5D Library › 说服文案代数、损失框架不对称性与结账微界面工程 · English

第八章 赢家诅咒、免费报告页的巴纳姆效应与预注册实验

多少「胜出」的文案是真的

推导

# Chapter 8. p14: How many "winning" copy tests are real?
# True relative lifts of copy variants are drawn from a prior: 70% have no
# effect, 30% have a lift ~ Normal(+5%, 3%). Each test has 13,914 visitors
# per arm at a 3% baseline (p08's size for 3.0% -> 3.6%). We count the
# share of significant winners that are real, and how much the observed
# lift of a winner overstates its true lift (the winner's curse).
import random
from math import sqrt

def run(tests=4000, n=13914, base=0.03, seed=77014):
    rng = random.Random(seed)
    wins = real = 0
    over = []
    se = sqrt(2 * base * (1 - base) / n)
    for _ in range(tests):
        lift = 0.0 if rng.random() < 0.7 else rng.gauss(0.05, 0.03)
        diff = base * lift + rng.gauss(0, se)
        if diff / se > 1.959964:
            wins += 1
            real += lift > 0
            over.append(diff / base - lift)
    return wins, real, sum(over) / len(over)

if __name__ == "__main__":
    w, r, o = run()
    print("   tests 4000; significant winners {}; truly positive {} ({:.3f})"
          .format(w, r, r / w))
    print("   a winner's observed lift exceeds its true lift by {:+.1%}"
          " on average".format(o))
#    tests 4000; significant winners 210; truly positive 131 (0.624)
#    a winner's observed lift exceeds its true lift by +12.4% on average

程序假设:所有文案改动中七成实际没有效果,三成有平均 5% 的相对提升;每次测试每组 13,914 名访客(第四章算出的 3.0% 对 3.6% 的样本量)。在 4,000 次测试中,显著「胜出」的有 210 次,其中真正有正效果的 131 次,占 0.624:将近四成的「胜出」是假阳性。而且胜出版本的观察提升平均比真实提升高出 12.4 个百分点,这就是赢家诅咒:被选中的恰恰是运气好的那些。上线之后效果「回落」通常不是因为环境变了,而是因为当初的估计偏高。

免费报告页

推导

# Chapter 8. p15: The free report page and the Barnum effect.
# A report built from statements that each fit a share f of all readers
# feels personal to everyone. With 8 statements, the chance a reader
# accepts at least 6 (and so feels "this is me") is binomial in the
# acceptance probability 1 - (1 - f)^M (p13 of Vol. 071, M = 10 episodes).
from math import comb

def at_least(k, n, p):
    return sum(comb(n, i) * p ** i * (1 - p) ** (n - i)
               for i in range(k, n + 1))

if __name__ == "__main__":
    print("   f      acceptance   P(6+ of 8 accepted)")
    for f in (0.05, 0.10, 0.20):
        a = 1 - (1 - f) ** 10
        print("   {:.2f}   {:10.3f}   {:.3f}".format(f, a, at_least(6, 8, a)))
    print("   the same report shown to a stranger scores the same:")
    print("   'feels accurate' measures vagueness, not personalisation.")
#    f      acceptance   P(6+ of 8 accepted)
#    0.05        0.401   0.051
#    0.10        0.651   0.431
#    0.20        0.893   0.954
#    the same report shown to a stranger scores the same:
#    'feels accurate' measures vagueness, not personalisation.

