A Survey on Data Pricing: from Economics to Data Science

1 Introduction

• Data pricing provides a scalable approach for valuing shared and reused data, with the survey focusing on digitally distributed digital products and data products as information goods.Digital products include intangible electronically consumed goods, while data products include datasets and information services derived from datasets; methods for pricing digital products can generalize to data products.Source: §1 · pp. 1–3

1.1 Related Surveys

• The survey offers a comprehensive interdisciplinary account of data pricing across economics, marketing, e-commerce, databases, data management, operational research, management science, machine learning, and AI.It presents broader coverage than prior surveys, connecting economics, digital product pricing, and data product pricing while examining arbitrage-freeness, revenue maximization, fairness, truthfulness, and privacy preservation.Source: §1.1 · pp. 3–4

1.2 Structure of This Survey

• The article proceeds from economics and fundamental pricing principles to digital-product and data-product pricing, then concludes with challenges and future directions.Its roadmap covers cost reduction and product distinctions, versioning and desirable pricing properties, digital-product revenues, bundling, subscriptions and auctions, and data-marketplace pricing including dynamic, online, federated, and collaborative-learning settings.Source: §1.2 · pp. 5–6

2 Economics of Data Pricing

• The economics of data pricing focuses on cost reductions specific to information goods and their differences from digital products and data as products.The section frames cost reduction as the basis for examining economic factors in data pricing.Source: §2 · p. 6

2.1 Cost Reduction in Information Goods

• Information goods reduce search, production, replication, transportation, and tracking or verification costs, requiring adjustments to standard economic models.The survey largely follows Goldfarb and Tucker’s framework for these five cost reductions.Source: §2.1 · p. 6

2.1.1 Search Costs

• Low search costs help users discover and compare digital products and datasets, lower online book and CD prices, and support diverse or long-tail consumption.Online book and CD prices are lower than offline prices, although price dispersion does not shrink accordingly. Low search costs also support variety, recommender-mediated consumption, platform matching, and trade efficiency.Source: §2.1.1 · p. 7

2.1.2 Production Costs

• Information goods substantially reduce material, semi-finished-product, sharing, supply-chain, and customization costs, enabling pay-as-you-go, query-based consumption, and long-tail business models.Unit costs can approach zero through sufficient reuse and sales volume, while reduced customization costs support smaller and more diverse customer groups.Source: §2.1.2 · pp. 7–8

2.1.3 Replication Costs

• Because information goods are non-rival and have zero marginal costs, bundling can efficiently serve diverse preferences, while public availability supports contribution and complementary sales.Bundling thousands of digital products may be economically optimal, but low replication costs also challenge copyright, affect innovation, and facilitate privacy breaches, spamming, and online crime.Source: §2.1.3 · pp. 8–9

2.1.4 Transportation Costs

• Near-zero transportation costs can produce a flat-world effect, but local tastes, regulation, and copyright policies still shape regional availability and consumption.Local tastes remain relevant for music and content consumption, while regional restrictions can alter contribution and media access.Source: §2.1.4 · p. 9

2.1.5 Tracking and Verification Costs

• Low tracking costs enable personalized markets, behavioral price discrimination, versioning, personalized advertising, and auctions for advertising or information-good price discovery.Tracking also creates privacy concerns, while privacy-sensitive customers may withhold information from monopolies; versioning is discussed further in Section 3.1.Source: §2.1.5 · pp. 9–10
• Low verification costs reduce the costs of establishing producer and user identity and reputation, thereby facilitating online transactions and lowering trust costs.These verification gains accompany the broader tracking-cost reductions in information goods.Source: §2.1.5 · p. 10

Units and consumption

• Data products are priced and consumed in customer-specific aggregates, whereas digital products typically have fixed units consumed independently.A single low-granularity data unit may have little value; customers commonly combine and aggregate many units for their needs.Source: §2.2 · pp. 10–11

Aggregateability

• Data sets support flexible aggregation across dimensions, creating business opportunities and challenges such as ensuring arbitrage-freeness; digital-product bundling is not aggregation.Bundling combines independently consumed digital products to exploit low replication costs and meet diverse demands.Source: §2.2 · p. 11

