Chapter 6: Online Platforms

Algorithmic Game Theory
Two-sided marketsNetwork effects and critical massPlatform pricingSponsored search: GSP and VCGReputation systems
Sercan KÜLCÜ · Giresun University · Computer Engineering

Platforms and Markets

Warm-up Game: The Sponsored Search Auction

The rules

Advertisers bid per click; ads are ranked into three slots by bid × quality score, and higher slots get more clicks.

Think about it
  • If a click earns you 8 TL, is bidding 8 TL best?
  • Does the highest bidder always get the top slot?
cheap flightsAd · slot 1rank: bid × quality12%Ad · slot 2rank: bid × quality7%Ad · slot 3rank: bid × quality3.5%Bar: the slot's click-through rate (Search Ads scenario)

Platforms and Markets

Learning Outcomes

Modeling platforms

Describe two-sided markets, cross-side network effects, and the platform as a mechanism designer.

Computing critical mass

Find the equilibria, the threshold and the “tipping” dynamics in a simple adoption model.

Explaining price structure

Show in a two-sided numerical example which side is subsidized, and why.

Ad auctions

Compute GSP and VCG payments from the same bids; show why truthful bidding is not an equilibrium of GSP.

Measuring reputation

Rank with Bayesian averages; compute when fake reviews pay off.

Discussing competition

Interpret how tipping, switching costs, multi-homing and regulation (the EU Digital Markets Act, DMA) affect platform competition.

Platforms and Markets

What Is an Online Platform?

  • An online platform: a digital ecosystem (an intermediary) that brings users, sellers and service providers together.
  • The value comes less from the platform itself than from the sides' access to each other: a two-sided market.
  • The platform sets the rules: who joins, who sees whom, who pays what → a mechanism.
TypeSide 1Side 2Examples
E-commerce marketplacebuyerssellersTrendyol, Hepsiburada, Amazon
Sharing economyguests / ridershosts / driversAirbnb, BlaBlaCar, Uber
Social mediausersadvertisers, creatorsInstagram, TikTok, Facebook
Search enginesearchersadvertisersGoogle, Bing

Platforms and Markets

Two-Sided Markets and Cross-Side Network Effects

Buyers ASellers BPlatformeach buyer adds value for sellers (eB)each seller adds value for buyers (eA)pApB
Cross-side network effect

The number of participants on one side determines the value the other side gets (more sellers → more choice for buyers).

Price structure

The platform chooses not only the total price but how the fee is split between the sides (pA, pB).

The chicken-and-egg problem

No sellers, no buyers; no buyers, no sellers: one side has to be “pulled in” first.

Platforms and Markets

The Platform Is a Mechanism, Users Are Players

  • Platform design (fees, ranking and rating rules) steers users' strategic behavior.
  • Users choose strategies according to incentives; the resulting behavior is a Nash equilibrium under the rules.

The platform's incentives

  • Bonuses: extra pay for drivers in busy areas (Uber)
  • Discounts, coupons, free trials
  • Ranking and recommendation algorithms → visibility
  • Ratings and badges (trusted seller)

Users' strategic responses

  • Fake reviews and rating inflation
  • Multiple accounts, bonus abuse
  • Gaming the algorithm (click farms)
  • Retaliatory negative reviews
Right incentives → cooperation and efficiency. Wrong incentives → manipulation and loss of trust.

Network Effects

Types of Network Effects

Direct (same side)

WhatsApp: the more of your friends use it, the more valuable it is to you.

Indirect / cross-side

Uber: more drivers → shorter waits → more riders → more work for drivers.

Negative

Crowding → congestion, information overload; sellers on the same side compete with each other.

Local

Value comes from your neighborhood (city, circle of friends), not from all users; competition is won locally.

A positive network effect creates a feedback loop: winners grow, losers shrink. The market can “tip” one way.

Network Effects

A Simple Adoption Model

  • N = 1000 potential users; user i's type θi ~ U[0, 1] measures how much they care about the platform.
  • If a fraction x of the population joins, i's utility is θi·x − p, where p = 0.16 (a fee or the effort of joining).
  • A user joins ⟺ θi·x ≥ p ⟺ θi ≥ p/x. Since θ is uniform on [0, 1], the fraction with θ above p/x is 1 − p/x.
expect x = 0.5 → threshold θ* = p/x = 0.32stay outjoin: 0.6801θ ~ U[0, 1]expect x = 0.25 → threshold θ* = p/x = 0.64stay outjoin: 0.3601θ ~ U[0, 1]
Fraction joining if x is expected: F(x) = 1 − p/x  (0 if x ≤ p)
  • Equilibrium: a self-fulfilling expectation, F(x) = x.
  • Dynamics: this period's adopters set next period's expectation, xt+1 = F(xt).

A simplified version of the network-effects model in Easley and Kleinberg (2010, Chapter 17).

Network Effects

Step by Step: Critical Mass and Tipping

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x
tx₀ = 0.25x₀ = 0.19
0 
1 
2 
3 
4 
5 
6 

The curve F(x) = 1 − 0.16/x: the fraction who actually join when “a fraction x will join” is expected. Nobody joins while x ≤ 0.16.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
0 
1 
2 
3 
4 
5 
6 

Equilibria satisfy F(x) = x: multiplying 1 − 0.16/x = x by x gives x² − x + 0.16 = 0 → x = 0.2 and x = 0.8; also x = 0 (nobody). The hollow point is an unstable equilibrium: a small deviation pushes the dynamics away from it.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.25
10.36
2 
3 
4 
5 
6 

250 users are expected (x₀ = 0.25): those with θ ≥ 0.16/0.25 = 0.64 join → x₁ = 0.36.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.25
10.36
20.556
3 
4 
5 
6 

Adoption beat expectations; with the new expectation 0.36, those with θ ≥ 0.444 join → x₂ = 0.556.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.25
10.36
20.556
30.712
4 
5 
6 

The feedback continues: x₃ = 1 − 0.16/0.556 = 0.712.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.25
10.36
20.556
30.712
40.775
5 
6 

x₄ = 0.775; the steps shrink as the path approaches the stable equilibrium.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.25
10.36
20.556
30.712
40.775
50.794
60.798

x₆ ≈ 0.798 → 0.8: the market has tipped in the platform's favor (800 users).

