5D Library › 超临界病毒级联、信息缺口风险率与多层付费墙优化 · English
第三章 分佣、分享解锁与假账号的激励相容
用户会选哪条路
第一章把分享率与付费率当作给定的数。实际上它们是用户的选择:每个人比较「拉朋友」的代价与「付钱」的代价,挑便宜的那条路,或者干脆离开。机制设计把这类问题称为激励相容:规则必须让每个人按自己的利益行事时,恰好做出设计者希望的事 (Myerson 1981)。第 50 本详细讨论过这一框架;本章只把它用于两条通道。
推导
两条路的代价。 用户对完整测算的估值为 (帕累托分布,最小值 3,尾指数 1.5),每发出一份邀请的个人代价为 (0 到 2 之间均匀分布,即开口求朋友的社交成本)。分享通道需要在接受率 0.15 下凑够 2 人,约发 13.3 份邀请,代价 ;付费通道的代价是 9.9 元减去预期的分佣返还。用户在 超过较小的代价时选那条路,否则离开。
第五段程序在 100,000 个模拟用户上比较四种分佣比例 。
推导
# Chapter 3. Share to unlock, or pay and earn a commission?
# 100,000 simulated users. Each values the full reading at v, Pareto with
# minimum 3 and tail index 1.5, and has a personal cost e per invitation
# sent, uniform on [0, 2] (the social cost of asking a friend).
# share path: needs m = 2 joiners at acceptance 0.15, so about
# 2 / 0.15 = 13.3 invitations: cost 13.3 e
# pay path: price 9.9; a payer sends n2 = 20 alpha commission links,
# each joins with prob 0.10 and pays with prob 0.05 - the
# commission alpha x 9.9 per paying referral is a rebate.
# Each user takes the cheaper path if v exceeds its cost, else leaves.
# The seller keeps 9.9 per payer minus commissions paid out. K counts
# joiners per user (sharers bring min(Bin(20, 0.15), 2) on average).
# For a single seed the expected lifetime number of users is 1 / (1 - K)
# when K < 1, so revenue per seed = net revenue per user / (1 - K).
import math
import random
PRICE, Q, M = 9.9, 0.15, 2
JOIN_SHARE = sum(min(k, M) * math.comb(20, k) * Q ** k * (1 - Q) ** (20 - k)
for k in range(21))
def simulate(alpha, users):
n2 = 20 * alpha
paying_refs = n2 * 0.10 * 0.05
pay_cost = PRICE - alpha * PRICE * paying_refs
share = pay = 0
for v, e in users:
sc = e * M / Q
best = min(sc, pay_cost)
if v < best:
continue
if sc <= pay_cost:
share += 1
else:
pay += 1
n = len(users)
ps, pp = share / n, pay / n
k = ps * JOIN_SHARE + pp * n2 * 0.10
from_links = pp * n2 * 0.10 / k if k else 0.0
net = pp * PRICE * (1 - alpha * from_links)
return ps, pp, k, net
if __name__ == "__main__":
rng = random.Random(63005)
users = [(3.0 * (1 - rng.random()) ** (-1 / 1.5), rng.uniform(0, 2))
for _ in range(100000)]
print(" alpha share pay K net per user per seed")
for alpha in (0.0, 0.25, 0.5, 0.75):
ps, pp, k, net = simulate(alpha, users)
seed = "{:7.3f}".format(net / (1 - k)) if k < 1 else " grows"
print(" {:4.2f} {:5.3f} {:5.3f} {:5.3f} {:7.3f}"
" {}".format(alpha, ps, pp, k, net, seed))
print("\nReading: the commission barely changes who pays - the expected")
print("rebate is tiny - but it makes payers send links, so K rises. Each")
print("step up in alpha raises K and the revenue per seed while lowering")
print("the take per payer. Of the users who stay, about two in three")
print("share rather than pay: asking friends costs them less than 9.9.")
# alpha share pay K net per user per seed
# 0.00 0.212 0.105 0.379 1.038 1.671
# 0.25 0.212 0.106 0.431 1.021 1.796
# 0.50 0.211 0.111 0.487 0.970 1.890
# 0.75 0.209 0.118 0.550 0.887 1.969
#
# Reading: the commission barely changes who pays - the expected
# rebate is tiny - but it makes payers send links, so K rises. Each
# step up in alpha raises K and the revenue per seed while lowering
# the take per payer. Of the users who stay, about two in three
# share rather than pay: asking friends costs them less than 9.9.分佣几乎不改变谁付费:每个付费者预期能拿到的返还很小,付费比例只从 0.105 升到 0.118。它改变的是付费者发出的推荐链接,所以 K 从 0.379 升到 0.550。结果是一个清楚的权衡: 越高,每个用户的净收入越低(从 1.038 降到 0.887),但因为 K 更高,每个种子带来的总净收入更高(从 1.671 升到 1.969)。在这个模型里,50% 的分佣并不是最优点,而是这条单调曲线上的一个点;真正的约束来自下面的假账号。留下的用户中,大约三分之二选择分享而不是付费:对他们来说,开口求朋友比 9.9 元便宜。
证据
推荐奖励的实证。 实验研究发现,推荐奖励会提高推荐的可能性,而且对关系较弱的朋友提高得更多 (Ryu and Feick 2007);一家银行的数据显示,经推荐获得的客户贡献的利润与留存都高于其他渠道获得的客户 (Schmitt et al. 2011)。这些结果说明推荐值得付费,但并不说明 50% 的分佣比例本身是最优的。
当假账号比朋友便宜
任何按人数奖励的机制都会引来作假。第六段程序允许用户以每个 的代价开假账号,比较两种捷径:用 2 个假账号代替朋友来解锁,以及通过一个假账号的推荐链接购买、把 50% 的分佣返给自己。
