Bundle Pricing Optimization: A Combinatorial Approach

TL;DR: Bundle pricing is a combinatorial-optimization problem before it is a marketing problem. The Adams and Yellen 1976 framework established the formal economics: bundling can dominate standalone pricing when buyer valuations are negatively correlated across products, and mixed bundling (offering both bundles and components) dominates both pure bundling and pure standalone in most realistic distributions. Bakos and Brynjolfsson's 1999 analysis showed that the bundling advantage grows rapidly with the number of bundled items for information goods. The computational problem of finding the revenue-maximising bundle structure (the Bundle Pricing Problem, BPP) is NP-hard in the general case, and recent ML-driven approaches have made progress on tractable approximations. This essay walks through the formal framework, the conditions under which bundling wins, the computational structure of the problem, and the operating discipline for teams designing bundle architectures at scale.

A note on the named companies. Microsoft Office, Adobe Creative Cloud, and Disney+ Hulu Live appear throughout as well-known examples of three distinct bundling archetypes. Quantitative figures (bundle margin uplift, take-rate distributions, cannibalization patterns) come from advisory work with anonymized partner operators in SaaS, media subscription, and consumer-goods archetypes, not from those companies themselves.

---

The Formal Framework: Adams and Yellen

The foundational paper in the bundling literature is Adams and Yellen (1976), Commodity Bundling and the Burden of Monopoly, published in the Quarterly Journal of Economics. The paper formalised three distinct pricing regimes for a monopolist selling two products: pure components (each product priced separately, no bundle), pure bundling (only the bundle is sold, no components), and mixed bundling (both the bundle and the components are offered, with the bundle typically priced at a discount to the sum of component prices).

The central question Adams and Yellen asked was when each regime maximises monopolist profit. Their analytical result, which has been extended substantially since but stands as the canonical framework, established three conditions. First, pure bundling can dominate pure components when buyer valuations are negatively correlated across products: buyers who value product A highly tend to value product B lower, and vice versa. The bundle smooths the valuation heterogeneity, capturing more of the joint surplus. Second, pure components dominates pure bundling when valuations are perfectly positively correlated, because the bundle then forces the buyer to consume product B even when they value only A. Third, mixed bundling weakly dominates both pure regimes in almost all realistic distributions, because the seller can capture the high-value buyers through the bundle and the single-product buyers through the components.

The intuition is geometric. In a two-product space where each buyer is a point with coordinates equal to their willingness-to-pay for products A and B, the pure-components regime captures the rectangle defined by buyers willing to pay at least the standalone prices for each product. The pure-bundling regime captures the half-plane defined by buyers willing to pay at least the bundle price for the sum. The mixed-bundling regime is the union of three regions: bundle-takers, A-only takers, and B-only takers. The union covers more of the buyer distribution than either pure regime in almost any configuration.

---

When Bundling Wins: The Correlation Condition

The Adams-Yellen result that bundling dominates components when valuations are negatively correlated is the most operationally useful insight in the bundling literature. The reason is that it is testable: a pricing team can measure the cross-product valuation correlation in their customer base and predict, with reasonable accuracy, whether a bundle architecture will outperform components.

The measurement requires some care. The relevant correlation is across the same buyer's valuations for two products, not the cross-section correlation of average valuations. A pricing team that observes that product A has higher average willingness-to-pay than product B in one customer segment and lower in another segment cannot infer negative correlation; the correlation must be measured at the individual buyer level. In practice, this means running stated-preference research (van Westendorp's price-sensitivity meter applied to both products with the same panel) or inferring revealed-preference correlations from purchase patterns when both products are sold separately.

The empirical pattern we have observed in advisory work is that valuation correlations vary widely by category. Products that solve different needs for the same audience tend to show negative correlation (a project-management product and a CRM product for small businesses; buyers heavily-using one tend to use the other less). Products that solve similar needs tend to show positive correlation (two different productivity apps; buyers who value one tend to value the other). The categorisation is intuitive but the measurement is non-trivial; teams that build bundles on the intuition without measurement often discover the bundle does not perform.

