Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk

arXiv:2510.27112 · econ.TH · Submitted 2025-10-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Targeted Advertising Platforms".

Warren: Here's the gist of the paper. E-commerce platforms are rolling out ambitious targeted advertising initiatives that rely on merchants sharing customer data with each other via the platform The paper proposes a model of designing targeted advertising platforms that incentivizes merchants to voluntarily share customer data despite poaching concerns The optimal mechanism can be implemented through the design of three markets: i) selling market, where merchants can sell all their data at a posted price p, ii) exchange market, where merchants share all their data in exchange for high click-through rate (CTR) ads, and iii) buying market, where high-value merchants buy high CTR ads at the full price.

Ruth: Model Formulation and Objectives The designer's objective is to maximize a weighted sum of platform revenue, customer engagement and merchants’ surplus The objective function is given by ψη(θ, x, t):= ηvV (θ, x) + ηwW(x) + ηrR(t), for some non-negative welfare weight η = (ηv, ηw, ηr) such that ηv +ηw +ηr = 1 and ηr > 0 The designer faces incentive compatibility (IC) constraints: merchants must truthfully report their private information (θ) to implement (x F B(θ), t F B(θ)) Mechanism Design Framework The paper invokes the Revelation Principle and restricts the designer’s problem to direct mechanisms which are measurable maps given by targeted ads rule (henceforth, ads rule) x = (xi)i∈N such that xi = (x ω i)ω∈omega and transfer rule t = (ti)i∈N such that ti: Θ → R The designer’s problem is to design a campaign (x, t) that maximizes a weighted sum of platform revenue, customer engagement and merchants’ surplus subject to participation constraints Optimal Mechanism Characterization The main result is the existence of an optimal scoring mechanism (g, p) that attains the value Π(α, η) in (P) This mechanism involves a single merchant-specific scoring rule gi: Θi → R+ such that for given type profile θ, a customer ω ∈ omega is targeted by the merchant i with the highest weighted score ωigi(θi) The tie breaking rule p is chosen such that the critical type ˆθi(zi)’s IR constraint binds Large Market Simplification In sufficiently large platforms, the optimal mechanism can be implemented through the design of three markets: i) selling market, where merchants can sell all their data at a posted price pS = ϕ S η,i−1(1) per unit of CTR ii) exchange market, where merchants share all their data in exchange for high click-through rate (CTR) ads iii) buying market, where high-value merchants buy high CTR at the full price pB = 1 Asymptotic Efficiency The asymptotic relative efficiency (ARE) is measured by the ratio of the value of design Π∞(α, η) to the total surplus TS∞(α, ζ) The main source of inefficiency comes from the observation that merchants are compensated at a price pS per unit of clicks on their individual ads, i.e.

Warren: Eµi

pSωi: = pSEµi

ωi: , instead of at a price pS per unit of expected clicks on efficient ads, i.e.

Ruth: Eµi

pS maxj (ωj): = pSEµi

maxj (ωj): Applications and Extensions The model is broad in scope with applications in other market design settings like the greenhouse gas credit market and reallocating public resources The paper suggests new directions in combinatorial market exchange designs It also relates to the bundling literature since the optimal design trades data in bundles The model is applicable to other economic settings that involve trading of goods among a set of agents This paper proposes a model of targeted advertising platforms that host merchants who share their data and compete with one another for targeted ads In a large market, the optimal design simplifies to the design of three markets: a selling, a buying and an exchange market The designer can mitigate customer poaching by compensating merchants with either higher CTR ads or transfers The model's broader applicability in market design settings such as the GHG credit markets and reallocation of public resources is highlighted The paper concludes by describing the model’s scope and its applications in other market design settings like greenhouse gas credit market and reallocation of public resources References Abrache, J., T.

Paper summary: Ruth: So we're talking about this paper called "Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk" by Klajdi Hoxha and others. Basically, e-commerce platforms are starting to let merchants share customer data with each other on the platform, right? But the authors point out that current ways of setting up these platforms don't really stop merchants from just poaching customers away.

