arxiv: 2604.14260 · v1 · submitted 2026-04-15 · 💰 econ.TH

Recognition: unknown

How do you know you won't like it if you've (never) tried it? Preference discovery and data design

Sebastiano Della Lena , Alessio Muscillo , Paolo Pin

Authors on Pith no claims yet

Pith reviewed 2026-05-10 11:43 UTC · model grok-4.3

classification 💰 econ.TH

keywords preference discoverydata designbundlingco-consumption networkbelief updatingplatform designconsumer learningexposure structure

0 comments

The pith

The structure of consumption bundles shapes whether consumers learn their true preferences quickly or remain stuck in misperceptions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Consumers discover their preferences through trying different goods, yet platforms and firms often control the sequence and combinations of those trials. The paper shows that bundling creates a co-consumption network where surprises about one good's utility affect beliefs about others. This structure can be designed to either slow down learning or speed it up. When firms know the errors in estimates, they can target bundles to lock in misperceptions. In contrast, designs that break correlations between goods allow faster and more accurate discovery of true preferences.

Core claim

The authors establish that in their data-design framework, the geometry of co-consumption determines learning outcomes: popularity-biased bundles slow preference discovery while correlation-breaking bundles accelerate it, and known estimation errors allow bias-targeted designs to shut down learning and amplify misperceptions.

What carries the argument

The co-consumption network generated by bundling, through which utility surprises propagate to update beliefs.

Load-bearing premise

Consumers update their beliefs about preferences according to utility surprises that propagate through the network created by past co-consumption patterns.

What would settle it

A controlled experiment that exposes participants to different bundling regimes and measures whether correlation-breaking bundles produce faster convergence to independently verified true preferences compared to popularity-biased bundles.

Figures

Figures reproduced from arXiv: 2604.14260 by Alessio Muscillo, Paolo Pin, Sebastiano Della Lena.

**Figure 1.** Figure 1: Line Given this path of consumption, if the consumer tastes again good 1 alone and experiences a unitary positive surprise, ∆ut = +1, the updating of the estimated preferences of the directly connected good and the good at distance three decrease, whereas those of the good at distance two increases. See [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

**Figure 2.** Figure 2: Updating after tasting good 1 and experiencing a unitary positive surprise. 2.3 Data Design and Learning The previous section and examples show how consumption surprises propagate across connected goods, leading to preference revisions that depend on the structure of the network of goods that summarizes the consumer’s consumption history. We now build on that mechanism and discuss how, at each point in tim… view at source ↗

**Figure 3.** Figure 3: Example with ∆β1 < 0 and ∆β2 > 0. (a) Orthogonal bundles xt+1 ⊥ ∆βt do not generate expected surprise and thus prevent learning; (b) Bundles such that x ′ t+1∆βt < 0 will generate positive expected surprise. Proposition 2 shows that a positive (or negative) surprise generated by the bundle xt+1 is transmitted across goods through the covariance structure encoded in Wt . In particular, the expected revision… view at source ↗

**Figure 4.** Figure 4: Updates E[βˆt+1] − βˆt after consuming the bundle xt+1 = (x1,t+1, 1 − x1,t+1), when w12,t < 0, and the vector of biases in estimated preferences is ∆βt = (−0.2, 0.2). If x1,t+1 = 1 2 , xt+1 ⊥ ∆βt. Long-run Having characterized the local geometry of updating, we now analyze long-run learning, identifying the conditions under which βˆ t converges to β and the consumption patterns that make learning faster. W… view at source ↗

**Figure 5.** Figure 5: Learning dynamics with two goods (|I| = 2) under different bundle designs, starting from βˆ 0 = (0.9, 1.2), with β = (1, 1) and σ 2 = 0. 17 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

**Figure 6.** Figure 6: Co–starring network of the 121 actors in our database, with larger font size for more central nodes according to network centrality v N . The link width is proportional to the number of movies two actors co-star in. Samuel L. Jackson, Robert Downey Jr., Cate Blanchett and Leonardo DiCaprio are in red for readability, as we focus on them in Section 4.2. 29 [PITH_FULL_IMAGE:figures/full_fig_p030_6.png] view at source ↗