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.250.19
10.360.158
20.556
30.712
40.775
50.794
60.798

Start just below the threshold: x₀ = 0.19 → x₁ = 1 − 0.16/0.19 = 0.158. Adoption fell short of expectations.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.250.19
10.360.158
20.5560
30.712
40.775
50.794
60.798

x₁ = 0.158 ≤ p = 0.16 → x₂ = 0: nobody joins and the platform collapses.

000.20.20.40.40.60.60.80.811expected fraction xfraction joining F(x)F(x) = xF(x) = 1 − 0.16/x0.2 threshold0.8
tx₀ = 0.25x₀ = 0.19
00.250.19
10.360.158
20.5560
30.712
40.775
50.794
60.798

Result: 0 and 0.8 are stable, 0.2 is unstable. Critical mass is 0.2 (200 users): above it the market goes to 0.8, below it to 0.

Network Effects

Reaching Critical Mass: Design Lessons

pequilibriathreshold
0.090, 0.1, 0.90.1
0.160, 0.2, 0.80.2
0.210, 0.3, 0.70.3
0.250 and 0.5 (tangent)0.5
0.3only 0—

A lower fee means a lower threshold. For x² − x + p = 0 to have real roots, the discriminant 1 − 4p ≥ 0, i.e. p ≤ 0.25; if p > 0.25 the market never forms.

Launch subsidies

Free periods, bonuses: p falls, and so does the critical mass.

Seeding

The platform supplies the first content, sellers or drivers itself.

Managing expectations

An “everyone's here” perception lifts x₀ above the threshold.

Standalone value

Start with a tool that is useful without the network; build the network later.

Example: WhatsApp's value grew rapidly as users joined; Friendster collapsed through the same loop in reverse once users started leaving.

Platform Pricing

Principles of Two-Sided Pricing

  • Besides the overall price level, the platform chooses the price structure: which side pays, and how much? (Rochet and Tirole, 2003)
  • The side that adds more value to the other side, or is more price-sensitive, gets a lower price — even a subsidy (a negative price).
  • The loss is recovered from fees on the other side, which the subsidized side attracts.
Search / social media

Users are free; advertisers pay.

Credit cards

Rewards for cardholders, fees from merchants.

Marketplaces

Coupons for buyers, commissions from sellers.

Revenue models

Commissions, subscriptions, ads, per-transaction fees.

In a simple linear model (Armstrong, 2006) the rule is: pA = markup − the value one A participant adds to side B. If the second term is large, the price turns negative. Markup: the gap between cost and the price the platform would charge a participant without network effects.

Platform Pricing

A Numerical Example: A Marketplace Model

BuyersnA, pASellersnB, pBeB = 0.75eA = 0.25platform
  • n: participants (thousands), p: monthly fee (TL). Serving one more participant costs the platform nothing (marginal cost 0).
  • Buyers are very valuable to sellers (eB = 0.75); seller variety matters less to buyers (eA = 0.25).
nA = 20 + 0.25·nB − pA
nB = 80 + 0.75·nA − pB
Π = pA·nA + pB·nB

Question: which (pA, pB) maximizes profit? Which side is subsidized?

Platform Pricing

Step by Step: Which Side Gets Subsidized?

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
 
 
 
 

Choosing participation instead of prices simplifies the algebra: each (nA, nB) pair corresponds to exactly one price pair.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
 
 
 
 

Substitute into profit. The nAnB term shows that the platform can capture the value the two sides add to each other.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
 
 
 
 

Hold nB fixed and differentiate Π with respect to nA (a partial derivative, written ∂); at the optimum both derivatives are 0: two linear equations.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
Optimum0.25 / 0.75−55040602800
 
 
 

Solution: buyers are paid 5 TL (pA = −5), sellers pay 50 TL. Profit Π = −200 + 3000 = 2800.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
Optimum0.25 / 0.75−55040602800
Free for buyers0.25 / 0.75047.534.658.52776.9
 
 

Compare: if buyers are free (pA = 0), the best pB ≈ 47.5; fewer buyers, profit 2776.9 < 2800.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
Optimum0.25 / 0.75−55040602800
Free for buyers0.25 / 0.75047.534.658.52776.9
Buyers pay 10 TL0.25 / 0.751042.523.855.42592.3
 

Charging buyers 10 TL cuts them to 23.8 thousand, and sellers shrink too: profit 2592.3.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
Optimum0.25 / 0.75−55040602800
Free for buyers0.25 / 0.75047.534.658.52776.9
Buyers pay 10 TL0.25 / 0.751042.523.855.42592.3
 

Why? One extra buyer adds 0.75·60 = 45 TL of value for sellers, which the platform recovers through pB. 45 > 40 (the markup) → subsidy.

1. Make participation the decision variable: pA = 20 + 0.25 nB − nA,  pB = 80 + 0.75 nA − nB
2. Π = 20 nA + 80 nB + (0.25 + 0.75) nAnB − nA² − nB²
3. ∂Π/∂nA = 20 + nB − 2nA = 0,   ∂Π/∂nB = 80 + nA − 2nB = 0
4. nA = 40, nB = 60 → pA = 20 + 15 − 40 = −5, pB = 80 + 30 − 60 = 50
5. Rule: pA = nA − eB·nB = 40 − 0.75·60 = −5  (markup nA − the value a buyer adds for sellers)
ScenarioeA / eBpApBnAnBΠ
Optimum0.25 / 0.75−55040602800
Free for buyers0.25 / 0.75047.534.658.52776.9
Buyers pay 10 TL0.25 / 0.751042.523.855.42592.3
Effects reversed0.75 / 0.25253040602800

Counterfactual: with the cross effects reversed (buyers adding little value for sellers), pA = 25 and pB = 30; the subsidy disappears.

