关于「感谢页的注意衰减:没有多巴胺十秒,只有一条半衰期」,《离场降单与感谢页加购》给出的定量结论是什么?
结论:「多巴胺峰值十秒」三个成分,一个都不能照搬。多巴胺编码的是预测误差,付款成功如期出现时误差接近零,信号早在点击付款那一刻就出现了。十秒不是生理常数,而是接受时间分布上的一个分位数,示意参数下前 10 秒集中了 0.830 的接受,但这个数完全取决于自家页面的半衰期。真正可以写成规则的,是一条注意衰减曲线:提示与感谢内容同时出现,不打断,不催促。
数据口径:光标轨迹确实带有意图信息,但模型输出的是概率,合成桌面数据上的 AUC 为 0.809,触屏会话失去鼠标信号后降到 0.618。 多巴胺编码的是奖赏预测误差,没有文献证明付款后存在十秒窗口;感谢页上可测的是加购接受率的半衰期,示意参数下估计为 6.50 秒,前 10 秒集中了 0.830 的接受。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.830, 10
| sessions 3,000 acceptances 171 rate 0.0570 | |
| fitted A0 0.0114 H 6.50 s (true 0.012, 6.00) | |
| bootstrap 90% interval for H | 5.25 to 7.75 s |
| share of acceptances in 0-10 s | 0.830 |
| offer delayed by d s (true model): acceptance per session | |
| d | 0: 0.0566 (relative 1.000) |
| d | 5: 0.0229 (relative 0.405) |
| d 10 | 0.0100 (relative 0.176) |
查看完整验算程序
# Chapter 5. p52: half-life of add-on acceptance on a thank-you page.
# Each second t on the page a buyer either accepts the add-on offer
# (hazard a(t) = A0 2^(-t / H), true A0 = 0.012, H = 6 s), or leaves
# (the Weibull hazard of p51, shape 0.7, scale 14 s), or stays.
# 3,000 synthetic sessions, 60 s window. We fit A0 and H by maximum
# likelihood on per-second counts (leaving acts as censoring), and
# bootstrap 200 resamples for an interval. The "ten seconds" claim is
# then read as a quantile: what share of acceptances fall in 0-10 s?
import math
import random
rng = random.Random(39852)
A0, H, K, LAM = 0.012, 6.0, 0.7, 14.0
def leave_h(t):
s0 = math.exp(-(((t - 1) / LAM) ** K))
s1 = math.exp(-((t / LAM) ** K))
return 1.0 - s1 / s0
sessions = []
for _ in range(3000):
out = (60, 0)
for t in range(1, 61):
if rng.random() < A0 * 2.0 ** (-t / H):
out = (t, 1)
break
if rng.random() < leave_h(t):
out = (t, 0)
break
sessions.append(out)
def counts(rows):
n = [0] * 61
d = [0] * 61
for t, e in rows:
for u in range(1, t + 1):
n[u] += 1
d[t] += e
return n, d
def fit(rows):
n, d = counts(rows)
best = None
for hi in range(8, 61):
h = hi / 4.0
num = math.fsum(d[t] for t in range(1, 61))
den = math.fsum(n[t] * 2.0 ** (-t / h) for t in range(1, 61))
a = num / den
ll = math.fsum(d[t] * math.log(a * 2.0 ** (-t / h)) +
(n[t] - d[t]) * math.log(1.0 - a * 2.0 ** (-t / h))
for t in range(1, 61) if n[t])
if best is None or ll > best[0]:
best = (ll, a, h)
return best[1], best[2]
acc = [t for t, e in sessions if e]
a_hat, h_hat = fit(sessions)
print("sessions 3,000 acceptances {:,} rate {:.4f}".format(
len(acc), len(acc) / 3000.0))
print("fitted A0 {:.4f} H {:.2f} s (true 0.012, 6.00)".format(
a_hat, h_hat))
boot = []
for _ in range(200):
rs = [sessions[rng.randrange(3000)] for _ in range(3000)]
boot.append(fit(rs)[1])
boot.sort()
print("bootstrap 90% interval for H: {:.2f} to {:.2f} s".format(
boot[9], boot[189]))
early = [1.0 for t in acc if t <= 10]
print("share of acceptances in 0-10 s: {:.3f}".format(
math.fsum(early) / len(acc)))
print("offer delayed by d s (true model): acceptance per session")
base = None
for delay in (0, 5, 10):
alive, terms = 1.0, []
for t in range(1, 61):
a = A0 * 2.0 ** (-t / H) if t > delay else 0.0
terms.append(alive * a)
alive *= (1.0 - a) * (1.0 - leave_h(t))
p = math.fsum(terms)
base = p if base is None else base
print("d {:2d}: {:.4f} (relative {:.3f})".format(delay, p, p / base))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《离场轨迹微降单挽留、零门槛诱饵捕获与付款成功页黄金十秒连带母典》第 5 章「感谢页的注意衰减:没有多巴胺十秒,只有一条半衰期」(版本 1.0.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-398 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026398,
author = {Fan, Yafu},
title = {Exit-Trajectory Micro-Downsells, Zero-Friction Lead Magnet Capture and Thank-You-Page Ten-Second Monetisation},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}