免费报告页是测算产品销售流程的起点。若报告由每句符合一成情境的笼统陈述组成,读者在对照自己十段经历时接受每句的概率为 0.651,8 句中接受 6 句以上(「这就是我」)的概率为 0.431;每句符合两成情境时,这个概率升到 0.954。同一份报告给陌生人看,分数完全一样 (Forer 1949)。所以「免费报告让人觉得很准」度量的是陈述的笼统程度,而不是个性化程度;以它为付费转化的依据,也就是以巴纳姆效应为依据。

注

四步核查。 第一,文案中的每个「损失」都要能回答:这个损失是购买可以避免的吗?第二,每个精确数字都要能被核查。第三,结账中删去的只能是非必要字段,退款与取消条款必须可见。第四,测试先算样本量、固定停止规则,并把退款率作为护栏。

预注册的文案实验

推导

# Chapter 8. p16: A pre-registered copy test with a refund guardrail.
# Primary outcome: paid conversion, 3.0% vs 3.6% (two-sided 0.05, power
# 0.80). Guardrail: 30-day refund rate, base 3%; the variant fails if
# refunds rise by more than 2 points (one-sided 0.05, power 0.80 to
# exclude a rise when the true rates are equal). Both computed exactly
# with normal approximations; the test runs until both are met.
from math import sqrt, ceil
from statistics import NormalDist

N = NormalDist()

def n_two(p1, p2, alpha=0.05, power=0.80):
    za, zb = N.inv_cdf(1 - alpha / 2), N.inv_cdf(power)
    p = (p1 + p2) / 2
    return ceil((za * sqrt(2 * p * (1 - p))
                 + zb * sqrt(p1 * (1 - p1) + p2 * (1 - p2))) ** 2
                / (p2 - p1) ** 2)

def n_noninf(p, margin, alpha=0.05, power=0.80):
    za, zb = N.inv_cdf(1 - alpha), N.inv_cdf(power)
    return ceil((za + zb) ** 2 * 2 * p * (1 - p) / margin ** 2)

if __name__ == "__main__":
    nv = n_two(0.03, 0.036)
    buyers = n_noninf(0.03, 0.02)
    print("   visitors per arm for the conversion test: {:,}".format(nv))
    print("   buyers per arm for the refund guardrail: {:,}".format(buyers))
    print("   visitors per arm to reach those buyers at 3%: {:,}".format(
        ceil(buyers / 0.03)))
    print("   binding constraint: {}".format(
        "conversion" if nv >= buyers / 0.03 else "guardrail"))
#    visitors per arm for the conversion test: 13,914
#    buyers per arm for the refund guardrail: 900
#    visitors per arm to reach those buyers at 3%: 30,000
#    binding constraint: guardrail

主要结局为付费转化,3.0% 对 3.6%,双侧 0.05、把握度 0.80,每组需要 13,914 名访客。护栏为 30 天退款率,基线 3%:若新文案使退款率上升超过 2 个百分点即判失败;要以 0.80 的把握度排除这样的上升,每组需要 900 名买家,即约 30,000 名访客。因此真正约束实验规模的是退款护栏,而不是转化率本身:一个只看转化率就宣布胜出的测试,样本量还不到判断它是否在制造退款所需的一半。

假说 · 尚未验证

预注册的预测(尚未验证): 「优惠截止」式的真实损失框架相对于收益框架,付费转化提升不超过 10%,退款率无显著上升;「暗示损失好运」式框架的退款率显著上升。什么会证明它错: 真实损失框架提升显著超过 10%,或暗示损失框架的退款率不上升。

全书小结

损失框架的 2.25 是价值函数的斜率比(p01),现实的转化效应只有几个到十几个百分点。左位数字效应可以计算,最优价格落在整十之下;精确数字的可信度来自可核查性。结账摩擦是存活率的乘积。把退款、拒付与复购放进同一本账,真实的损失框架胜过暗示的损失框架。测试需要事先的样本量、固定的停止规则和退款护栏。

参考文献

Forer, Bertram R. 1949. “The Fallacy of Personal Validation: A Classroom Demonstration of Gullibility.” The Journal of Abnormal and Social Psychology 44 (1): 118–23.

Fan, Y. (2026). Persuasive Copywriting Algebra, Loss-Framing Asymmetry and Checkout Micro-Architecture: Prospect-Theory Valuation Ratios, Numerical Specificity Credibility and Cognitive-Fluency Conversion Bounds [说服文案代数、损失框架不对称性与结账微界面工程] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090205 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).