A script scanned the (pA, pB) grid in 0.5 TL steps: the profit peaks at (−5, 50).

Platform Pricing

Code Trace: participation(pA, pB)

0204060012345678round tthousands4060nA = 0nB = 0pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Call with the optimal prices pA = −5, pB = 50. Instead of solving for participation “by hand”, let the sides respond to each other.

0204060012345678round tthousands4060nA = 0nB = 0pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Start: the platform is empty, nA = nB = 0.

0204060012345678round tthousands4060nA = 25nB = 0pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 1: buyers respond to the current number of sellers: nA = 20 + 0.25·0 + 5 = 25.

0204060012345678round tthousands4060nA = 25nB = 48.8pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Sellers respond to the new number of buyers: nB = 80 + 0.75·25 − 50 = 48.8. The buyer subsidy drew in sellers.

0204060012345678round tthousands4060nA = 37.2nB = 48.8pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 2: buyers respond to the current number of sellers: nA = 20 + 0.25·48.8 + 5 = 37.2.

0204060012345678round tthousands4060nA = 37.2nB = 57.9pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Sellers respond to the new number of buyers: nB = 80 + 0.75·37.2 − 50 = 57.9. The cross effect grows both sides together.

0204060012345678round tthousands4060nA = 39.5nB = 57.9pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 3: buyers respond to the current number of sellers: nA = 20 + 0.25·57.9 + 5 = 39.5.

0204060012345678round tthousands4060nA = 39.5nB = 59.6pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Sellers respond to the new number of buyers: nB = 80 + 0.75·39.5 − 50 = 59.6. The cross effect grows both sides together.

0204060012345678round tthousands4060nA = 39.9nB = 59.9pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 4: nA = 39.9, nB = 59.9. The change shrinks by a factor of ≈ 0.25·0.75 ≈ 0.19 each round.

0204060012345678round tthousands4060nA = 40nB = 60pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 5: nA = 40, nB = 60. The change shrinks by a factor of ≈ 0.25·0.75 ≈ 0.19 each round.

0204060012345678round tthousands4060nA = 40nB = 60pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 6: nA = 40, nB = 60. The change shrinks by a factor of ≈ 0.25·0.75 ≈ 0.19 each round.

0204060012345678round tthousands4060nA = 40nB = 60pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 7: nA = 40, nB = 60. The change shrinks by a factor of ≈ 0.25·0.75 ≈ 0.19 each round.

0204060012345678round tthousands4060nA = 40nB = 60pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Round 8: nA = 40, nB = 60. The change shrinks by a factor of ≈ 0.25·0.75 ≈ 0.19 each round.

0204060012345678round tthousands4060nA = 40nB = 60pA = −5 (buyers are paid)pB = 50dashed: fixed point (40, 60)

Fixed point (40, 60). Profit = −5·40 + 50·60 = 2,800. This is the participation equilibrium where each side's expectations confirm the other's.

def participation(pA, pB, rounds=8):
  nA, nB = 0, 0
  for t in range(rounds):
    nA = 20 + 0.25 * nB - pA   # buyers
    nB = 80 + 0.75 * nA - pB   # sellers
  return nA, nB

Convergence condition: the product of the cross effects eA·eB = 0.1875 < 1. If it exceeded 1, participation would explode or collapse: the network effect is too strong (tipping).

Platform Pricing

A Richer Model: Two-Sided Platform Pricing

The same model, two more parameters

Model: nA = baseA − sA·pA + eA·nB, nB = baseB − sB·pB + eB·nA.

  • base: users who join at a zero fee when the other side is empty.
  • s: price sensitivity (marketplace: sA = 11, sB = 5); s = 1 in the numerical model.
  • Participation is capped at 0–100 users; fees are integers from −5 to 10.
  • Profit Π = pA·nA + pB·nB; participation is again the fixed point of mutual responses.
LeveleA / eBMost profitable (pA, pB)nA / nBΠ
10.10 / 0.30(2, 4)25.3 / 22.6140.8
20.15 / 0.60(0, 6)50.3 / 35.2211.0
30.20 / 0.85(−2, 9)78.3 / 56.6352.5

Marketplace scenario (base 45 / 35). At level 3, fees of (3, 3) earn only 167.0: subsidizing A and charging B a high fee more than doubles the profit.

The more side B values A (the larger eB), the lower A's fee, until it turns negative: the model's version of the rule pA = markup − the value A adds to side B.

Sponsored Search

Sponsored Search Auctions

search: car rentalAd · slot 1300 clicksAd · slot 2200 clicksAd · slot 3100 clicks
  • Every search query triggers an auction among the advertisers bidding on that keyword.
  • m slots with click counts α1 > α2 > … from top to bottom (here 300, 200, 100 per 1000 impressions).
  • Advertiser i values a click at vi, bids bi per click, and in slot j pays pj per click.
  • The utility of the advertiser in slot j: αj·(vi − pj).
Example: A, B, C, D value a click at 10, 8, 5, 2 TL (and bid that). Three slots, four advertisers.

Sponsored Search

Two Payment Rules: GSP and VCG

GSP (generalized second price)

Rank by bid (b(k): the k-th highest bid). The advertiser in slot j pays the next bid down per click:

pj = b(j+1)

With one slot it equals Vickrey; with several slots truthfulness is lost.

VCG

Same ranking. Each advertiser pays the loss its presence imposes on others:

αj·pj = Σk≥j (αk − αk+1)·b(k+1)

Truthful bidding is a dominant strategy.

GSP: per click, the next bid downslot 1A: 10pays 8slot 2B: 8pays 5slot 3C: 5pays 2—D: 2VCG: A's total payment = the losses it causes800B in slot 2, not 1: (300 − 200)·8500C in slot 3, not 2: (200 − 100)·5200D out, not in slot 3: (100 − 0)·2= 1.500
So under VCG, the advertiser in slot j pays the total click value lost by those below as each slides down one slot. αj·pj is the total payment, pj the per-click payment; αm+1 = 0.