Consumption modes

• Digital products are consumed directly by people, while data products are typically analyzed, summarized, or used by computers to train models and support decisions.Model outputs can automate operations or support human decision making.Source: §2.2 · p. 11

Reuse and resale

• Data can be reused through different aggregations and analyses and transformed for hard-to-detect resale, unlike digital products that are more readily reused or resold whole.These transformation and resale possibilities materially distinguish data-product pricing considerations.Source: §2.2 · p. 11

Context-dependent classification

• The same information may be a digital product when read directly but a data product when systematically collected and processed for analytics.Tweets and customer reviews illustrate this shift from online reading to batch analysis for event detection, customer profiling, and recommender systems.Source: §2.2 · pp. 11–12

2.3 Summary

• Information goods reduce several key costs and differ internally across consumption, aggregateability, consumption methods, and reuse and resale.These cost reductions have significant implications for pricing information goods.Source: §2.3 · p. 12

3 Fundamental Principles of Data Pricing

• Versioning is presented as a fundamental framework for designing information goods and pricing them.The section then turns to important properties in cost models for digital and data products.Source: §3 · p. 12

3.1 Versioning

• Versioning links price to customer value by offering different versions that appeal to customers valuing different features.Software versions may vary in features, while movies may vary in image resolution and sound effects.Source: §3.1 · pp. 12–13
• Versions can differ by delivery delay, convenience, comprehensiveness, manipulation rights, community access, advertising, and customer support.For example, stock-market services may offer real-time quotes or the same information 20 minutes later; most versions subtract value from the most complete version.Source: §3.1 · p. 13
• Free versions let customers test information goods whose value is initially unclear, supporting awareness, follow-on sales, customer networks, attention, and competitive advantage.Source: §3.1 · p. 13
• The needed number of versions depends on how many ways the information can be used and how much customers’ valuations vary.Source: §3.1 · p. 13
• Versioning has been applied to relational data sets and query results, where relational views flexibly generate versions amid challenges including arbitrage and fine-grained pricing.Further challenges include pricing updates, integrated data, and competing data sources.Source: §3.1 · p. 13

3.2.1 Truthfulness

• Truthfulness requires buyers to act selfishly by offering only the price that maximizes their true utility, enabling mechanisms such as auctions.Different buyers may assign different utility values to the same product.Source: §3.2.1 · p. 14

3.2.2 Revenue Maximization

• Revenue maximization is a central pricing objective, but low replication costs make revenue and profit maximization differ from physical-product markets.Pricing models may instead optimize lowest cost, highest profit, or largest sales.Source: §3.2.2 · p. 14

3.2.3 Fairness

• Fairness requires balanced, symmetric, zero-contribution, and additive payments; the Shapley value is the unique allocation satisfying these requirements [184].For information goods, near-zero replication costs may let sellers produce extra units to obtain unjustifiably larger Shapley payments.Source: §3.2.3 · p. 15

3.2.4 Arbitrage-free Pricing

• Arbitrage-free pricing is desirable because customers can exploit inconsistent prices or aggregation effects to obtain equivalent information more cheaply.With variance-based query pricing, ten independently perturbed variance-10 answers costing $50 can be averaged into variance 1, saving $50 versus one $100 answer.Source: §3.2.4 · p. 16

3.2.5 Privacy-preservation

• Privacy preservation is essential because marketplace transactions can expose buyers, information-good providers, and third parties.Protection directions include hiding purchase details, decentralized trustworthy marketplaces, payment–accuracy tradeoffs, and privacy-sensitive aggregation [50].Source: §3.2.5 · pp. 16–17

3.2.6 Computational Efficiency

• Computational efficiency requires pricing to scale to many goods and buyers, with marketplace price computation expected to be polynomial in the number of sellers.Auction-based pricing also requires fast bid processing.Source: §3.2.6 · pp. 17–18

3.3 Summary

• Versioning links prices for different information-good versions to values placed by different customer groups.Pricing information goods also requires truthfulness, revenue maximization, fairness, arbitrage-free pricing, privacy preservation, and computational efficiency, creating technical challenges for pricing models.Source: §3.3 · p. 18

• The review uses digital-product pricing to identify ideas that can be borrowed and extended to data products, whose boundary with digital products is sometimes blurry.Source: §4 · p. 18
• It first examines the three major streams of revenue for digital products.Source: §4 · p. 18
• It then reviews bundling and subscription, and auctions, two pricing-model types popularly adopted by digital-product marketplaces.Source: §4 · p. 18