Warren: Yeah, they argue that this paper proposes a model for designing these targeted advertising platforms that actually gives merchants an incentive to voluntarily share their customer data even when they worry about losing their customers to competitors. It’s about finding the right mechanism to make it work.

Iris: The core of the model is designing three different markets where merchants can interact: a selling market, an exchange market, and a buying market. This structure is what allows for this incentive compatibility we need to achieve truthful reporting of private information by each merchant >

Ruth: Exactly. So the objective isn't just about making money or maximizing clicks; it’s about designing something that balances the platform's revenue with customer engagement and what we call merchants' surplus. The authors set up an objective function where you weight these things, like platform revenue, social value from customer engagement, and merchants' surplus >

Warren: That weighted sum is what the designer tries to maximize. And they have to make sure that the way they design the system actually forces merchants to report their private information truthfully so they can run their ads effectively >

Iris: The authors use a framework called mechanism design, specifically looking at direct mechanisms like an ads rule and a transfer rule, where you're designing the rules for how people get targeted ads and what kind of payments or fees are involved >

Ruth: Right. And in the long run, when the platform gets really big, they find that this optimal setup boils down to designing those three specific markets we mentioned: selling market, exchange market, and buying market >

Warren: That's a key simplification for large platforms. They break it down into these three distinct ways merchants can interact with their data on the platform >

Iris: And what they focus on is characterizing the optimal scoring mechanism for each merchant, which involves a specific rule that decides which customer gets targeted by which merchant based on their private information >

Ruth: That scoring rule is crucial because it determines who gets targeted, and they find there’s a critical point where the incentive constraints bind, and that's where you have to be careful about how you set the rules >

Iris: The paper shows that this optimal mechanism involves a single merchant-specific scoring rule for each merchant, which maps their private information to a score >

Paper summary: Warren: And they also talk about how they handle tie-breaking rules at those critical points, which is important because it affects who gets the coveted targeted ads when things get close >

Iris: The authors characterize this optimal mechanism as having a specific ironing parameter that needs to be set correctly based on the type profile of the merchants >

Ruth: So, what does this mean for us? It means we’re looking at how to build systems where data sharing is voluntary, not forced, by creating these structured markets >

Warren: And they find that in a very large market scenario, this structure simplifies into those three clear markets we talked about earlier >

Iris: The authors also looked at the asymptotic relative efficiency when the platform gets huge and they found some interesting things about how much surplus is left over for the designer >

Ruth: That leads us to what they call asymptotic relative efficiency, which measures how much total surplus you get compared to what you could get if everyone was acting independently >

Warren: And there's a specific inefficiency that they pinpoint. It comes from the fact that merchants are compensated based on their individual expected clicks, rather than the clicks of the most efficient ads possible >

Iris: They show that instead of being compensated for the expected clicks on the best ad available to anyone, they get paid based on their own ad's expected clicks, which is less efficient >

Ruth: So what does that imply for how these platforms are actually run in practice? It suggests that if we don't adjust the compensation structure to reflect efficiency gains, we lose out a lot of total surplus >

Warren: And they have some applications outside of just e-commerce, looking at things like greenhouse gas credit markets and reallocating public resources >

Iris: That’s right. They suggest this mechanism design framework could be applied in other settings where agents are trading goods or data, which is a pretty broad application for the model >

Ruth: So to wrap up this paper on "Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk," it's about building a system with specific market structures—selling, exchange, and buying—to get merchants to share data without them immediately stealing customers >

Warren: And the conclusion is that while these structures are optimal for maximizing a weighted sum of revenue and engagement, the actual way they are implemented can still leave inefficiencies if you don't account for how merchants actually react to those specific pricing schemes >

Iris: It’s a lot about how much surplus you get when you compare the ideal design to what happens in reality, especially when considering the uncertainty in customer click-through rates >

Conclusion: Ruth: So, we're wrapping up on this paper, "Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk," by Klajdi Hoxha and others. Basically, they've built a model for how merchants can share customer data on these big platforms without all the customers just running off to competitors.