Sponsored Search

Step by Step: GSP and VCG Payments from the Same Bids

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)
2200B (b = 8)
3100C (b = 5)
—0D (b = 2)
Revenue
Ranked by bid: A → slot 1, B → slot 2, C → slot 3; D is left out.

Both rules get the same bids (= values); only the payments differ.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)82400
2200B (b = 8)
3100C (b = 5)
—0D (b = 2)—0
Revenue
GSP: A pays B's bid, the one below: 8 TL/click × 300 = 2400.

Under GSP each slot pays the next bid down, per click.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)82400
2200B (b = 8)51000
3100C (b = 5)
—0D (b = 2)—0
Revenue
GSP: B pays C's bid: 5 × 200 = 1000.

The second slot.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)82400
2200B (b = 8)51000
3100C (b = 5)2200
—0D (b = 2)—0
Revenue3600
GSP: C pays D's bid: 2 × 100 = 200. GSP revenue is 3600.

GSP done.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)82400
2200B (b = 8)51000
3100C (b = 5)22002002
—0D (b = 2)—00—
Revenue3600
VCG, C: without C, D moves up to slot 3: welfare below is 200; with C it is 0 → C pays 200.

VCG payment = others' welfare “if it were absent” − others' welfare with it present. Those above are unaffected; only those below count.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)82400
2200B (b = 8)510007003.5
3100C (b = 5)22002002
—0D (b = 2)—00—
Revenue3600
VCG, B: without B, C takes slot 2 (200·5) and D slot 3 (100·2) → 1200; with B, C gets 100·5 = 500 → 700.

Removing B would move C up one slot and bring D into slot 3.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)8240015005
2200B (b = 8)510007003.5
3100C (b = 5)22002002
—0D (b = 2)—00—
Revenue36002400
VCG, A: without A, 300·8 + 200·5 + 100·2 = 3600; with A, 200·8 + 100·5 = 2100 → 1500.

The formula gives the same: (300−200)·8 + (200−100)·5 + 100·2 = 1500.

SlotClicksAdvertiserGSP /clickGSP totalVCG totalVCG /click
1300A (b = 10)8240015005
2200B (b = 8)510007003.5
3100C (b = 5)22002002
—0D (b = 2)—00—
Revenue36002400
With the same bids, GSP charges at least as much as VCG in every slot: 8 ≥ 5, 5 ≥ 3.5, 2 ≥ 2 TL/click.

Revenue: GSP 3600, VCG 2400. But the comparison is unfair: under GSP the bids change.

Sponsored Search

Code Trace: gsp_vcg(bids, ctr)

b0123456agspvcgb: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG payment

m = 3 slots.

b100815223040506agspvcgb: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG payment

Sort bids from high to low and append m zeros so that a “next bid down” always exists.

b100815223040506a3002001000gspvcgb: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG payment

Append a dummy (m+1)-th slot with 0 clicks to a: αm+1 = 0.

b100815223040506a3002001000gsp000vcg0000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG payment

Zero the result arrays; vcg's extra last element is the recursion's starting value.

b100815223040506a3002001000gsp240000vcg0000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 0

gsp[0] = a[0]·b[1] = 300·8 = 2400: slot 1 pays the next bid down, per click.

b100815223040506a3002001000gsp240010000vcg0000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 1

gsp[1] = a[1]·b[2] = 200·5 = 1000: slot 2 pays the next bid down, per click.

b100815223040506a3002001000gsp24001000200vcg0000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 2

gsp[2] = a[2]·b[3] = 100·2 = 200: slot 3 pays the next bid down, per click.

b100815223040506a3002001000gsp24001000200vcg0000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 2

extra = (100 − 0)·2 = 200: the click value lost by the advertiser below, stuck in slot 4 instead of 3.

b100815223040506a3002001000gsp24001000200vcg002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 2

vcg[2] = 0 + 200 = 200: losses imposed further down carry upward too.

b100815223040506a3002001000gsp24001000200vcg002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 1

extra = (200 − 100)·5 = 500: the click value lost by the advertiser below, stuck in slot 3 instead of 2.

b100815223040506a3002001000gsp24001000200vcg07002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 1

vcg[1] = 200 + 500 = 700: losses imposed further down carry upward too.

b100815223040506a3002001000gsp24001000200vcg07002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 0

extra = (300 − 200)·8 = 800: the click value lost by the advertiser below, stuck in slot 2 instead of 1.

b100815223040506a3002001000gsp24001000200vcg15007002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG paymentj = 0

vcg[0] = 700 + 800 = 1500: losses imposed further down carry upward too.

b100815223040506a3002001000gsp24001000200vcg15007002000b: sorted bids (TL/click)a: slot clicksgsp[j]: total GSP paymentvcg[j]: total VCG payment

Result: GSP [2400, 1000, 200] (total 3600), VCG [1500, 700, 200] (total 2400).

def gsp_vcg(bids, ctr):
  m = len(ctr)
  b = sorted(bids, reverse=True) + [0]*m
  a = ctr + [0]     # dummy slot: 0 clicks
  gsp, vcg = [0]*m, [0]*(m + 1)
  for j in range(m):
    gsp[j] = a[j] * b[j+1]   # next bid
  for j in range(m-1, -1, -1):  # bottom up
    extra = (a[j] - a[j+1]) * b[j+1]
    vcg[j] = vcg[j+1] + extra
  return gsp, vcg[:m]

Sponsored Search

Step by Step: Truthful Bidding Is Not an Equilibrium of GSP

Adv.vbidslotpays/clickutility
A101018600
B8825600
C5532300
D22——0
Platform revenue3600
024681012040080012001600A's bid (TL/click)A's utilitybid 10

Everyone bids their value. A is in slot 1 and pays B's bid (8): utility 300·(10 − 8) = 600.

Adv.vbidslotpays/clickutility
A101018600
B8825600
C5532300
D22——0
Platform revenue3600
024681012040080012001600A's bid (TL/click)A's utility1000bid 10

A's utility is a step function of its bid: 2–5 gives slot 3 (800), 5–8 gives slot 2 (1000), above 8 slot 1 (600).