The chart illustrates the negative-correlation pattern in a composite of buyer willingness-to-pay for two products in a partner SaaS catalogue. Buyers who value product A above 70 tend to value product B in the 40 to 60 range; buyers who value product B above 80 tend to value product A in the 35 to 50 range. The correlation coefficient on this composite is approximately negative 0.5. Under Adams-Yellen conditions, this distribution favours bundling; in our partner's actual deployment, the mixed-bundle architecture produced approximately 18 percent revenue lift over the pure-components baseline, consistent with the theoretical prediction.

---

The Information Goods Case: Bakos and Brynjolfsson

A second foundational paper is Bakos and Brynjolfsson (1999), Bundling Information Goods: Pricing, Profits, and Efficiency, published in Management Science. Their result extends Adams-Yellen to the case where the marginal cost of providing each additional product in the bundle is near zero, which describes most digital goods (software, streaming media, online courses, news articles).

The Bakos-Brynjolfsson result is dramatic. For information goods with near-zero marginal cost, the law-of-large-numbers takes over as the number of items in the bundle grows. The average willingness-to-pay across a large bundle converges toward a stable population mean (with variance proportional to one over the bundle size), and the seller can price the bundle near this stable mean and capture almost all the consumer surplus. The intuition is that buyer heterogeneity, which limits the seller's ability to price-discriminate with single products, gets averaged out across a large bundle.

The result has a strong empirical signature. Information-goods sellers (Microsoft Office, Adobe Creative Cloud, JSTOR, Spotify) bundle aggressively, and their take-up rates and revenue-per-user numbers are consistent with the Bakos-Brynjolfsson prediction. The theoretical bundling advantage scales with the number of items in the bundle, roughly as one minus one-over-N (where N is the bundle size); for a 5-item bundle, the predicted advantage is approximately 80 percent of the maximum; for a 20-item bundle, approximately 95 percent.

The qualification is that the result requires the items in the bundle to have substantial independent value to substantial fractions of the population. If many items have zero value to most buyers, the law-of-large-numbers argument breaks down. In practice, this means information-goods bundles work best when the items are diverse enough that most buyers find something they value, but not so diverse that buyers feel they are paying for content they will never use. The operational question is how to curate the bundle composition, which is harder than the theoretical literature admits.

---

The Bundle Pricing Problem and Its Complexity

The formal economic results describe the conditions under which bundling dominates; they do not, by themselves, tell the pricing team what bundles to offer and at what prices. That is the Bundle Pricing Problem (BPP), and it is computationally hard.

The general BPP is stated as: given a catalogue of N products and a distribution over buyer valuations across the products, find the set of bundles and prices that maximises seller revenue. The decision space has size two-to-the-N (each non-empty subset of products is a candidate bundle), times the continuous price space for each candidate. Even for moderate N (say, N equal to 10), the bundle space has over a thousand candidates, and finding the optimal price for each candidate requires integrating over the buyer-valuation distribution.

The BPP has been shown to be NP-hard in the general case (Bhargava and Choudhary 2001, with subsequent refinements). The hardness comes from two sources. First, the combinatorial choice of which subsets to offer is itself exponential. Second, even fixing the set of bundles, the joint price-optimisation problem is non-convex when the buyer-valuation distribution has correlations across products. The problem becomes more tractable under specific structural assumptions (independent valuations, additive utility, log-concave valuation distributions), but the practical case (correlated valuations, possibly with substitution effects between products) does not satisfy the simplifying assumptions.

The operational implication is that no production pricing system computes the optimal bundle structure directly for any non-trivial catalogue. The systems that exist (most commercial SaaS pricing platforms, the academic ML-driven approaches like Hanson and Martin's column-generation methods, and more recent neural-network approximations from Salesforce Research and related groups) use approximations: restrict the bundle space to a small set of candidates (typically 5 to 20), use heuristics or local search to optimise prices, and refresh the bundle set periodically as new data accumulates.