Streams of Revenues

• Digital products generate revenue through money, information/privacy, and time/attention streams.Money comes from selling content or services; information/privacy from collecting and selling customer information; and time/attention from selling advertising space.Source: §4.1 · pp. 18–19
• Firms often combine revenue streams because they compete, requiring tradeoffs such as charging for ad-free or premium content versus expanding free, attention-based access.Paywalls can reduce customer visits, while free samples may stimulate long-term sales.Source: §4.1 · p. 19
• Money-stream models include fixed prices, pricing tiers, subscription durations, freemium models, and micropayments, which affect consumer behavior differently from subscriptions.Micropayments enable very small payments that are impractical in traditional credit-card transactions because of network service fees.Source: §4.1 · p. 19
• Information/privacy revenue uses customer identities, behaviors, preferences, and needs for direct marketing, audience segmentation, customized discounts, and individually priced queries.Selling customer data connects digital-product pricing with data-product and privacy-information pricing, which the next section examines.Source: §4.1 · p. 20
• Time/attention revenue mainly comes from embedded advertising, whose effects are difficult to measure, while retargeting combines online and offline behavior data to focus ads on previously interested customers.Advertisers also customize online display ads.Source: §4.1 · pp. 20–21

Bundling and Subscription Planning

• Bundling is a combinatorial optimization problem enabled by low replication costs, encompassing static single- or multi-product pricing and substitution or separate-seller scenarios.Traditional bundling is also common in telecommunications, financial services, healthcare, and consumer electronics.Source: §4.2 · p. 21
• Selling items separately or as one grand bundle achieves at least a constant fraction of optimal revenue under independent values and additive buyers.This supports choosing either individual prices or a grand-bundle price based on expected prices.Source: §4.2 · pp. 21–22
• For unlimited supplies and unknown preferences, a random single price achieves revenue within a logarithmic factor for general valuations.Uniform pricing is practically convenient, while larger bundles improve valuation prediction through the law of large numbers.Source: §4.2 · p. 22
• Subscription pricing jointly sets fees and periods for product sets and rental prices, balancing heterogeneous users’ usage rates and product values [13].Platforms may encourage lower-usage customers to subscribe while higher-usage customers rent, and often offer both options.Source: §4.2 · p. 22
• Grand subscription has necessary and sufficient optimality conditions, while cardinality-proportional fees achieve 1 4 log 2m+log n of optimal revenue.The approximation cannot be improved by more than Ω( 1 log n) in polynomial time for n customer types and m product types.Source: §4.2 · p. 22
• Modeling bundling and subscriptions is computationally challenging, while dynamic pricing through promotions and coupons has rarely been studied.The limitation follows from the combinatorial nature of these pricing problems.Source: §4.2 · pp. 22–23

4.3.1 Basics about Auctions

• The four basic auction types are ascending-bid, descending, first-price sealed-bid, and second-price sealed-bid auctions, with the winner paying the final, accepted, named, or second-highest price, respectively.Auction values may follow private-value or pure common-value models.Source: §4.3.1 · p. 23

4.3.2 Sponsored Search Auctions

• Sponsored search auctions allocate ad positions using advertiser keyword bids and may charge per impression, click, or action; generalized second-price ranking incorporates CTR, relevance, and site quality.Early generalized first-price auctions were unstable, costly for advertisers, and reduced search-engine profits, motivating later mechanisms.Source: §4.3.2 · pp. 24–25

4.3.3 Auctions on Digital Products with Unlimited Supplies

• Unlimited-supply digital-good auctions use random sampling for competitive pricing, while asymmetric deterministic auctions can approximately match optimal single-price revenue.Random sampling auctions are 15-competitive in the worst case and 4-competitive when at least 6 bids are as good as the optimal sale price.Source: §4.3.3 · pp. 25–27

4.3.4 Envy-free Auctions

• No auction can simultaneously be truthful, competitive, and envy-free; CORE-based randomized mechanisms trade either truthfulness or envy-freeness probabilistically while retaining competitiveness.One mechanism is truthful with probability 1−ϵ and always envy-free, whereas the other is envy-free with probability 1−ϵ and always truthful.Source: §4.3.4 · p. 27