Adv.vbidslotpays/clickutility
B8816600
A106251000
C5532300
D22——0
Platform revenue3000
024681012040080012001600A's bid (TL/click)A's utility1000bid 6

If A bids 6, it drops to slot 2 and pays C's 5: 200·(10 − 5) = 1000 > 600. Truthful bidding is not a best response → not an equilibrium.

Adv.vbidslotpays/clickutility
A1010151500
B8523.5900
C53.532300
D22——0
Platform revenue2400
024681012040080012001600A's bid (TL/click)A's utility1500bid 10

An equilibrium: B bids 5, C 3.5, D 2. A's best is still slot 1 (1500); a script tried every deviation, and nobody gains.

Adv.vbidslotpays/clickutility
A1010151500
B8523.5900
C53.532300
D22——0
Platform revenue2400
024681012040080012001600A's bid (TL/click)A's utility1500bid 10

In this equilibrium the per-click payments 5 / 3.5 / 2 are the VCG payments, revenue 2400. It is the lowest-revenue locally envy-free equilibrium of GSP. Locally envy-free: no advertiser would rather take the slot just above at that slot's current per-click price.

Sponsored Search

Code Trace: profitable_dev(i, b, v, ctr, cands)

ranking under current bidsslot 1A: 10slot 2B: 8slot 3C: 5outD: 2v_A = 10 · clicks 300/200/100

Truthful bids b = v = [10, 8, 5, 2]. Should A (i = 0) deviate? Candidate bids 9, 6, 4, 3, 1.

ranking under current bidsslot 1A: 10slot 2B: 8slot 3C: 5outD: 2v_A = 10 · clicks 300/200/100old utility = 600

Currently A is in slot 1 and pays the next bid down (8): old = 300·(10 − 8) = 600.

ranking under trial bids b2slot 1A: 9slot 2B: 8slot 3C: 5outD: 2v_A = 10 · clicks 300/200/100old utility = 600x = 9 → slot 1, u = 600u > old fails → continue

x = 9: slot 1, u = 600 ≤ 600.

ranking under trial bids b2slot 1B: 8slot 2A: 6slot 3C: 5outD: 2v_A = 10 · clicks 300/200/100old utility = 600x = 6 → slot 2, u = 1,000u > old → profitable deviation!return: (6, 1,000)

x = 6: A drops to slot 2 and pays less: u = 1,000 > 600 → profitable deviation found; return (x, u).

ranking under current bidsslot 1A: 10slot 2B: 5slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100

Now the proposed equilibrium: B bids 5, C 3.5, D 2. Same candidate list for A.

ranking under current bidsslot 1A: 10slot 2B: 5slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500

Currently A is in slot 1 and pays the next bid down (5): old = 300·(10 − 5) = 1,500.

ranking under trial bids b2slot 1A: 9slot 2B: 5slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500x = 9 → slot 1, u = 1,500u > old fails → continue

x = 9: slot 1, u = 1,500 ≤ 1,500.

ranking under trial bids b2slot 1A: 6slot 2B: 5slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500x = 6 → slot 1, u = 1,500u > old fails → continue

x = 6: slot 1, u = 1,500 ≤ 1,500.

ranking under trial bids b2slot 1B: 5slot 2A: 4slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500x = 4 → slot 2, u = 1,300u > old fails → continue

x = 4: slot 2, u = 1,300 ≤ 1,500.

ranking under trial bids b2slot 1B: 5slot 2C: 3.5slot 3A: 3outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500x = 3 → slot 3, u = 800u > old fails → continue

x = 3: slot 3, u = 800 ≤ 1,500.

ranking under trial bids b2slot 1B: 5slot 2C: 3.5slot 3D: 2outA: 1v_A = 10 · clicks 300/200/100old utility = 1,500x = 1 → no slot, u = 0u > old fails → continue

x = 1: the ad is left out, u = 0 ≤ 1,500.

ranking under current bidsslot 1A: 10slot 2B: 5slot 3C: 3.5outD: 2v_A = 10 · clicks 300/200/100old utility = 1,500return: None

No candidate is profitable → None. A script ran the same test for all four advertisers on a bid grid with step 0.25: nobody gains, so this profile is a Nash equilibrium.

def profitable_dev(i, b, v, ctr, cands):
  old = gsp_utility(i, b, v, ctr)
  for x in cands:
    b2 = b[:]; b2[i] = x   # only i changes
    u = gsp_utility(i, b2, v, ctr)
    if u > old:
      return x, u   # profitable
  return None     # no deviation

gsp_utility(i, b, v, ctr): sort the bids; if i is in slot j, ctr[j]·(v[i] − the next bid down), and 0 if it gets no slot. Bids are assumed distinct.

Sponsored Search

GSP and VCG Compared

GSPVCG
Is truthful bidding dominant?No (A bids 6 instead of 10)Yes
Revenue with the same bids3600 (≥ VCG in every slot)2400
Revenue in equilibrium≥ 2400 in locally envy-free equilibria; the lowest equals VCG2400 (with truthful bids)
Explaining itEasy: “the next bid down”A sum of externalities, harder
In practiceLong the norm in search ads (with quality scores)Used in Facebook's ad auction (Varian and Harris, 2014)
Quality score

In practice, ranking uses the ad's click quality (its quality score) together with the bid.

Lesson

The name “second price” does not guarantee truthfulness: what holds for one item can break with many slots.

Sponsored Search

A Given Bid Profile ≠ Equilibrium Bids

GSP outcomes are computed for given bids. 3600 is the revenue when everyone bids truthfully — but truthful bidding is not a dominant strategy in GSP. Click counts 300, 200, 100; values A 10, B 8, C 5, D 2.

