5D Library › 短视频完播率风险函数、推荐反馈环与自动化获客漏斗 · English
第四章 钩子宽度的逆向选择与落地时机
播放更多,买家更少
一个「宽」的钩子对所有人说话,一个「窄」的钩子只对目标买家说话。宽钩子带来更多播放,但带进来的多半是不会买的人。设钩子宽度 从 0(最窄)到 1(最宽):日播放为 ,付费转化率为 ,于是 时播放是 10 倍,转化率低 20 倍。转化率按 下降,表示稍微放宽时先带进来的是「近买家」,再放宽才是陌生人。入门产品 RM 49;买家中有 的比例日后购买 RM 298 的课程;退款率为 ;每位买家的服务成本 RM 6,每日固定成本 RM 120。
推导
# Chapter 4. p07: Broad hooks, more views, fewer buyers.
# Hook breadth b runs from 0 (a hook that speaks only to the target
# buyer) to 1 (a hook that speaks to everyone). Illustrative model:
# daily views V(b) = 2,000 x 10^b (10 times at b = 1);
# paid conversion c(b) = 0.01 x 20^(-b^2) (20 times lower at b = 1;
# a slightly broader hook first adds near-buyers, then strangers);
# entry product RM 49; a share u(b) = 0.2 x 2^(-b) of buyers later
# buys the RM 298 programme; refunds r(b) = 0.03 + 0.05 b;
# support cost RM 6 per buyer; fixed cost RM 120 per day.
# Daily net cash = buyers x (49 + u 298) x (1 - r) - 6 buyers - 120.
# The program scans b and finds the best breadth on a fine grid.
def day(b):
views = 2000 * 10 ** b
buyers = views * 0.01 * 20 ** (-b * b)
u = 0.2 * 2 ** (-b)
r = 0.03 + 0.05 * b
cash = buyers * (49 + u * 298) * (1 - r) - 6 * buyers - 120
return views, buyers, cash
if __name__ == "__main__":
print(" breadth b views/day buyers/day net cash RM/day")
for i in range(0, 11, 2):
b = i / 10
v, n, c = day(b)
print(" {:9.1f} {:13,.0f} {:12.1f} {:17,.0f}".format(b, v, n, c))
best = max((day(i / 1000)[2], i / 1000) for i in range(1001))
v, n, c = day(best[1])
print("\n best breadth b = {:.3f}: {:,.0f} views, {:.1f} buyers,"
.format(best[1], v, n))
print(" RM {:,.0f} a day".format(c))
v1, n1, c1 = day(1.0)
print(" the broadest hook: {:.1f} times the views of the best,"
.format(v1 / v))
print(" {:.2f} of its buyers, RM {:,.0f} a day".format(n1 / n, c1))
print("\nReading: the cash optimum sits at a modest breadth, well")
print("short of the hook that maximises views. Past it, each extra")
print("view brings more strangers than buyers, refunds rise and")
print("the premium programme sells to fewer of them. The broadest")
print("hook earns about a fifth of the optimum with ten times the")
print("views of the narrowest.")
# breadth b views/day buyers/day net cash RM/day
# 0.0 2,000 20.0 1,867
# 0.2 3,170 28.1 2,435
# 0.4 5,024 31.1 2,476
# 0.6 7,962 27.1 1,966
# 0.8 12,619 18.6 1,205
# 1.0 20,000 10.0 545
#
# best breadth b = 0.314: 4,121 views, 30.7 buyers,
# RM 2,534 a day
# the broadest hook: 4.9 times the views of the best,
# 0.33 of its buyers, RM 545 a day
#
# Reading: the cash optimum sits at a modest breadth, well
# short of the hook that maximises views. Past it, each extra
# view brings more strangers than buyers, refunds rise and
# the premium programme sells to fewer of them. The broadest
# hook earns about a fifth of the optimum with ten times the
# views of the narrowest.最窄的钩子每天带来 2,000 次播放、20.0 位买家、RM 1,867 的净现金。现金最优的宽度为 0.314:4,121 次播放、30.7 位买家、每天 RM 2,534。最宽的钩子带来 20,000 次播放,是最优宽度的 4.9 倍,买家却只有其 0.33,每天只剩 RM 545,约为最优的五分之一。逆向选择的机制很直接:过了最优点,每一次额外的播放带来的陌生人多于买家,退款上升,高价课程卖给了更少的人。
注意力残留与落地时机
观众从信息流点进落地页时,心思有一部分还留在信息流里。Leroy 的实验表明,在一项任务未完成时切换到另一项任务,会留下「注意力残留」,使下一项任务的表现变差;在切换前完成前一项任务,残留会减少 (Leroy 2009)。程序设到达后第 秒的注意力为 ,观众仍在页面上的概率为 ,在第 秒出价的转化与两者之积成正比。