**Figure 7.** Figure 7: History of the βˆ’s of four actors (in million US$): Robert Downey Jr., Leonardo DiCaprio, Cate Blanchett and Samuel L. Jackson. Dashed lines denote movies where at least one of them acted, for readability. Cate Blanchett stars in Thor: Ragnarok (2017) which explains the jump in her βˆ. Notably, however, even if the other actors do not appear in such a movie, their βˆ’s are also updated because they are li… view at source ↗

**Figure 8.** Figure 8: Bundles that do not generate expected surprise and thus prevent learning. (a) If ˆα = α, any bundle orthogonal to ∆βt; (b) if ˆα ̸= α, bundles lying on a shifted hyperplane orthogonal to ∆βt, with the shift determined by the sign of (ˆα − α). Finally, note that the permanent bias in βˆ t induced by intercept misspecification has direct implications for the provider’s profits. When the consumer underestimat… view at source ↗

**Figure 9.** Figure 9: Co–starring network of the 121 actors in our database, with larger font size for more central nodes according to correlation centrality v C . The two groups induced by this eigenvector are in different colors. As in [PITH_FULL_IMAGE:figures/full_fig_p056_9.png] view at source ↗

**Figure 10.** Figure 10: On the left, the 14 actors among the 121 in the dataset who appear in at least 4 movies both before and after the adjacency matrix Zt reaches full rank. On the right, the 17 movies in which they co-star together (except Robert Downey Jr. who does not co-appear with others) [PITH_FULL_IMAGE:figures/full_fig_p058_10.png] view at source ↗

**Figure 11.** Figure 11: Estimated βˆ’s for 14 selected actors and for their complementarities, indicated by an “X” between actors’ last names. 57 [PITH_FULL_IMAGE:figures/full_fig_p058_11.png] view at source ↗

read the original abstract

Consumers discover their preferences through experience, yet the sequence and composition of those experiences are often designed by firms, digital platforms, or policymakers. We introduce a ``data-design'' framework for preference discovery, in which the structure of consumption data shapes learning. Bundling generates correlated exposure across goods, so utility surprises propagate through the co-consumption network. When estimation errors are known, bias-targeted design can shut down learning and amplify misperceptions. Conversely, robust design uses only the geometry of past co-consumption: popularity-biased bundles slow learning, while correlation-breaking bundles accelerate preference discovery. The framework thus explains how dominant platforms can sustain biased demand through exposure design, and why effective regulation may need to intervene on the structure of exposure itself rather than only on prices or market shares.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This introduces a data-design framework for preference discovery through co-consumption networks, but the unspecified learning mechanisms make the robustness claims questionable.

read the letter

Here's the quick take: this work introduces a data-design framework that shows how bundling creates correlated experiences, letting utility surprises move through a co-consumption network and shape what consumers learn about their preferences. The key claims are that bias-targeted designs can block learning while robust ones speed it up using only the geometry of past bundles. It does a solid job laying out why platforms might keep demand biased through exposure patterns rather than prices alone, and the policy suggestion to regulate the structure of data exposure is a useful shift in thinking. The main weakness is that the model details are thin in the abstract. No exact learning rule or utility function is given, so the results on how popularity-biased bundles slow things down or correlation-breaking ones speed them up might not hold up under standard assumptions. The concern that these geometry effects could be sensitive to the belief update process is fair, and without the derivations it's tough to see if the framework is as general as claimed. The thinking here engages honestly with the experience goods and platform literature. No obvious circularity, and the new angle on data structure as a design variable stands out. This would be useful for people working on theoretical models of consumer learning in digital markets or on regulation of big platforms. A reader who wants new ways to think about how firms control information flow through consumption sequences would get something from it. I'd put it through peer review. The core idea is worth checking out formally, even if it needs work on the foundations.

Referee Report

2 major / 2 minor

Summary. The paper introduces a data-design framework in which the structure of consumption experiences—particularly through bundling—shapes consumers' preference discovery. Bundling creates correlated exposures that generate a co-consumption network through which utility surprises propagate. When estimation errors are known, bias-targeted design can shut down learning and amplify misperceptions. In contrast, robust design relies only on the geometry of past co-consumption: popularity-biased bundles slow learning while correlation-breaking bundles accelerate preference discovery. The framework is used to explain how platforms can sustain biased demand via exposure design and to argue that regulation may need to target the structure of exposure rather than only prices or shares.

