5D Library › Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization · 中文

Chapter 6. Bundles, Decoy Tiers and Paying per Question

Why bundling works

Chapter 5 showed that the extra money a multi-tier paywall takes comes from selling more goods. These goods can be sold separately or together. Adams and Yellen long ago pointed out that when buyers’ values for several goods are not perfectly correlated, bundling lets the seller collect more (Adams and Yellen 1976); Bakos and Brynjolfsson showed further that for information goods with zero marginal cost, the more items in the bundle, the more each buyer’s average value per item concentrates near its mean, and the larger the share of value a single bundle price can collect (Bakos and Brynjolfsson 1999).

The eleventh program makes the value of each question independent (log-normal, median 2) and compares separate pricing with one bundle.

Calculation

# Chapter 6. Why "ask as many questions as you like" can pay.
# A buyer's value for each answered question is independent across
# questions: log-normal with median 2 and log-sd 0.8. Selling n questions
# separately at the best single price per question, against selling the
# n as one bundle at the best bundle price (Bakos and Brynjolfsson: the
# law of large numbers makes the bundle's value per question concentrate
# near its mean, so one price captures more of it). 20,000 buyers.
import math
import random

def best_price(vals):
    vs = sorted(vals, reverse=True)
    return max(v * (i + 1) for i, v in enumerate(vs)) / len(vs)

if __name__ == "__main__":
    rng = random.Random(63011)
    print("   questions   separate (per q)   bundle (per q)   value per q"
          "   share")
    for n in (1, 2, 5, 10, 20):
        vals = [[2 * math.exp(rng.gauss(0, 0.8)) for _ in range(n)]
                for _ in range(20000)]
        sep = best_price([v for row in vals for v in row])
        bun = best_price([sum(row) for row in vals]) / n
        mean = sum(sum(row) for row in vals) / (20000 * n)
        print("     {:3d}          {:6.3f}             {:6.3f}"
              "          {:6.3f}        {:5.3f}".format(
                  n, sep, bun, mean, bun / mean))
    print("\nReading: priced one by one, each question earns the same share")
    print("of its value however many there are. As one bundle, the share")
    print("captured climbs towards the whole value as n grows. This is an")
    print("argument for bundles of independent questions, not for tiers of")
    print("one reading.")
#    questions   separate (per q)   bundle (per q)   value per q   share
#        1           1.003              1.003           2.770        0.362
#        2           1.000              1.187           2.753        0.431
#        5           1.001              1.463           2.749        0.532
#       10           0.999              1.670           2.758        0.605
#       20           0.998              1.861           2.746        0.678
#
# Reading: priced one by one, each question earns the same share
# of its value however many there are. As one bundle, the share
# captured climbs towards the whole value as n grows. This is an
# argument for bundles of independent questions, not for tiers of
# one reading.

Priced separately, each question collects the same 1.000 or so however many there are; as one bundle, revenue per question rises with the number of items, from 1.003 for 1 question to 1.861 for 20, and the share of value collected rises from 0.362 to 0.678. This is the logic of an “ask as many as you like” plan — but it rests on the questions being valued independently. It is a reason to bundle independent questions, not a reason to split one reading into tiers.

Share of value collected per question, priced separately and as one bundle, as the number of questions n grows.

The decoy tier

Designs such as an “anchoring joint-purchase package” often rely on another effect: adding a clearly inferior option to a menu so that the expensive option beside it looks like a bargain. Huber, Payne and Puto called this the asymmetric dominance effect (Huber et al. 1982); research on anchoring also shows that an irrelevant number can change what people are willing to pay (Ariely et al. 2003). The twelfth program adds to a multinomial logit choice model a decoy fully dominated by the premium tier, with a parameter λ\lambda for the size of the attraction effect.

Calculation

# Chapter 6. The decoy tier: an effect that has to be measured.
# Visitors choose among: nothing (utility 0), a basic reading A (quality
# 1, price 9.9) and a premium reading B (quality 2, price 29.9). A decoy D
# (quality 1.7, price 29.9) is dominated by B and is not meant to sell.
# Choice is multinomial logit with utility 4 q - 0.15 price - 1.5 for
# each product, plus lambda for every option it dominates (the
# "attraction" or asymmetric-dominance effect of Huber, Payne and Puto).
# With lambda = 0 the decoy works only through the logit's arithmetic.
# The program reports choice shares and revenue per visitor with and
# without the decoy for lambda from 0 to 1.
import math