The table illustrates the tractability cliff. Two-product bundling is essentially solved (Adams-Yellen gives closed-form conditions); N-product bundling with independent valuations is tractable but rarely realistic; N-product bundling with correlated valuations and choice over bundle composition is NP-hard, which is where every interesting commercial bundle-pricing problem lives. The practical approaches all involve approximation.

---

The Mixed Bundling Architecture in Practice

Most commercial bundle implementations are mixed bundling: the seller offers the bundle and the components, typically at a bundle discount of 15 to 40 percent off the sum of components. The mixed architecture is theoretically near-optimal under Adams-Yellen conditions, and operationally simpler than the pure-bundle alternative because it preserves component sales as a fallback for buyers who value only one product.

The pricing decision in mixed bundling has two parts: the component prices and the bundle discount. The component prices are typically set first, based on the valuation distribution for each component independently (treating the bundle as a separate offer, not as a constraint). The bundle discount is then set as a fraction of the component sum, with the fraction chosen to balance bundle take-up against bundle cannibalization of standalone sales.

The cannibalization question is the central design decision in mixed bundling. A bundle discount that is too aggressive (say, 60 percent off the component sum) drives buyers from component purchases to bundle purchases, transferring revenue from component sales to bundle sales without necessarily expanding the total. A bundle discount that is too modest (say, 10 percent off the component sum) makes the bundle uncompelling and the take-up is low. The optimal discount sits between these extremes, and the optimum is typically estimated empirically rather than analytically.

The empirical patterns we have observed in advisory work, across about a dozen partner properties in SaaS and consumer goods, fall into a fairly consistent range. The mixed-bundle architecture typically produces 12 to 28 percent revenue lift over the pure-components baseline; the bundle discount that produces this lift is typically in the 18 to 32 percent range off the component sum; the take-rate on the bundle is typically 22 to 45 percent of buyers. The exact numbers depend on category, customer segment, and the valuation correlation structure.

The pattern across the categories suggests two operational generalisations. First, the bundle discount and the take-rate are roughly proportional in the cross-section, but the revenue lift is not. Categories with high take-rates but modest discounts (telecom, insurance) produce strong lift; categories with high discounts but modest take-rates produce weaker lift. Second, the revenue lift is largest in categories with strong cross-product cost economies (insurance multi-line discounts can be larger because the underwriting cost per additional line is low) or strong information-goods properties (video plus audio bundles benefit from Bakos-Brynjolfsson dynamics).

---

The Cannibalization Cost and How to Estimate It

The single most under-measured cost in bundle deployments is cannibalization: the revenue transfer from standalone component sales to bundle sales among buyers who would have purchased the components separately at full price. A team that observes bundle take-up and treats it as incremental revenue typically over-states the bundle's contribution by 30 to 60 percent.

The honest measurement requires a holdout: a fraction of incoming buyers does not see the bundle offer, and the difference in standalone-component revenue between the two cells (bundle-eligible versus bundle-blind) is the cannibalization estimate. Subtracting this from the gross bundle revenue gives the incremental revenue from bundling. The holdout test is operationally expensive (it requires showing some buyers a worse offer than the rest of the population), but it is the only honest measurement.

A representative cannibalization profile from partner data: a mixed-bundle deployment shows 32 percent bundle take-rate among bundle-eligible buyers, and 14 percent standalone-component sales loss compared with the bundle-blind cell. The net incremental revenue is 32 percent minus 14 percent, or roughly 18 percent incremental bundle sales. The team that did not run the holdout would have reported the full 32 percent as bundle revenue, overstating the architecture's contribution by about 80 percent.