推导
# Chapter 4. p08: Attention residue and the landing page delay.
# A viewer who taps through from the feed arrives with part of the
# mind still on the feed (attention residue). Illustrative model:
# attention at time t after arrival a(t) = 1 - rho exp(-t / tau);
# chance the viewer is still there exp(-beta t);
# and the offer, shown at time t, converts in proportion to both:
# f(t) = exp(-beta t) (1 - rho exp(-t / tau)).
# Setting f'(t) = 0 gives the best delay
# t* = tau ln(rho (1 + beta tau) / (beta tau)), or 0 if negative.
# tau = 8 s, beta = 0.03 per second. rho = 0.5 when the video leaves
# its story open (a cliffhanger) and 0.2 when it closes the story
# before the call to action. The program also checks t* on a grid.
import math
def f(t, rho, tau=8.0, beta=0.03):
return math.exp(-beta * t) * (1 - rho * math.exp(-t / tau))
def t_star(rho, tau=8.0, beta=0.03):
x = rho * (1 + beta * tau) / (beta * tau)
return max(0.0, tau * math.log(x))
if __name__ == "__main__":
print(" delay t (s) open story rho 0.5 closed story rho 0.2")
for t in (0, 2, 4, 8, 15, 30):
print(" {:11d} {:20.3f} {:22.3f}".format(t, f(t, 0.5), f(t, 0.2)))
for rho in (0.5, 0.2):
ts = t_star(rho)
grid = max((f(i / 100, rho), i / 100) for i in range(6001))
print("\n rho = {}: t* = {:.2f} s, f(t*) = {:.3f}; grid {:.2f} s"
.format(rho, ts, f(ts, rho), grid[1]))
gain = f(t_star(0.2), 0.2) / f(t_star(0.5), 0.5)
print("\n closing the story first: {:.3f} times the conversion"
.format(gain))
print("\nReading: an offer shown the instant a viewer lands meets a")
print("divided mind; a short decompression of several seconds helps")
print("more than it loses to bounces. Better still is to let the")
print("video finish its story before the call to action, so that")
print("less of the mind stays behind.")
# delay t (s) open story rho 0.5 closed story rho 0.2
# 0 0.500 0.800
# 2 0.575 0.795
# 4 0.618 0.779
# 8 0.642 0.729
# 15 0.589 0.618
# 30 0.402 0.405
#
# rho = 0.5: t* = 7.59 s, f(t*) = 0.642; grid 7.59 s
#
# rho = 0.2: t* = 0.26 s, f(t*) = 0.800; grid 0.26 s
#
# closing the story first: 1.246 times the conversion
#
# Reading: an offer shown the instant a viewer lands meets a
# divided mind; a short decompression of several seconds helps
# more than it loses to bounces. Better still is to let the
# video finish its story before the call to action, so that
# less of the mind stays behind.视频以悬念结尾()时,到达即出价的相对转化为 0.500;最优延迟为 7.59 秒,转化升到 0.642。视频在号召行动之前先讲完故事()时,最优延迟只有 0.26 秒,转化为 0.800,是前者最优值的 1.246 倍。两条结论:有残留时,几秒钟的「缓冲」得大于失;更好的做法是让视频先把故事讲完,减少被留在身后的注意力。
假说 · 尚未验证
可检验的推论。 在随机分配的落地页实验中,以悬念结尾的视频引来的访客,其转化应随出价延迟先升后降;先讲完故事的视频引来的访客,到达即出价的转化应已接近最高。什么会证明这一推论错:若两类访客的转化都随延迟单调下降,则注意力残留不是影响落地转化的主要因素。
小结
钩子宽度存在现金最优,远小于播放最大的宽度;越过最优点,播放增加而现金减少。落地页的出价时机取决于观众带来的注意力残留,最有效的办法是让视频先完成它的故事。
参考文献
Fan, Y. (2026). Short-Video Completion Hazards, Recommender Exploration Loops and Automated Acquisition Funnels: Weibull Swipe-Away Kinetics, Multi-Stage Pool Promotion Thresholds and Traffic-to-Cashflow Conversion Bounds [短视频完播率风险函数、推荐反馈环与自动化获客漏斗] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090213 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).