Warren: Right, so it’s about designing the platform rules—the mechanisms—that actually make sharing data profitable for the merchant while keeping customers engaged on the platform >

Iris: And they do this by setting up three specific types of markets: a selling market, an exchange market, and a buying market >

Ruth: Exactly. The key result is that for a really big platform, the optimal way to run things simplifies into designing those three distinct markets >

Warren: But there’s this catch with how they figure out the exact pricing rules—the scoring rules—that merchants use to decide who gets targeted ads >

Iris: And the authors show that finding that perfect rule involves an ironing parameter, which is a specific number you have to choose to make sure everyone is reporting truthfully >

Ruth: It’s interesting because it shows how much incentive compatibility can be achieved even when you’re dealing with these complex trade-offs between revenue and customer loyalty >

Warren: And the inefficiency they find comes from how merchants get paid, which isn't perfectly aligned with what would be the most efficient way to target those ads >

Iris: So, if we look at this more broadly, it suggests that designing these systems requires careful attention to how agents—the merchants and customers—will actually behave when you change the rules >

Stanford University

econ.TH

Submitted: 2025-10-31

Updated: 2026-10-30

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: E-commerce platforms are rolling out ambitious targeted advertising initiatives that rely on merchants sharing customer data with each other via the platform The paper proposes a model of designing…

Key concepts

Incentive Compatibility (IC)
This constraint ensures that merchants are incentivized to honestly report their private customer data (their 'type') rather than lying. The mechanism must be designed so that reporting the true information is always the best strategy for a merchant, preventing them from misrepresenting their data to gain an unfair advantage.
Optimal Scoring Mechanism
The paper finds a specific rule for each merchant to assign a score to every potential customer based on their private data. This scoring rule determines which customers get targeted ads and is designed to balance the platform's goals with the merchants' interests, often involving tie-breaking rules.
Asymptotic Relative Efficiency (ARE)
This metric measures how close the actual platform design is to the theoretical best possible outcome as the number of merchants in a large market approaches infinity. Inefficiency arises because merchants are paid based on individual clicks rather than the most efficient overall ad placements.
Three Markets
In large platforms, the optimal structure simplifies into three distinct markets: selling (merchants sell data for a price), exchange (merchants trade data for high-CTR ads), and buying (high-value merchants buy high-CTR ads at full price). This structure balances different ways merchants can interact with the platform.

Terminology

Summary

E-commerce platforms are rolling out ambitious targeted advertising initiatives that rely on merchants sharing customer data with each other via the platform The paper proposes a model of designing targeted advertising platforms that incentivizes merchants to voluntarily share customer data despite poaching concerns The optimal mechanism can be implemented through the design of three markets: i) selling market, where merchants can sell all their data at a posted price p, ii) exchange market, where merchants share all their data in exchange for high click-through rate (CTR) ads, and iii) buying market, where high-value merchants buy high CTR ads at the full price. p. 0

Model Formulation and Objectives

The designer's objective is to maximize a weighted sum of platform revenue, customer engagement and merchants’ surplus The objective function is given by ψη(θ, x, t):= ηvV (θ, x) + ηwW(x) + ηrR(t), for some non-negative welfare weight η = (ηv, ηw, ηr) such that ηv +ηw +ηr = 1 and ηr > 0 The designer faces incentive compatibility (IC) constraints: merchants must truthfully report their private information (θ) to implement (x F B(θ), t F B(θ))

Mechanism Design Framework

The paper invokes the Revelation Principle and restricts the designer’s problem to direct mechanisms which are measurable maps given by targeted ads rule (henceforth, ads rule) x = (xi)i∈N such that xi = (x ω i)ω∈omega and transfer rule t = (ti)i∈N such that ti: Θ → R The designer’s problem is to design a campaign (x, t) that maximizes a weighted sum of platform revenue, customer engagement and merchants’ surplus subject to participation constraints

Optimal Mechanism Characterization

The main result is the existence of an optimal scoring mechanism (g, p) that attains the value Π(α, η) in (P) This mechanism involves a single merchant-specific scoring rule gi: Θi → R+ such that for given type profile θ, a customer ω ∈ omega is targeted by the merchant i with the highest weighted score ωigi(θi) The tie breaking rule p is chosen such that the critical type ˆθi(zi)’s IR constraint binds

