blaming the hook. Check: did the audience match the question in the plan, or did the platform expand it silently?

Measurement gap. Broken links, slow pages, mismatched promises, and unattributed actions all masquerade as creative failure. If every variant fails on value while attention looks normal, stop testing hooks and inspect the destination. One Research Desk team once "proved" three hooks failed before discovering the sample-brief button loaded in nine seconds on mobile. The hooks were never the test.

Add confirmation bias: the human habit of seeing the winner you hoped for. Pre-written decision rules, a second reader, and the sentence "what would change my mind?" are the antidotes.

The outcome table and the next move

Use this map when you sit down with any small report:

Apply it to the tailor. Demo earns more clicks but bookings stay flat because the booking page hides Friday availability. Verdict: warning on the destination, not a creative loss. Fix the page, re-run the same question, then judge.

Seasonality deserves one line: a test run over a holiday or local event may not repeat. Note the calendar so future-you does not treat December behavior as a law.

Exercise: write the readout

Create TEST-READOUT.md with exactly these sections:

• Observed result: numbers in a small table, with dates, spend, and audience noted. No adjectives yet.
• Interpretation: one of promising / inconclusive / warning, in one sentence.
• Confidence: low / medium / higher, with reason — sample depth, consistency across days, guardrail status.
• Confounds checked: sample, fatigue, mismatch, measurement, seasonality — one line each, even if the answer is "no evidence."
• Next test: the single follow-up question, or the non-creative fix (page, audience, offer) that must come first.
• What not to conclude: the overclaim you are refusing to make. Example: "Do not conclude demo hooks always win; do not apply this to a different audience."

Worked example for the table above:

Finish line: a TEST-READOUT.md that lets a stranger see the numbers, the verdict, the confidence, and the next step — without inheriting your optimism.

Verify quickly: cover the numbers and read only the interpretation. Then uncover the numbers. If the interpretation claims more than the numbers support, shorten it.

Common failure mode: promoting a one-day spike to permanent control and doubling spend. Promising means confirm, then scale deliberately — the subject of Lesson 74.4.

Check your understanding

1. What distinguishes a promising result from an inconclusive one and from a warning? 2. Why can high clicks with low actions point to the landing page rather than the creative? 3. What belongs in the "what not to conclude" section, and why does it matter?

Next

You can read a result honestly. Lesson 74.3 makes that honesty compound — a creative library in Markdown and JSON that preserves what you tried, what happened, and what to reuse next launch.