Significance. If the results hold, the paper provides a useful theoretical lens on how platforms and firms can influence demand formation through the design of consumption data rather than solely through pricing. The distinction between bias-targeted and geometry-based robust designs offers a clean way to think about interventions that either hinder or facilitate learning. The emphasis on network geometry as a sufficient statistic for robust design is potentially valuable for both positive and normative analysis in digital markets, though its force depends on the robustness of the propagation mechanism to standard updating rules.

major comments (2)

[Framework description (model setup)] The propagation of utility surprises through the co-consumption network is the load-bearing mechanism for all qualitative claims (slowing vs. accelerating discovery, shutdown of learning). Yet the manuscript does not specify the belief-update rule or the form of the utility function. It is therefore unclear whether the geometry-based results survive standard Bayesian updating with quadratic utilities or hold only under particular linear approximations or heuristics. A derivation or invariance result showing that the directional effects on learning speed are robust to the precise update rule would be required to support the central contrast between bias-targeted and robust design.
[Results on bias-targeted design] The claim that bias-targeted design 'can shut down learning' when estimation errors are known is presented as a sharp result, but the conditions under which shutdown occurs (versus mere slowing) are not derived explicitly. Without a precise statement of how targeting is implemented and the resulting fixed point of beliefs, it is difficult to evaluate the quantitative force of the contrast with robust design or to assess whether the shutdown is knife-edge or generic.

minor comments (2)

[Abstract] The abstract is concise but introduces several terms (co-consumption network, geometry of past co-consumption, bias-targeted vs. robust design) without brief definitions; a short parenthetical clarification would improve accessibility.
[Notation and definitions] Notation for the network and for the propagation operator should be introduced once and used consistently; early definitions would reduce ambiguity when the geometry-based claims are stated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of the framework's foundations and the precision of our claims. We address each major point below and will incorporate clarifications and derivations in the revised manuscript.

read point-by-point responses

Referee: [Framework description (model setup)] The propagation of utility surprises through the co-consumption network is the load-bearing mechanism for all qualitative claims (slowing vs. accelerating discovery, shutdown of learning). Yet the manuscript does not specify the belief-update rule or the form of the utility function. It is therefore unclear whether the geometry-based results survive standard Bayesian updating with quadratic utilities or hold only under particular linear approximations or heuristics. A derivation or invariance result showing that the directional effects on learning speed are robust to the precise update rule would be required to support the central contrast between bias-targeted and robust design.

Authors: We agree that an explicit connection to standard updating rules strengthens the results. The current presentation emphasizes network geometry as the sufficient statistic, but the manuscript does not derive the propagation under Bayesian updating. In the revision we will add a new proposition (with proof in the appendix) showing that, under quadratic utilities and Bayesian updating, the directional effects on learning speed are preserved: popularity-biased bundles continue to slow discovery while correlation-breaking bundles accelerate it. The invariance follows because the co-consumption matrix governs the covariance of signals regardless of the exact posterior mean update, provided the utility surprises are mean-zero conditional on true preferences. revision: yes
Referee: [Results on bias-targeted design] The claim that bias-targeted design 'can shut down learning' when estimation errors are known is presented as a sharp result, but the conditions under which shutdown occurs (versus mere slowing) are not derived explicitly. Without a precise statement of how targeting is implemented and the resulting fixed point of beliefs, it is difficult to evaluate the quantitative force of the contrast with robust design or to assess whether the shutdown is knife-edge or generic.

Authors: We accept that the shutdown claim requires a sharper characterization. Bias-targeted design shuts down learning when the designer selects bundles that exactly offset the known vector of estimation errors, producing a fixed point at which the expected utility surprise is zero in every direction. In the revision we will state this fixed-point condition formally (as a corollary to the main propagation result) and delineate the parameter region in which complete shutdown occurs versus mere slowing. We will also note that shutdown is not knife-edge but requires precise knowledge of the error vector; robust design, by contrast, operates without that knowledge and relies only on the geometry of past co-consumption. revision: yes