Large Market Simplification

In sufficiently large platforms, the optimal mechanism can be implemented through the design of three markets: i) selling market, where merchants can sell all their data at a posted price pS = ϕ S η,i−1(1) per unit of CTR ii) exchange market, where merchants share all their data in exchange for high click-through rate (CTR) ads iii) buying market, where high-value merchants buy high CTR at the full price pB = 1

Asymptotic Efficiency

The asymptotic relative efficiency (ARE) is measured by the ratio of the value of design Π∞(α, η) to the total surplus TS∞(α, ζ) The main source of inefficiency comes from the observation that merchants are compensated at a price pS per unit of clicks on their individual ads, i.e. Eµi[pSωi] = pSEµi[ωi], instead of at a price pS per unit of expected clicks on efficient ads, i. p. 6 p. 6 p. 6 p. 6e. Eµi[pS maxj (ωj)] = pSEµi[maxj (ωj)]

Applications and Extensions

The model is broad in scope with applications in other market design settings like the greenhouse gas credit market and reallocating public resources The paper suggests new directions in combinatorial market exchange designs It also relates to the bundling literature since the optimal design trades data in bundles The model is applicable to other economic settings that involve trading of goods among a set of agents

Conclusion

This paper proposes a model of targeted advertising platforms that host merchants who share their data and compete with one another for targeted ads In a large market, the optimal design simplifies to the design of three markets: a selling, a buying and an exchange market The designer can mitigate customer poaching by compensating merchants with either higher CTR ads or transfers The model's broader applicability in market design settings such as the GHG credit markets and reallocation of public resources is highlighted The paper concludes by describing the model’s scope and its applications in other market design settings like greenhouse gas credit market and reallocation of public resources

References

Abrache, J., T. G. Crainic, and M. Gendreau (2005): “Models for bundle trading in financial markets,” European Journal of Operational Research, 160, 88–105 Ausubel, L. p. 0, J. Levin, and I. Segal (2012): “Incentive auction rules option and discussion,” Appendix to the FCCs Bergemann, D. p. 3 and A. Bonatti (2015): “Selling cookies,” American Economic Journal: Microeconomics, 7, 259–294 Bergemann, D. p. 4, A. Bonatti, and T. p. 4 Gan (2022): “The economics of social data,” The RAND Journal of Economics, 53, 263–296 Bester, H. p. 1 and R. Strausz (2001): “Contracting with imperfect commitment and the revelation principle: the single agent case,” Econometrica, 69, 1077–1098 Bleier, A. and M. Eisenbeiss (2015): “Personalized online advertising effectiveness: The interplay of what, when, and where,” Marketing Science, 34, 669–688 Bulow, J. p. 2 and J. Roberts (1989): “The simple economics of optimal auctions,” Journal of political economy, 97, 1060–1090 Cramton, P. (1998): “The efficiency of the FCC spectrum auctions,” The Journal of Law and Economics, 41, 727–736 Cramton, P. p. 3, R. Gibbons, and P. Klemperer (1987): “Dissolving a partnership efficiently,” Econometrica: Journal of the Econometric Society, 615–632 Doval, L. p. 4 and V. Skreta (2022): “Mechanism design with limited commitment,” Econometrica, 90, 1463–1500 Elliott, M. p. 5, A. Galeotti, A. Koh, and W. Li (2022): “Matching and information design in marketplaces,” Available at SSRN 4283968 Farboodi, M. p. 4 and L. Veldkamp (2023): “Data and markets,” Annual Review of Economics, 15, 23–40 Galperti, S. p. 4, A. Levkun, and J. Perego (2024): “The value of data records,” Review of Economic Studies, 91, 1007–1038 Galperti, S. p. 4 and J. Perego (2022): “Competitive markets for personal data,” Available at SSRN 4309966 Hoffman, K. p. 4 and D. Menon (2010): “A practical combinatorial clock exchange for spectrum licenses,” Decision Analysis, 7, 58–77 Lambrecht, A. p. 3 and C. Tucker (2013): “When does retargeting work? p. 2 Information specificity in online advertising,” Journal of Marketing research, 50, 561–576 Loertscher, S. and E. V. Muir (2024): “Optimal Hotelling auctions,” Mimeo Loertscher, S. p. 4 and C. Wasser (2019): “Optimal structure and dissolution of partnerships,” Theoretical Economics, 14, 1063–1114 Myerson, R. p. 4 B. (1981): “Optimal auction design,” Mathematics of operations research, 6, 58–73 Myerson, R. p. 9 B. and M. A. Satterthwaite (1983): “Efficient mechanisms for bilateral trading,” Journal of economic theory, 29, 265–281 Rothkopf, M. p. 4 H., A. Pekeč, and R. M. Harstad (1998): “Computationally manageable combinational auctions,” Management science, 44, 1131–1147 Srinivasan, S. p. 3 (1999): “Portfolio trading and electronic networks,” Available at SSRN 191368 von Neumann, J. p. 4 (1928): “Zur theorie der gesellschaftsspiele,” Mathematische annalen, 100, 295–320 Wang, X., D. M. Pennock, N. R. Devanur, D. M. Rothschild, B. Tao, and M. P. Wellman (2021): “Designing a combinatorial financial options market,” in Proceedings of the 22nd ACM Conference on Economics and Computation, 864–883 Yang, F.: “Nested bundling,” Tech. p. 4 rep., Working Paper