ProfileBids (A, B, C, D)A's slot and utilityPlatform revenue
Truthful10, 8, 5, 2slot 1: 300·(10 − 8) = 6003600
A bids lower6, 8, 5, 2slot 2: 200·(10 − 5) = 10003000
Locally envy-free equilibrium10, 5, 3.5, 2slot 1: 300·(10 − 5) = 15002400 = VCG
  • In a lower slot A gets fewer clicks but pays less per click: 1000 > 600.
  • Locally envy-free equilibrium (Edelman, Ostrovsky, Schwarz, 2007): no advertiser wants the slot just above at that slot's current price. B in slot 2 gets 200·(8 − 3.5) = 900; taking slot 1 at 5 would give 300·(8 − 5) = 900: no gain.
In these equilibria GSP revenue is at least 2400, and the lowest equals VCG. The 3600 from truthful bids is not an equilibrium revenue.

Sponsored Search

Budgets and Online Allocation

  • Advertisers have daily budgets; searches arrive one by one, and the platform must assign each to an advertiser immediately, without knowing the future (an online decision).
  • Example: A and B each have a 100 TL budget. Both bid 1 TL on “running shoes”; only A bids 1 TL on “gym bag”. First 100 shoe searches arrive, then 100 bag searches.
RuleShoe searchesBag searchesRevenue
Greedy (highest bid, ties to A)all 100 to A; A's budget runs outA has no money left: 0100
Balance: to the larger remaining budgetA 50, B 50A's remaining 50150
Optimum knowing the futureall 100 to Ball 100 to A200
Common mistake

“Always picking the highest bid is best.” With budgets, the greedy rule exhausts the flexible advertiser early; in the worst case it guarantees only 1/2 of the optimum.

Guarantee

When bids are small relative to budgets, the MSVV rule — weighting bids by remaining budget — earns at least 1 − 1/e ≈ 63% of the optimum in any search order (Mehta, Saberi, Vazirani and Vazirani, 2007).

Sponsored Search

Step by Step: GSP with Quality Scores

bid b (TL/click)AA: 10BB: 8CC: 5DD: 2

The same four advertisers. Clicks depend on slot and ad quality: i gets αj·qi clicks in slot j (α = 300, 200, 100).

score b·qAA: 4 (b = 10, q = 0.4)BB: 6.4 (b = 8, q = 0.8)CC: 3 (b = 5, q = 0.6)DD: 2 (b = 2, q = 1)

Quality q: how likely the ad is to be clicked (e.g. A 0.4, B 0.8). The platform ranks by the score b·q: expected revenue per impression.

score b·qslot 1B: 6.4 (b = 8, q = 0.8)slot 2A: 4 (b = 10, q = 0.4)slot 3C: 3 (b = 5, q = 0.6)—D: 2 (b = 2, q = 1)

Ranked by score: B (6.4) > A (4) > C (3) > D (2). A, the highest bidder, is no longer first.

score b·qslot 1B: 6.4 (b = 8, q = 0.8)pays 4 / 0.8 = 5slot 2A: 4 (b = 10, q = 0.4)pays 3 / 0.4 = 7.5slot 3C: 3 (b = 5, q = 0.6)pays 2 / 0.6 = 3.33—D: 2 (b = 2, q = 1)

Payment: the lowest bid that keeps the position. biqi ≥ the score below ⟺ bi ≥ scorej+1/qi.

score b·qslot 1B: 6.4 (b = 8, q = 0.8)pays 4 / 0.8 = 5≤ 8 ✓slot 2A: 4 (b = 10, q = 0.4)pays 3 / 0.4 = 7.5≤ 10 ✓slot 3C: 3 (b = 5, q = 0.6)pays 2 / 0.6 = 3.33≤ 5 ✓—D: 2 (b = 2, q = 1)

No payment exceeds the bid (5 ≤ 8, 7.5 ≤ 10, 3.33 ≤ 5). A high-quality ad gets the same slot for less: quality is rewarded.

Rankingslots 1–3total clicks (per 1000 impressions)revenue (TL)
By bid onlyA, B, C3401,880
With quality scores (b·q)B, A, C3802,000

Here quality scores raised both clicks (340 → 380) and revenue (1,880 → 2,000): users see relevant ads, and the platform sells more clicks.

Sponsored Search

Back to the Game: The Sponsored Search Auction

Theory's answer: no

With the “Search Ads” rates (12%, 7%, 3.5%), 1000 impressions give 120, 70, 35 clicks. Your click value is 8, rival bids 6, 4, 2, all qualities 1; GSP:

BidSlotPayment/clickUtility
8 (truthful)16120 × 2 = 240
52470 × 4 = 280
33235 × 6 = 210

Dropping to a lower slot pays more; a best-response search finds exactly this deviation.

In practice

Under first-price ad auctions, bids repeatedly ratcheted up and crashed (Edelman and Ostrovsky, 2007). Under the GSP used by Google, truthful bidding is not dominant either (Edelman, Ostrovsky and Schwarz, 2007).

The name “second price” brings truthfulness with one slot, not with many. The rule that guarantees truthfulness is VCG.

Reputation and Competition

Reputation Systems and Information Asymmetry

  • Buyers don't know a seller's quality in advance: information asymmetry. Buyers then offer an average price, good sellers exit, and only bad products remain (Akerlof's “market for lemons”).
  • A reputation system does two jobs: selection (find good sellers) and discipline (punish bad behavior in the future, a repeated game).
Fake reviews

Bought positive reviews, or negative ones for a rival.

Fear of retaliation

With mutual ratings, bad experiences go unreported.

Grade inflation

If everyone has 4.8+, ratings stop discriminating.

Selective participation

Moderately satisfied users stay silent; new sellers get no reviews at all (cold start).

Design measures: require verified purchases, down-weight old reviews, account for the uncertainty of few reviews (Bayesian averages), and reveal both sides' ratings simultaneously in mutual rating.

Reputation and Competition

Step by Step: 5.0 from a Few Reviews or 4.7 from Many?

ProductreviewsaveragerankBayesian scorerank
X251
Y2004.72
Z204.23
Ranked by raw average: X (5.0) > Y (4.7) > Z (4.2). But X has only 2 reviews!

The average of a product with few reviews is very noisy; two fake reviews put it on top.