OPTS = {"A": (1.0, 9.9), "B": (2.0, 29.9), "D": (1.7, 29.9)}

def dominates(x, y):
    (qx, px), (qy, py) = OPTS[x], OPTS[y]
    return qx >= qy and px <= py and (qx, px) != (qy, py)

def shares(menu, lam):
    u = {"none": 0.0}
    for o in menu:
        q, p = OPTS[o]
        u[o] = 4 * q - 0.15 * p - 1.5 + lam * sum(
            dominates(o, other) for other in menu if other != o)
    z = sum(math.exp(v) for v in u.values())
    return {k: math.exp(v) / z for k, v in u.items()}

def revenue(sh):
    return sum(sh[o] * OPTS[o][1] for o in sh if o != "none")

if __name__ == "__main__":
    print("   lambda   menu      none    A       B       D       revenue")
    for lam in (0.0, 0.5, 1.0):
        for menu in (("A", "B"), ("A", "B", "D")):
            sh = shares(menu, lam)
            print("   {:4.1f}     {:8s}  {:5.3f}   {:5.3f}   {:5.3f}   {}"
                  "   {:6.3f}".format(lam, "+".join(menu), sh["none"],
                                       sh["A"], sh["B"],
                                       "{:5.3f}".format(sh["D"])
                                       if "D" in sh else "  -  ",
                                       revenue(sh)))
    print("\nReading: without an attraction effect (lambda = 0) revenue")
    print("rises only because some visitors buy the dominated decoy itself:")
    print("they pay the premium price for less. The further rise at lambda")
    print("> 0 comes entirely from one assumed parameter. Experiments with")
    print("realistic products have often found it small or absent, so it")
    print("has to be measured, not assumed.")
#    lambda   menu      none    A       B       D       revenue
#     0.0     A+B       0.089   0.245   0.666     -     22.343
#     0.0     A+B+D     0.074   0.204   0.555   0.167   23.606
#     0.5     A+B       0.089   0.245   0.666     -     22.343
#     0.5     A+B+D     0.054   0.150   0.673   0.123   25.272
#     1.0     A+B       0.089   0.245   0.666     -     22.343
#     1.0     A+B+D     0.038   0.104   0.772   0.086   26.678
#
# Reading: without an attraction effect (lambda = 0) revenue
# rises only because some visitors buy the dominated decoy itself:
# they pay the premium price for less. The further rise at lambda
# > 0 comes entirely from one assumed parameter. Experiments with
# realistic products have often found it small or absent, so it
# has to be measured, not assumed.

With no attraction effect (λ=0\lambda = 0), adding the decoy raises revenue per visitor from 22.343 to 23.606 — but only because 0.167 of visitors buy the decoy itself, paying the premium price for less. This revenue comes from buyers’ mistakes, not from clever design. As λ\lambda grows revenue keeps rising (26.678 at λ=1.0\lambda = 1.0), and this increase comes entirely from that one assumed parameter.

Evidence

The decoy effect is not stable for real products. Frederick, Lee and Baskin ran a series of experiments with realistic products and descriptions and found the attraction effect small in many cases, and sometimes absent or reversed (Frederick et al. 2014). Book 47 discussed evidence of this kind on choice architecture. Whether a decoy tier works can only be measured with a particular menu and audience.

Pay per question, or an annual pass?

“One matter, one reading, priced per question” is paying per use; its opposite is a yearly plan. The thirteenth program assumes that how often users ask in a year varies from person to person (gamma-distributed; the smaller the shape kk, the more a few people ask a lot and most ask little) and compares paying per question, an annual pass, and offering both.

Calculation

# Chapter 6. Pay per question, or an annual pass?
# Each user has questions at a personal yearly rate lam, gamma-distributed
# with mean 3 and shape k (small k: a few heavy askers and many light
# ones). Each question is worth v, log-normal with median 6, log-sd 0.6.
# Over a year a user with rate lam asks Poisson(lam) questions.
#   per question  price p per answer; a question is bought if v >= p
#   annual pass   price P for unlimited answers; bought if the user's
#                 expected yearly value lam E[v] is at least P
#   both          the user takes whichever leaves more expected surplus
# Prices are chosen by grid search; 20,000 users; revenue per user-year.
import math
import random

def users(k, rng, n=20000):
    out = []
    for _ in range(n):
        lam = rng.gammavariate(k, 3 / k)
        q = sum(1 for _ in range(200) if rng.random() < lam / 200)
        out.append((lam, [6 * math.exp(rng.gauss(0, 0.6)) for _ in range(q)]))
    return out