推导
# Chapter 3. When a fake account is cheaper than a friend.
# Same 100,000 users and funnel as p05, commission alpha = 0.5. Now a user
# can also open fake accounts at a cost f each (a spare phone number, the
# time it takes). Two shortcuts:
# fake unlock a would-be sharer opens m = 2 fake accounts instead of
# inviting friends, if 2 f < 13.3 e (their invitation cost)
# self-referral a payer buys through a fake account's commission link
# and pockets 0.5 x 9.9 = 4.95, if f < 4.95
# Fake joiners are counted as new users by the platform but bring no one.
# For f from 0.5 to 10 the program reports the share of sharers who fake,
# the share of payers who self-refer, the K the platform measures against
# the real K, and the seller's net take per payer.
import math
import random
PRICE, Q, M, ALPHA = 9.9, 0.15, 2, 0.5
JOIN = sum(min(k, M) * math.comb(20, k) * Q ** k * (1 - Q) ** (20 - k)
for k in range(21))
N2 = 20 * ALPHA
if __name__ == "__main__":
rng = random.Random(63005)
users = [(3.0 * (1 - rng.random()) ** (-1 / 1.5), rng.uniform(0, 2))
for _ in range(100000)]
rebate = ALPHA * PRICE
pay_cost = PRICE - ALPHA * PRICE * N2 * 0.10 * 0.05
print(" f fake unlock self-referral K measured K real"
" take per payer")
for f in (0.5, 2.0, 5.0, 10.0):
sharers = fakers = payers = 0
for v, e in users:
sc = min(e * M / Q, M * f)
pc = pay_cost - (rebate - f if f < rebate else 0.0)
if v < min(sc, pc):
continue
if sc <= pc:
sharers += 1
fakers += M * f < e * M / Q
else:
payers += 1
n = len(users)
real_share = (sharers - fakers) / n
k_real = real_share * JOIN + payers / n * N2 * 0.10
k_meas = k_real + fakers / n * M
self_ref = 1.0 if f < rebate else 0.0
take = PRICE * (1 - ALPHA * self_ref)
print(" {:4.1f} {:5.3f} {:5.3f} {:5.3f}"
" {:5.3f} {:5.2f}".format(
f, fakers / max(sharers, 1), self_ref, k_meas, k_real,
take))
print("\nReading: when fake accounts are cheap the funnel's own metrics")
print("lie - the measured K exceeds 1 while the real K is far lower -")
print("and every payer self-refers, so a 50 % commission becomes a 50 %")
print("discount. Above f = 4.95 both shortcuts stop in this example:")
print("two fake accounts then cost more than simply paying.")
# f fake unlock self-referral K measured K real take per payer
# 0.5 0.963 1.000 1.992 0.065 4.95
# 2.0 0.796 1.000 1.357 0.253 4.95
# 5.0 0.000 0.000 0.487 0.487 9.90
# 10.0 0.000 0.000 0.487 0.487 9.90
#
# Reading: when fake accounts are cheap the funnel's own metrics
# lie - the measured K exceeds 1 while the real K is far lower -
# and every payer self-refers, so a 50 % commission becomes a 50 %
# discount. Above f = 4.95 both shortcuts stop in this example:
# two fake accounts then cost more than simply paying.假账号很便宜时(),0.963 的分享者用假账号解锁,平台测得的 K 为 1.992——看起来是强烈的超临界增长——而真实的 K 只有 0.065。同时每个付费者都自我推荐,平台从每个付费者那里只收到 4.95:50% 的分佣变成了 50% 的折扣。只有当开假账号的代价超过 4.95 时,两种捷径都停止。
注
合规与平台规则(需按所在地核对,不是法律意见)。 「分享后才能查看结果」一类的设计,在一些平台上被明确列为禁止的「诱导分享」:据报道,微信的《微信外部链接内容管理规范》把诱导分享类内容与 H5 测试类内容都列为受限制的外链类型 (ifanr, n.d.)。按下线的下线计酬的多级分佣,在中国可能落入《禁止传销条例》所禁止的「团队计酬」等情形 (State Council of the People’s Republic of China 2005);本书的模型只考虑单层分佣。在中国大陆,《广告法》第九条禁止广告含有迷信内容 (Standing Committee of the National People’s Congress 2015)。各条规定的现行文本与适用范围需由作者按经营所在地核对。
小结
用户在拉朋友与付费之间按自己的代价选择。提高分佣比例几乎不改变谁付费,却让付费者发出更多推荐链接,从而提高 K 与每个种子的总收入,同时降低每个用户的净收入。按人数奖励的机制会被假账号利用:假账号便宜时,平台测得的 K 可以远大于 1 而真实的 K 远小于 1,分佣变成自我折扣。合规上,分享解锁、多级分佣与算命类广告都受具体规则约束。
参考文献
Fan, Y. (2026). Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization: Galton-Watson Referral Thresholds, Share-to-Unlock Incentives and Second-Degree Price Discrimination [超临界病毒级联、信息缺口风险率与多层付费墙优化] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090174 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).