ProductreviewsaveragerankBayesian scorerank
X2513.753
Y2004.724.641
Z204.233.972
Bayesian average: (C·m + Σ ratings)/(C + n), with m = 3.5 (the platform mean), C = 10, n: the product's number of reviews. X: (35 + 10)/12 = 3.75; Y: (35 + 940)/210 ≈ 4.64.

It is as if every product got 10 “average” virtual reviews: products with few reviews are pulled toward the mean. New ranking: Y > Z > X.

ProductreviewsaveragerankBayesian scorerank
X + 3 fake55142
Y2004.724.641
Z204.233.973
Adding 3 fake 5★ reviews to X keeps its raw average at 5.0 (still on top); its Bayesian score is (35 + 25)/15 = 4.0 → still below Y.

Bayesian averages blunt fake reviews: producing the same effect takes many more of them.

Reputation and Competition

Step by Step: When Do Fake Reviews Pay?

04812163.43.63.844.2top-rank threshold 4.0number of fake 5★ kraw
A seller: 20 reviews, average 3.6. Ratings ≥ 4.0 rank at the top → extra monthly profit G = 600 TL. A fake 5★ review costs 15 TL.
Raw average: (72 + 5k)/(20 + k) ≥ 4 ⟺ k ≥ 8. Cost 8 × 15 = 120 TL.
Detection probability ρ; if caught, loss L = 3000 TL. Profitable ⟺ 600 − 120 > ρ·3000 ⟺ ρ < 0.16.
With the Bayesian score (m = 3.5, C = 10): (107 + 5k)/(30 + k) ≥ 4 ⟺ k ≥ 13; cost 195 TL → ρ < 0.135.
Design levers: raise ρ (verified purchases, anomaly detection), raise L (account bans), lower G (robust scoring).

Model the fake review as an investment decision: return, cost and expected penalty.

04812163.43.63.844.2top-rank threshold 4.0number of fake 5★ krawk = 8
A seller: 20 reviews, average 3.6. Ratings ≥ 4.0 rank at the top → extra monthly profit G = 600 TL. A fake 5★ review costs 15 TL.
Raw average: (72 + 5k)/(20 + k) ≥ 4 ⟺ k ≥ 8. Cost 8 × 15 = 120 TL.
Detection probability ρ; if caught, loss L = 3000 TL. Profitable ⟺ 600 − 120 > ρ·3000 ⟺ ρ < 0.16.
With the Bayesian score (m = 3.5, C = 10): (107 + 5k)/(30 + k) ≥ 4 ⟺ k ≥ 13; cost 195 TL → ρ < 0.135.
Design levers: raise ρ (verified purchases, anomaly detection), raise L (account bans), lower G (robust scoring).

8 fake 5★ reviews lift the raw average to exactly 4.0.

04812163.43.63.844.2top-rank threshold 4.0number of fake 5★ krawk = 8
A seller: 20 reviews, average 3.6. Ratings ≥ 4.0 rank at the top → extra monthly profit G = 600 TL. A fake 5★ review costs 15 TL.
Raw average: (72 + 5k)/(20 + k) ≥ 4 ⟺ k ≥ 8. Cost 8 × 15 = 120 TL.
Detection probability ρ; if caught, loss L = 3000 TL. Profitable ⟺ 600 − 120 > ρ·3000 ⟺ ρ < 0.16.
With the Bayesian score (m = 3.5, C = 10): (107 + 5k)/(30 + k) ≥ 4 ⟺ k ≥ 13; cost 195 TL → ρ < 0.135.
Design levers: raise ρ (verified purchases, anomaly detection), raise L (account bans), lower G (robust scoring).

Expected gain 600 − 120 − ρ·3000. If the platform's detection probability is below 16%, cheating pays.

04812163.43.63.844.2top-rank threshold 4.0number of fake 5★ krawBayesiank = 8k = 13
A seller: 20 reviews, average 3.6. Ratings ≥ 4.0 rank at the top → extra monthly profit G = 600 TL. A fake 5★ review costs 15 TL.
Raw average: (72 + 5k)/(20 + k) ≥ 4 ⟺ k ≥ 8. Cost 8 × 15 = 120 TL.
Detection probability ρ; if caught, loss L = 3000 TL. Profitable ⟺ 600 − 120 > ρ·3000 ⟺ ρ < 0.16.
With the Bayesian score (m = 3.5, C = 10): (107 + 5k)/(30 + k) ≥ 4 ⟺ k ≥ 13; cost 195 TL → ρ < 0.135.
Design levers: raise ρ (verified purchases, anomaly detection), raise L (account bans), lower G (robust scoring).

Reaching the same threshold under the Bayesian score takes 13 fakes: cheating costs more, and the break-even drops to ρ < 0.135.

04812163.43.63.844.2top-rank threshold 4.0number of fake 5★ krawBayesiank = 8k = 13
A seller: 20 reviews, average 3.6. Ratings ≥ 4.0 rank at the top → extra monthly profit G = 600 TL. A fake 5★ review costs 15 TL.
Raw average: (72 + 5k)/(20 + k) ≥ 4 ⟺ k ≥ 8. Cost 8 × 15 = 120 TL.
Detection probability ρ; if caught, loss L = 3000 TL. Profitable ⟺ 600 − 120 > ρ·3000 ⟺ ρ < 0.16.
With the Bayesian score (m = 3.5, C = 10): (107 + 5k)/(30 + k) ≥ 4 ⟺ k ≥ 13; cost 195 TL → ρ < 0.135.
Design levers: raise ρ (verified purchases, anomaly detection), raise L (account bans), lower G (robust scoring).

The mechanism-design view: the platform doesn't just ban cheating, it pushes the expected return of cheating below zero.

Reputation and Competition

Platform Competition and Regulation

Competition
  • Winner takes all: through tipping, a single platform can capture most of the market.
  • Switching costs: friends, past ratings, data → keep users locked in.
  • Multi-homing: using several platforms at once (a driver on both Uber and a rival) weakens monopolization.
  • Examples: Hepsiburada – Trendyol, Uber – Lyft, Instagram – TikTok.
Ethics and regulation
  • Data privacy and security.
  • Monopolization risk and self-preferencing.
  • Transparency of ranking and recommendation algorithms.
  • Example: the EU Digital Markets Act (DMA) imposes obligations on large “gatekeeper” platforms.