4.3.5 Online Auctions

• Online auctions for unlimited supply achieve O(log h)-competitiveness with randomized thresholds, while bucket-based improvements achieve O(exp(√log log h))-competitiveness.Here h is the ratio of the highest to lowest bidder valuations; online auctions make decisions as bids arrive and require incentive compatibility.Source: §4.3.5 · p. 28

4.4 Summary

• Digital-product pricing considers revenue from money, information/privacy, and time/attention, alongside bundling and subscription planning shaped by low replication costs.It also reviews sponsored search, unlimited-supply, envy-free, and online auctions, whose ideas extend to data-product pricing.Source: §4.4 · p. 29

5.1 Data Markets and Pricing, What Are They?

• Data markets sell open, public, and private data to beneficiaries including analysts, vendors, developers, providers, consultants, certification entities, and market owners through free, usage-based, package, flat-fee, two-part, and freemium models.Markets may be monopolistic, oligopolistic, or strongly competitive, while data products can be versioned by quality, position, or query views.Source: §5.1 · pp. 30–32
• Data valuation for machine learning spans optimal mechanism design, privacy-compensating auctions, quality-based pricing, and Shapley or alternative methods that assess each record’s contribution to model performance.Exact data Shapley computation is computationally prohibitive for large sophisticated models, motivating Monte Carlo, gradient-based, and distributional estimators.Source: §5.1 · pp. 31–33

5.2 Arbitrage-free Pricing

• Arbitrage-free pricing requires derived queries or products not to cost more than the prices of the items from which they can be obtained, but practical computation remains difficult and often non-scalable.Query/view frameworks establish arbitrage-free and discount-free pricing functions, while Qirana uses uncertainty reduction and support-set approximations because complete support sets yield a #P-hard calculation.Source: §5.1 · p. 34, §5.2 · pp. 35–37
• Li et al. show that π(q, v) = Ω(1/v) is arbitrage-free when f is a positive semi-norm, and that applying a subadditive, nondecreasing function to arbitrage-free prices preserves arbitrage-freeness.Their framework nevertheless leaves bundle, biased-estimator, and profit-maximization challenges unresolved.Source: §5.2 · pp. 35–36
• Uniform user pricing is arbitrage-free and logarithmically approximates maximum revenue, while Xia and Muthukrishnan also provide a greedy non-uniform design retaining that approximation.For specified maximum and minimum purchase demands, the same work provides an O(D) approximation algorithm, where D is the largest minimal demand among buyers.Source: §5.2 · p. 38, §5.3 · p. 39

5.3 Revenue Maximization Pricing

• Revenue maximization for data products is less explored than digital-product pricing because data pricing must accommodate urgent concerns such as arbitrage, although existing work gives approximation algorithms for user and query/view markets.Chawla et al. study arbitrage-free revenue maximization with single-minded buyers and unlimited supply under uniform bundle, additive, and fractionally subadditive pricing.Source: §5.3 · p. 39
• For revenue maximization, relaxed subadditive pricing q satisfies p(x)^2 ≤ q(x) ≤ p(x) and is computable by dynamic programming in O(n^2) time.The relaxation addresses the coNP-hard optimization problem in machine-learning models.Source: §5.3 · pp. 39–40

5.4 Fair and Truthful Pricing

• Fair and truthful data markets use monotonic allocation to ensure truthfulness and randomized polynomial-time algorithms to approximate Shapley-fair seller payments within ϵ with probability 1−δ.Truthfulness holds if and only if increasing the price-rate/bid difference decreases prediction gain.Source: §5.4 · pp. 40–41
• Shapley-value approximations reduce utility evaluations from O(N^2 log N) to O(N(log N) log(log N)) by exploiting group testing and approximate sparsity.Shapley computation is generally exponential, motivating sampling and model-specific methods; replication-robust mechanisms down-weight duplicated data.Source: §5.4 · pp. 41–42