EV = 6 * math.exp(0.6 ** 2 / 2)

def per_q(us, p):
    return sum(p for _, vs in us for v in vs if v >= p) / len(us)

def both(us, p, pp):
    rev = 0.0
    for lam, vs in us:
        s_q = sum(max(v - p, 0) for v in vs)
        s_pass = lam * EV - pp
        if s_pass > s_q and s_pass >= 0:
            rev += pp
        else:
            rev += sum(p for v in vs if v >= p)
    return rev / len(us)

if __name__ == "__main__":
    rng = random.Random(63013)
    pq_grid = [3 + 0.5 * i for i in range(17)]
    pass_grid = [5 * i for i in range(1, 21)]
    print("   shape k   per question   annual pass   both offered")
    for k in (0.5, 2.0, 10.0):
        us = users(k, rng)
        best_q = max((per_q(us, p), p) for p in pq_grid)
        best_pass = max((sum(pp for lam, _ in us if lam * EV >= pp)
                         / len(us), pp) for pp in pass_grid)
        best_both = max((both(us, p, pp), p, pp)
                        for p in pq_grid[::2] for pp in pass_grid[::2])
        print("    {:4.1f}     {:6.2f} at {:4.1f}  {:6.2f} at {:3d}"
              "   {:6.2f} at {:4.1f} / {:3d}".format(
                  k, best_q[0], best_q[1], best_pass[0], best_pass[1],
                  best_both[0], best_both[1], best_both[2]))
    print("\nReading: when users ask at similar rates (large k) a pass")
    print("earns more than pricing each answer; when a few ask a great deal")
    print("(small k) pricing each question does better, and offering both")
    print("lets light askers pay per answer while heavy askers buy the pass.")
#    shape k   per question   annual pass   both offered
#      0.5       9.17 at  5.0    7.16 at  30     9.74 at  5.0 /  65
#      2.0       9.30 at  5.0    8.96 at  20    11.08 at  6.0 /  25
#     10.0       9.37 at  5.0   12.59 at  15    13.36 at  7.0 /  15
#
# Reading: when users ask at similar rates (large k) a pass
# earns more than pricing each answer; when a few ask a great deal
# (small k) pricing each question does better, and offering both
# lets light askers pay per answer while heavy askers buy the pass.

When people ask at similar rates (k=10.0k = 10.0) the annual pass earns more (12.59 against 9.37); when a few ask a very great deal (k=0.5k = 0.5) paying per question does better (9.17 against 7.16). Offering both is best in all three cases: light users pay per question and heavy users buy the pass, giving 11.08 at k=2.0k = 2.0, for example. This matches the conclusion of Book 58: the more dispersed the usage, the more a single yearly price lets heavy users take value away and drives light users off.

Note

Paying per question. The models of this chapter describe only how price structure affects revenue. A design that uses fading curiosity or anxiety to push users to pay again and again for the same matter may breach local consumer-protection rules; the compliance checklist in Chapter 8 lists what needs checking.

Summary

Bundling independent questions collects a growing share of value: 0.678 at 20 questions, while separate pricing stays at about 0.362. A decoy tier without an attraction effect raises revenue only because some people buy the dominated option, and the attraction effect itself is unstable in experiments and must be measured. Whether paying per question or an annual pass is better depends on how dispersed asking rates are; offering both is best in every case.

References

Adams, William James, and Janet L. Yellen. 1976. “Commodity Bundling and the Burden of Monopoly.” Quarterly Journal of Economics 90 (3): 475–98.
Ariely, Dan, George Loewenstein, and Drazen Prelec. 2003. “”Coherent Arbitrariness”: Stable Demand Curves without Stable Preferences.” Quarterly Journal of Economics 118 (1): 73–106.
Bakos, Yannis, and Erik Brynjolfsson. 1999. “Bundling Information Goods: Pricing, Profits, and Efficiency.” Management Science 45 (12): 1613–30.
Frederick, Shane, Leonard Lee, and Ernest Baskin. 2014. “The Limits of Attraction.” Journal of Marketing Research 51 (4): 487–507.
Huber, Joel, John W. Payne, and Christopher Puto. 1982. “Adding Asymmetrically Dominated Alternatives: Violations of Regularity and the Similarity Hypothesis.” Journal of Consumer Research 9 (1): 90–98.

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).