Reputation and Competition

Recommender Systems, Lock-in and Privacy

  • Recommendation and ranking algorithms allocate visibility (who gets featured?): they are an allocation rule, and participants choose strategies accordingly.
The platform's objective

Maximizing short-term engagement or revenue can conflict with users' long-term interest (clickbait, infinite scroll).

Creators are players too

Creators and sellers compete for visibility and converge on whatever form the algorithm rewards (titles, tags): content can become uniform.

Loyalty programs

Loyalty points, paid memberships (e.g. fast shipping), gamification (badges, streaks): deliberately raise switching costs and reduce multi-homing.

Data and privacy

The price of a “free” service is data: more data means better recommendations but also risks of leaks and misuse. Users choose strategically how much to share.

The mechanism-design question is the same: are the platform's goals aligned with the interests of users and creators?

Reputation and Competition

Discussion: Growth, Behavior and Fairness

Why do some platforms grow fast while others collapse?

Critical mass and tipping: if expectations are above the threshold, feedback drives growth; below it, collapse. Ease of use, trust and early incentives help clear the threshold.

How do platforms shape behavior?

Bonuses and discounts boost specific behaviors (driver bonuses in busy areas); recommendation algorithms steer preferences; spam and fake-review policing limits manipulation.

Can a design be fair to everyone?

Perfect fairness is hard: preferences differ, network effects favor large groups, and players manipulate strategically. Incentives, rules and regulation can improve things.

Assessment

Summary

ConceptExplanationExamples
Network effects, tippingF(x) = 1 − p/x; two stable equilibria with a threshold between themp = 0.16: 0 and 0.8 stable, threshold 0.2
Two-sided pricingpA = nA − eBnB: the side that adds more value to the other pays lesspA = −5, pB = 50, Π = 2800
GSPSlot j pays the next bid down; truthful bidding is not dominantrevenue 3600 with the same bids; A bids 6 instead of 10
VCGPays the loss imposed on others; truthful bidding is dominant1500 / 700 / 200, revenue 2400
GSP equilibriumThe lowest locally envy-free equilibrium gives the VCG outcomebids 10, 5, 3.5, 2 → 5 / 3.5 / 2 TL/click
ReputationBayesian averages make fake reviews more expensivefake reviews needed: 8 → 13

Assessment

Sample Problems

  1. In the adoption model let p = 0.21. What are the equilibria and the critical mass?
  2. In the marketplace model let eA = eB = 0.5 (all else equal). What are the optimal prices? Is there a subsidy?
  3. Two slots (100 and 60 clicks), bids 6, 4, 1. Find the total GSP and VCG payments.
  4. In the GSP of problem 3, with the others truthful, should the advertiser valued at 6 bid truthfully?
  5. 10 reviews, average 4.0. How many fake 5★ reviews raise the raw average to 4.5?
  6. Two slots; A bids 6 (q = 0.5), B 4 (q = 1), C 1 (q = 1). In GSP with quality scores, what are the ranking and per-click payments?
Show answers
  1. x² − x + 0.21 = 0 → x = 0.3 and 0.7 (and 0). Critical mass 0.3; stable equilibria 0 and 0.7.
  2. nA = 40, nB = 60; pA = 40 − 0.5·60 = 10, pB = 60 − 0.5·40 = 40. No subsidy.
  3. GSP: 100·4 + 60·1 = 460. VCG: slot 2 → 60·1 = 60; slot 1 → (100 − 60)·4 + 60 = 220; total 280.
  4. No: truthful → slot 1, 100·(6 − 4) = 200; a bid between 1 and 4 → slot 2, 60·(6 − 1) = 300.
  5. (40 + 5k)/(10 + k) ≥ 4.5 ⟺ k ≥ 10.
  6. Scores A 3, B 4, C 1 → ranking B, A. B: 3/1 = 3 TL; A: 1/0.5 = 2 TL (both below their bids).

Assessment

References

  • J.-C. Rochet, J. Tirole, “Platform Competition in Two-Sided Markets”, Journal of the European Economic Association, 2003.
  • M. Armstrong, “Competition in Two-Sided Markets”, RAND Journal of Economics, 2006.
  • B. Edelman, M. Ostrovsky, M. Schwarz, “Internet Advertising and the Generalized Second-Price Auction”, American Economic Review, 2007.
  • H. Varian, “Position Auctions”, International Journal of Industrial Organization, 2007.
  • D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge University Press, 2010 (Chapters 15 and 17).
  • N. Nisan, T. Roughgarden, É. Tardos, V. Vazirani (eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (Chapters 27–28).
  • T. Roughgarden, Stanford CS364A: Algorithmic Game Theory lecture notes, 2013 (Lectures 2–3: sponsored search auctions).
  • G. Akerlof, “The Market for ‘Lemons’”, Quarterly Journal of Economics, 1970.
  • M. Luca, G. Zervas, “Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud”, Management Science, 2016.

Assessment

References: Ad Auctions

  • H. Varian, C. Harris, “The VCG Auction in Theory and Practice”, American Economic Review: Papers & Proceedings 104(5), 2014.
  • B. Edelman, M. Ostrovsky, “Strategic Bidder Behavior in Sponsored Search Auctions”, Decision Support Systems 43(1), 2007. · M. Ostrovsky, M. Schwarz, “Reserve Prices in Internet Advertising Auctions: A Field Experiment”, ACM EC, 2011.
  • A. Mehta, A. Saberi, U. Vazirani, V. Vazirani, “AdWords and Generalized Online Matching”, Journal of the ACM 54(5), 2007.

The End

Chapter 6: Online Platforms
Previous: Chapter 5: Auctions · Next: Chapter 7: Social Choice and Voting
1 / 1 Sercan KÜLCÜ. All rights reserved.