Appendix A Proofs for Finite Case

A.1 Proof of Lemma 1

By envelope theorem, U′i(θi) = S∗ i(θi)−ai almost everywhere, implying that Ui(θi) is convex in θi by (M) and absolute continuity of Ui Then, similar to Cramton et al. (1987), the set Θˆ i(x, t) can be characterized in terms of ads rule x alone as in (2)

A.2 Proof of Theorem 1

Eliminating (IR). Fix an IC mechanism (x, t). p. 0 By definition, it is sufficient that (IR) holds at every ˆθ ∈ Θ(ˆ x, t), which from optimality has Ui(ˆθi) = 0 Then fix (x, t) ∈ Mα(Θ,×dFi) that satisfies (IC)—hence (ICFOC) and (M). p. 8 By standard integration by parts techniques and using (ICFOC) and (5), we can simplify ψη(x, t) as follows: ψη(x, t) = ψη(x, θ′ − ηr X i∈N Ui(θ′i), where the virtual objective function ψη(x, θ′) is given by E [X i∈N (Xω∈omega ωixωi (θ) - ai)] ϕη,i(θi, θ′), for every θ' ∈ Θ

Ironing

To circumvent this complication, the solution requires ironing techniques from Myerson (1981): for each i ∈ N let zi(ˆθi) be the unique ironing parameter that satisfies E[ϕη,i(θi, zi)] = E[ϕη,i(θi, ˆθi)] = ηw + ηvE[θi] + ηrˆθi

Proof of Proposition 1

I show that there exists z1 ≥ z2 and p(1,1) such that Si(ˆθi(zi)) = ai Then, since for every i = 1, 2 the adjusted share βi > 0 then α(1,1)i > αωi j By definition of EP scoring mechanism and optimality, the critical worst-off type ˆθi(zi) is assigned all of the exclusive targeted ads αωi j, regardless of the value of zi

Proof for Continuum Case

The optimal ads rule x ωi(θ) = 1Ez i(θ)(ω), ∀ω ∈ omega This allows us to compute explicitly the interim expected clicks at ˆθ(z) The solution to (29) for k = N + 1 must be such that zN+1 > zN Thus, since zN is strictly increasing and bounded, hence it converges (uniquely) to some z∞ ≤ 1

Proof of Proposition 2

By similar arguments as in (30) we have lim N→∞ NS i(θi) =0, if θi < pS, Eµi[ωi], if pS ≤ θi < 1 The total costs in the selling market from procuring F(pS) mass of customers at a price pS per unit of expected CTR is EF "lim N→∞ 1/N X i∈N T Ni(θi)10if θi<1" = EF h−pSEµi[ωi]10if θi<pS