Circularity Check

0 steps flagged

No circularity; framework derives implications from co-consumption geometry

full rationale

The paper introduces a data-design framework in which bundling creates correlated exposure and utility surprises propagate through a co-consumption network. Claims about bias-targeted design shutting down learning and robust design accelerating discovery via geometry follow directly from the stated network structure and updating process. No equations reduce predictions to fitted inputs by construction, no self-citations bear the central load, and no uniqueness theorems or ansatzes are smuggled in. The model is self-contained against its own assumptions; results are implications within the framework rather than tautological renamings or self-referential definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework relies on assumptions about how learning occurs via correlated exposures, but no specific free parameters or invented entities are mentioned in the abstract.

axioms (2)

domain assumption Consumers discover preferences through experience with utility surprises propagating through co-consumption networks.
Central to the framework as described in the abstract.
domain assumption The structure of consumption data can be designed to influence learning outcomes.
Assumed in the data-design approach.

pith-pipeline@v0.9.0 · 5436 in / 1220 out tokens · 29239 ms · 2026-05-10T11:43:00.089248+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 2 canonical work pages

[1]

Acemo g lu, D., Makhdoumi, A., Malekian, A., and Ozdaglar, A. (2024). When big data enables behavioral manipulation. American Economic Review: Insights , forthcoming

2024
[2]

Adams, W. J. and Yellen, J. L. (1976). Commodity bundling and the burden of monopoly. Quarterly Journal of Economics , 90(3):475--498

1976
[3]

What is product bundling and how to use it

Amazon (2024). What is product bundling and how to use it. Amazon Seller Blog. Available at https://sell.amazon.com/blog/product-bundling

2024
[4]

Discussion: Consumables bundling policy update

Amazon Seller Central (2024). Discussion: Consumables bundling policy update. Amazon Seller Forums. Available at https://sellercentral.amazon.com/seller-forums/discussions/t/44a087bd-f831-4635-9c46-21a153705af9

2024
[5]

Bergemann, D., Heumann, T., and Morris, S. (2022). Screening with persuasion. arXiv preprint arXiv:2212.03360

work page arXiv 2022
[6]

and Morris, S

Bergemann, D. and Morris, S. (2019). Information design: A unified perspective. Journal of Economic Literature , 57(1):44--95

2019
[7]

and Plemmons, R

Berman, A. and Plemmons, R. J. (1994). Nonnegative Matrices in the Mathematical Sciences . Society for Industrial and Applied Mathematics, Philadelphia, PA

1994
[8]

Bramoull \'e , Y., Kranton, R., and D'Amours, M. (2014). Strategic interaction and networks. American Economic Review , 104(3):898--930

2014
[9]

Busch, C. (2023). From algorithmic transparency to algorithmic choice: European perspectives on recommender systems and platform regulation. In Recommender Systems: Legal and Ethical Issues , pages 31--54. Springer International Publishing Cham

2023
[10]

Calvano, E., Calzolari, G., Denicolo, V., and Pastorello, S. (2020). Artificial intelligence, algorithmic pricing, and collusion. American Economic Review , 110(10):3267--3297

2020
[11]

Campbell, A. (2013). Word-of-mouth communication and percolation in social networks. American Economic Review , 103(6):2466--2498

2013
[12]

Campbell, A., Ushchev, P., and Zenou, Y. (2024). The network origins of entry. Journal of Political Economy , 132(11):000--000

2024
[13]

Cerreia-Vioglio, S., Maccheroni, F., Marinacci, M., and Rustichini, A. (2023). Multinomial logit processes and preference discovery: inside and outside the black box. Review of Economic Studies , 90(3):1155--1194

2023
[14]

Cooke, K. (2017). Preference discovery and experimentation. Theoretical Economics , 12(3):1307--1348

2017
[15]

Dasaratha, K., Golub, B., and Shah, A. (2024). Incentive design with spillovers. Mimeo

2024
[16]

Delaney, J., Jacobson, S., and Moenig, T. (2020). Preference discovery. Experimental Economics , 23(3):694--715

2020
[17]

Ely, J., Frankel, A., and Kamenica, E. (2015). Suspense and surprise. Journal of Political Economy , 123(1):215--260

2015
[18]

Case comp/c-3/39.530 --- microsoft (tying internet explorer to windows)

