5D Library › 一人公司自动化架构、数字人直播漏斗与任务级人工智能套利 · English
第三章 循环话术的马尔可夫模型
一套循环的脚本
直播带货常把一套固定脚本循环播放:先是点出观众痛点的钩子,然后是产品的机制演示,再是价格锚定(「平时……今天……」),最后是限时的逼单。观众在随机时刻进场,离开的速度取决于屏幕上正在播什么;主播在场与演示建立的信任,才是让观看变成购买的原因 (Wongkitrungrueng and Assarut 2020)。把这个循环当作带离开率的阶段链,就可以精确求解。基准情形下一个周期 6 分钟;从头看完本周期演示的观众在逼单时以 0.05 的概率购买,其他观众以 0.012 的概率购买。
推导
# Chapter 3. p05: A cyclic run-of-show and the viewer who arrives at a
# random moment. One cycle of T = 6 minutes: hook 25%, demonstration
# 35%, price anchor 15%, closing call 25% of T. Viewers leave at rates
# of 0.30, 0.20, 0.20 and 0.35 a minute in the four phases. At the end
# of each close, a viewer still present buys with probability 0.05 if
# she has seen this cycle's demonstration from its start, otherwise
# 0.012, and buyers leave (order value RM 89). Arrival times are spread
# evenly over the cycle (600 points); survival is exact (exponential
# within each phase).
import math
SHARE = (0.25, 0.35, 0.15, 0.25)
LEAVE = (0.30, 0.20, 0.20, 0.35)
FULL, PART, AOV = 0.05, 0.012, 89.0
def viewer(a, T, full=FULL, part=PART, cycles=60):
"""Arrival at time a in [0, T); returns (expected minutes, P(buy))."""
bounds = [0.0]
for s in SHARE:
bounds.append(bounds[-1] + s * T)
alive, minutes, bought = 1.0, 0.0, 0.0
first = True
for c in range(cycles):
for ph in range(4):
lo, hi = bounds[ph], bounds[ph + 1]
if first and hi <= a:
continue
start = max(lo, a) if first else lo
L, h = hi - start, LEAVE[ph]
minutes += alive * (1 - math.exp(-h * L)) / h
alive *= math.exp(-h * L)
if ph == 3:
saw = (not first) or a <= bounds[1]
p = full if saw else part
bought += alive * p
alive *= 1 - p
first = False
return minutes, bought
def average(T, **kw):
pts = [(i + 0.5) / 600 * T for i in range(600)]
rows = [viewer(a, T, **kw) for a in pts]
return (sum(r[0] for r in rows) / 600, sum(r[1] for r in rows) / 600)
if __name__ == "__main__":
m, b = average(6.0)
print(" T = 6 min: mean stay {:.2f} min, buys {:.4f} per viewer"
.format(m, b))
print(" per 1,000 viewers: {:.1f} orders, GMV RM {:,.0f}".format(
1000 * b, 1000 * b * AOV))
for name, a in (("during hook", 0.75), ("during demo", 2.25),
("during close", 5.25)):
m1, b1 = viewer(a, 6.0)
print(" arrives {:12s}: stays {:.2f} min, buys {:.4f}".format(
name, m1, b1))
# T = 6 min: mean stay 3.77 min, buys 0.0144 per viewer
# per 1,000 viewers: 14.4 orders, GMV RM 1,280
# arrives during hook : stays 3.92 min, buys 0.0161
# arrives during demo : stays 3.97 min, buys 0.0093
# arrives during close: stays 3.55 min, buys 0.0190平均每位观众停留 3.77 分钟,购买概率 0.0144:每 1,000 名观众 14.4 单、GMV RM 1,280。进场时刻很重要:钩子阶段进场的观众购买概率 0.0161,演示中途进场的只有 0.0093,逼单阶段进场的却有 0.0190,因为她很快到达一次逼单,随后又能看到完整的一轮。演示中途是最差的进场时刻:观众先碰到的是没有理由支撑的销售话术,要熬过一整轮才能看到理由。