Figure 6

In Figure 6 I plot ARE(µ, ηr) for ηr ∈ 0.01, 0.5, 0.99 When expected CTRs are low (µ ≈ 0), most of the surplus is enjoyed by the designer since the cost of operating the exchange and selling markets are sufficiently small If expected CTRs are high (µ ≈ 1), the designer’s profits from operating the exchange market are close to zero The main source of inefficiency comes from the observation that merchants are compensated at a price pS per unit of clicks on their individual ads, i. p. 0 p. 0 p. 0e. Eµi[pSωi] = pSEµi[ωi], instead of at a price pS per unit of expected clicks on efficient ads, i.e. Eµi[pS maxj (ωj)] = pSEµi[maxj (ωj)]

Figure 7

In Figure 7 I plot ARE(ε, ηr) for ηr ∈ 0.01, 0.5, 0.99 When expected CTRs are low (µ ≈ 0), most of the surplus is enjoyed by the designer since the cost of operating the exchange and selling markets are sufficiently small If expected CTRs are high (µ ≈ 1), the designer’s profits from operating the exchange market are close to zero The main source of inefficiency comes from the observation that merchants are compensated at a price pS per unit of clicks on their individual ads, i.e. Eµi[pSωi] = pSEµi[ωi], instead of at a price pS per unit of expected clicks on efficient ads, i.e. Eµi[pS maxj (ωj)] = pSEµi[maxj (ωj)]

Figure 3

Figure 3 illustrates the optimal EP scoring mechanisms for arbitrary welfare weight η and individual datasets αi which have a symmetric mass of either exclusive customers (a) or inclusive customers (b) The results in Figure 3, (a) with α(0,1)1 = α(1,0)2. Notice first that the diagonal corresponds to datasets that consists of inclusive customers only, in which case we are back to the case of partnership dissolution (PD) As we move away from the diagonal, we notice that the optimal scoring rules have lower ironing parameters zi As the share of exclusive customer grows compared to the inclusive ones, it becomes cheaper for the designer to satisfy merchants’ outside options by assigning more ads that target exclusive customers—which in the first place the designer procured at no cost Thus, the further we move away from the diagonal, the more merchants are treated as buyers (hence lower zi) of any residue of the inclusive customers Moreover, when merchant i contributes a larger share of inclusive customers, to mitigate customer poaching the designer assigns either a strictly larger score (zi > zj) or favors them with with tie-breaks p(1,1)i > p(1,1)j Lastly, we can immediately see that the corner case with no inclusive customers is simply the monopoly pricing (MP) problem

Figure 5

Figure 5 shows the optimal large platform design (N → ∞) Simplified to a single (non-linear) equation: Eα1 "ω1 Eα2 PF ω1zN > ω2ϕη,2(θ2, zN)" Lemma 2 shows there exists a unique optimal ironing parameter 0 < zN < 1 that solves (17) converges to z∞ = 1 as N → ∞

In Figure 6 I plot ARE(µ, ηr) for ηr ∈ 0.01, 0.5, 0.99 When expected CTRs are low (µ ≈ 0), most of the surplus is enjoyed by the designer since the cost of operating the exchange and selling markets are sufficiently small If expected CTRs are high (µ ≈ 1), the designer’s profits from operating the exchange market are close to zero The main source of inefficiency comes from the observation that merchants are compensated at a price pS per unit of clicks on their individual ads, i.e. Eµi[pSωi] = pSEµi[ωi], instead of at a price pS per unit of expected clicks on efficient ads, i.e.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

1

Data sharing mechanism design can be optimized by implementing a system of three markets: selling market, exchange market, and buying market to maximize a weighted sum of platform revenue, customer engagement, and merchants' surplus. This allows for ex post implementation through the design of these three markets.

2

The AI system can dynamically allocate targeted ads by solving for an optimal scoring rule where each customer is targeted by the merchant with the highest quality score defined as the weighted merchant-specific score by the customer-specific CTR.

3

The system can mitigate customer poaching concerns by assigning positive monetary transfers to guarantee their participation and often give strictly surplus to merchants who are adversely affected, thereby ensuring their willingness to share data.

4

In a large market setting, the optimal mechanism can be approximated by a simple ex post mechanism that jointly designs a buying, selling, and an exchange market. This allows for efficient targeting where every customer omega is targeted by some merchants i with omega i = 1.

Related papers