European Commission (2009). Case comp/c-3/39.530 --- microsoft (tying internet explorer to windows). Decision of 16 December 2009. Available at https://ec.europa.eu/competition/antitrust/cases/dec_docs/39530/39530_2671_5.pdf

2009
[19]

Fainmesser, I. P. and Galeotti, A. (2016). Pricing network effects. The Review of Economic Studies , 83(1):165--198

2016
[20]

Galeotti, A., Golub, B., and Goyal, S. (2020). Targeting interventions in networks. Econometrica , 88(6):2445--2471

2020
[21]

Galeotti, A., Golub, B., Goyal, S., Talam \`a s, E., and Tamuz, O. (2024). Robust market intervenctions. Mimeo

2024
[22]

and Goyal, S

Galeotti, A. and Goyal, S. (2009). Influencing the influencers: a theory of strategic diffusion. The RAND Journal of Economics , 40(3):509--532

2009
[23]

Golub, B. (2025). Eigenvalues in microeconomics. arXiv preprint arXiv:2502.12309

work page arXiv 2025
[24]

and Jackson, M

Golub, B. and Jackson, M. O. (2010). Naive learning in social networks and the wisdom of crowds. American Economic Journal: Microeconomics , 2(1):112--149

2010
[25]

and Jackson, M

Golub, B. and Jackson, M. O. (2012). How homophily affects the speed of learning and best-response dynamics. Quarterly Journal of Economics , 127(3):1287--1338

2012
[26]

Grenet, J., He, Y., and K \"u bler, D. (2022). Preference discovery in university admissions: The case for dynamic multioffer mechanisms. Journal of Political Economy , 130(6):1427--1476

2022
[27]

and Gentzkow, M

Kamenica, E. and Gentzkow, M. (2011). Bayesian persuasion. American Economic Review , 101(6):2590--2615

2011
[28]

Klimashevskaia, A., Jannach, D., Elahi, M., and Trattner, C. (2024). A survey on popularity bias in recommender systems. User Modeling and User-Adapted Interaction , 34(5):1777--1834

2024
[29]

and Muir, E

Loertscher, S. and Muir, E. V. (2022). Monopoly pricing, optimal randomization, and resale. Journal of Political Economy , 130(3):566--635

2022
[30]

P., McMillan, J., and Whinston, M

McAfee, R. P., McMillan, J., and Whinston, M. D. (1989). Multiproduct monopoly, commodity bundling, and correlation of values. The Quarterly Journal of Economics , 104(2):371--383

1989
[31]

Nalebuff, B. (2004). Bundling as an entry barrier. The Quarterly Journal of Economics , 119(1):159--187

2004
[32]

Nelson, P. (1970). Information and consumer behavior. Journal of P olitical E conomy , 78(2):311--329

1970
[33]

Algorithmic competition

OECD (2023). Algorithmic competition. Technical Report DAF/COMP(2023)3, Organisation for Economic Co-operation and Development, Paris

2023
[34]

A., Ortoleva, P., and Riella, G

Ok, E. A., Ortoleva, P., and Riella, G. (2012). Incomplete preferences under uncertainty: Indecisiveness in beliefs versus tastes. Econometrica , 80(4):1791--1808

2012
[35]

Pollock, D. S. G. (2003). Recursive estimation in econometrics. Computational Statistics & Data Analysis , 44:37--75

2003
[36]

and Morrison, W

Sherman, J. and Morrison, W. J. (1950). Adjustment of an inverse matrix corresponding to a change in one element of a given matrix. The Annals of Mathematical Statistics , 21(1):124--127

1950
[37]

Slovic, P. (1995). The construction of preference. American psychologist , 50(5):364

1995
[38]

Microsoft Corporation (1999)

United States v. Microsoft Corporation (1999). Findings of fact, civil action no. 98--1232 (tpj). Available at https://www.justice.gov/atr/us-v-microsoft-proposed-findings-fact-0

1999
[39]

and Zenou, Y

Ushchev, P. and Zenou, Y. (2018). Price competition in product variety networks. Games and Economic Behavior , 110:226--247

2018
[40]

Watts, D. J. (1999). Small worlds: the dynamics of networks between order and randomness . Princeton University Press

1999