一个周期该多长
演示越长越有说服力,但到逼单时还留着的观众越少。在下一个模型中,短于 0.8 分钟的演示完全没有说服力,超过之后说服力递增但边际递减。
推导
# Chapter 3. p06: Choosing the cycle length T between 3 and 8 minutes.
# Same phases and leaving rates as p05. A longer demonstration persuades
# more, with diminishing returns, and one under 0.8 minutes does not
# work at all: a viewer who saw a demonstration of length D buys at the
# close with probability 0.12 (1 - exp(-(D - 0.8) / 0.8)) (zero if D is
# under 0.8), others with 0.012. Reported per arriving viewer: mean
# stay, orders and GMV (RM 89 an order), for T in steps of half a minute.
import math
SHARE = (0.25, 0.35, 0.15, 0.25)
LEAVE = (0.30, 0.20, 0.20, 0.35)
PART, AOV = 0.012, 89.0
def viewer(a, T, full, cycles=80):
b = [0.0]
for s in SHARE:
b.append(b[-1] + s * T)
alive, minutes, bought, first = 1.0, 0.0, 0.0, True
for _ in range(cycles):
for ph in range(4):
lo, hi = b[ph], b[ph + 1]
if first and hi <= a:
continue
L = hi - (max(lo, a) if first else lo)
minutes += alive * (1 - math.exp(-LEAVE[ph] * L)) / LEAVE[ph]
alive *= math.exp(-LEAVE[ph] * L)
if ph == 3:
p = full if (not first or a <= b[1]) else PART
bought += alive * p
alive *= 1 - p
first = False
return minutes, bought
def per_viewer(T):
full = 0.12 * (1 - math.exp(-max(0.0, 0.35 * T - 0.8) / 0.8))
rows = [viewer((i + 0.5) / 600 * T, T, full) for i in range(600)]
return (sum(r[0] for r in rows) / 600, sum(r[1] for r in rows) / 600,
full)
if __name__ == "__main__":
print(" T P(buy | full demo) stay (min) orders/1,000 GMV")
best = None
for k in range(6, 17):
T = k / 2
m, b, f = per_viewer(T)
print(" {:4.1f} {:.4f} {:5.2f} {:5.1f}"
" RM {:5,.0f}".format(T, f, m, 1000 * b, 1000 * b * AOV))
if best is None or b > best[1]:
best = (T, b)
print(" best cycle: T = {:.1f} min, {:.1f} orders per 1,000 viewers"
.format(best[0], 1000 * best[1]))
# T P(buy | full demo) stay (min) orders/1,000 GMV
# 3.0 0.0322 3.70 28.2 RM 2,514
# 3.5 0.0495 3.69 31.9 RM 2,841
# 4.0 0.0633 3.69 32.0 RM 2,844
# 4.5 0.0745 3.70 30.3 RM 2,697
# 5.0 0.0834 3.71 28.0 RM 2,488
# 5.5 0.0906 3.72 25.4 RM 2,260
# 6.0 0.0964 3.73 22.9 RM 2,034
# 6.5 0.1010 3.74 20.5 RM 1,823
# 7.0 0.1047 3.75 18.3 RM 1,630
# 7.5 0.1077 3.76 16.4 RM 1,457
# 8.0 0.1101 3.77 14.6 RM 1,303
# best cycle: T = 4.0 min, 32.0 orders per 1,000 viewers3 分钟的周期里演示太短(看完整个演示的观众购买概率只有 0.0322),每 1,000 名观众 28.2 单。最佳周期是 4.0 分钟,每 1,000 名观众 32.0 单、GMV RM 2,844;3.5 分钟与 4.5 分钟相差不多(31.9 与 30.3)。到 8 分钟时,演示的说服力是三倍多(0.1101),但每 1,000 名观众只剩 14.6 单,因为大多数观众在逼单前就离开了。在这组离开率下,最优点落在 3 到 8 分钟范围的短端;离开得慢,最优点就会变长。
注
不用施压手段。 模型计算的是看过演示之后自愿购买的观众。虚假稀缺(明明还有很多却说「只剩 3 件」)、假倒计时或编造原价都是欺骗,在许多地方违法;它们不属于本章研究的循环话术。
小结
循环话术是各阶段之间的马尔可夫链。演示中途进场的观众买得最少。在示意性的离开率下,最佳周期约 4.0 分钟(每 1,000 名观众 32.0 单),更短或更长的周期都会损失订单。
参考文献
Fan, Y. (2026). One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage: Coasean Transaction-Cost Collapse, Cyclic Script Retention and High-Margin Solopreneur Portfolios [一人公司自动化架构、数字人直播漏斗与任务级人工智能套利] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090235 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).