5.5 Privacy Preserving Marketplaces of Data

• Privacy-preserving marketplaces compensate privacy loss through differential-privacy auctions, correlated-data contracts, dependent fairness, and sensitivity-based two-part tariffs.These mechanisms cover buyer accuracy or budgets, correlated sellers, and differentiated prices for sensitive versus nonsensitive attributes.Source: §5.5 · pp. 43, 44, 46
• With n honest buyer types and one adversarial type, query-bundle contracts need at most n contracts, while post-hoc fines and fast approximation reduce adversarial privacy leakage.In general, privacy loss cannot be properly compensated when correlations between private data and cost functions are unknown beforehand.Source: §5.5 · p. 45

5.6 Data Pricing in Novel Applications: Dynamic Data Pricing, Online Pricing and Federated Learning Pricing

• Dynamic and online pricing exploits query overlap and purchase histories: Qirana’s history-aware pricing is arbitrage-free, while mobile crowd-sensing pricing is arbitrage-free and constant-factor revenue-approximating.History-aware APIs use coupons and refunds, whereas Qirana charges only for newly supported data.Source: §5.6 · pp. 47–48
• Federated Shapley values account for participation order across training rounds and satisfy instantaneous group rationality, supporting data valuation in decentralized learning.The value of each participant is summed over all rounds in which the participant contributes.Source: §5.6 · pp. 49–50
• Federated Shapley values guarantee round-by-round fairness and assign zero value when a party adds no utility, while extending approximation methods for computation.For parties i and j, equality holds under the stated utility-equivalence condition; a party receives zero when adding it does not change utility.Source: §5.6 · p. 50
• Information gain serves as the utility function in collaborative machine learning, and ρ-Shapley rewards trade off Shapley fairness, individual rationality, coalition stability, and group welfare.The information gain is defined as I(θ; D) = H(θ) − H(θ|D), representing the reduction in model uncertainty; rewards are assigned as r_i = kψ_i^ρ.Source: §5.6 · p. 50
• Privacy-aware federated-learning pricing compensates users for privacy leakage through a two-stage Stackelberg game, while influence functions reward contributions to linear regression [176].The privacy mechanism targets users most likely to provide reliable data.Source: §5.6 · p. 50

5.7 Summary

• Data-product pricing in marketplaces considers product structures, players, production methods, arbitrage freedom, revenue, fairness, truthfulness, privacy, dynamics, and online pricing, typically integrating these considerations.Source: §5.7 · p. 50

6 Discussion and Open Challenges

• Data pricing remains a fast-growing interdisciplinary area with many unexplored questions, and the proposed challenges are explicitly non-exhaustive.The discussion aims to encourage broader research across economics, marketing, electronic commerce, data management, data mining, and machine learning.Source: §6 · p. 51

6.1 Data Supply Chain: A Grand Challenge

• Data supply chains should connect providers, processors, analysts, products, services, and consumers through value creation, rewards, and pricing feedback.Such end-to-end mechanisms are needed to match, coordinate, and balance data supply and consumption, but remain insufficiently studied.Source: §6.1 · p. 51
• Building sustainable data marketplaces requires domain customization, faster online queries, substitutable processing mashups, price transparency, customer focus, and mediation of privacy and welfare effects.Market platforms can combine market design and system development to connect supply and demand through sharing, discovery, integration, and data mashups.Source: §6.1 · pp. 51–52
• Cross-disciplinary communication and dialogue must be strengthened because data pricing spans economics, marketing, electronic commerce, data management, data mining, and machine learning.This interdisciplinary fragmentation is a major challenge for developing coherent data supply-chain solutions.Source: §6.1 · p. 52

6.2 Some Technical Challenges at the Micro Level

• Micro-level research should connect theoretical relative-pricing models with absolute prices, marketing effects, and experimental validation in practice.User behavior is difficult to model completely, making empirical studies important complements to theoretical investigations.Source: §6.2 · pp. 52–53
• Open technical challenges include valuing and negotiating among parties, enforcing models through accounting and auditing, and developing operational quality guarantees and adversary detection.These problems concern providers, owners, users, and brokers, while transparency and marketplace efficiency depend on effective accounting and auditing.Source: §6.2 · p. 53
• Application-specific pricing must address fairness, truthfulness, and privacy under domain mechanisms and constraints, while dynamic pricing must track changing data demand, supply, and value.Fair pricing is generally exponential in computation time but can be polynomial in kNN models.Source: §6.2 · p. 53