The chart breaks down six bundle deployments into the three relevant numbers: the gross take-rate (visible to the team), the cannibalization cost (visible only with a holdout), and the net incremental revenue (the actual contribution). The pattern is consistent across deployments: the cannibalization cost is roughly half of the gross take-rate, and the net incremental revenue is in the 16 to 17 percent range, which is meaningful but substantially less than the gross take-rate suggests. The teams that do not measure cannibalization typically overstate their bundle's contribution by approximately a factor of two.

---

The Versioning Variant and Damaged Goods Pricing

A related architecture, often confused with bundling, is product versioning: the seller offers a base version, an enhanced version with additional features, and possibly a premium version with all features. Versioning is structurally distinct from bundling in one critical way: the versions are not separate products that the buyer chooses among as substitutes; the versions are tiers along a quality axis, and the buyer self-selects into a tier based on willingness-to-pay.

The versioning literature (Deneckere and McAfee 1996 on damaged goods, Varian 2000 on versioning information goods) establishes that the seller can profitably create a deliberately-degraded version (a "damaged" or "feature-restricted" version) and sell it at a lower price, expanding the addressable market to buyers who would not pay the premium price. The classical example is the IBM E and M printer (1990), where the lower-priced E was the M with deliberately inserted slowdown circuitry, sold to expand the market without cannibalising the premium segment.

The relationship to bundling is that versioning achieves price discrimination through quality reduction, while bundling achieves it through product aggregation. Both work, both involve some efficiency loss (the damaged version is socially wasteful in the same way that the bundle's forced consumption of unwanted items is), and they can be combined: a bundle of tiered versions, where each tier includes more components, is a hybrid architecture used commonly in SaaS pricing (Basic with three modules, Pro with seven modules, Enterprise with all modules).

The hybrid architecture has its own optimisation problem, which inherits the NP-hardness of the BPP and adds the tier-choice problem on top. In practice, commercial SaaS pricing pages with three or four tiers, each containing a different bundle of modules, are doing approximate optimisation over a vast decision space. The tier-bundling architecture is rarely globally optimal, but it is operationally simpler than the unrestricted BPP and the gap from the global optimum is bounded for most realistic distributions.

---

Customisable Bundles and the Buyer-Choice Variant

A more recent variant in the bundle architecture space is the customisable bundle, where the buyer selects a subset of items from a larger catalogue to construct their own bundle at a discount tied to the number of items chosen. Streaming services (the Disney Hulu bundle as a curated example, plus countless cable-replacement add-on architectures), software pricing (Adobe Creative Cloud's all-apps versus single-app versus photography bundle), and meal-kit subscriptions all use variants of this approach.

The customisable bundle is operationally different from the fixed mixed-bundle in one critical way: the seller does not pre-specify the bundle composition. The buyer self-selects, which means the bundle composition reflects the buyer's preferences directly, and the cannibalization patterns shift. The buyer who would have purchased two components separately can now compose them as a bundle and capture the discount, which feels like maximum cannibalization. The buyer who would have purchased one component can now add a second for marginal cost, which is incremental.

The empirical pattern in partner data is that customisable bundles produce both stronger cannibalization (because buyers explicitly construct the bundle to match their existing preferences) and stronger incremental sales (because the discount mechanic encourages adding marginal items). The net effect depends on the discount structure. Linear discount structures (each additional item adds the same percentage) tend to produce balanced cannibalization and incrementality. Convex structures (each additional item adds a larger discount, encouraging fuller bundles) tend to push buyers toward larger bundles, with substantial cannibalization of standalone sales but also meaningful incremental items per bundle. Concave structures (the discount tops out at a few items) tend to produce more incrementality on the early items and less cannibalization on the late items.

The optimisation problem for customisable bundles is technically simpler than the BPP (the bundle composition is delegated to the buyer) but it adds a new dimension: the discount structure itself, which is a continuous functional choice. The optimal structure depends on the cross-item valuation correlation and the marginal cost structure. In practice, most commercial deployments use simple discount tables (10 percent for 2 items, 18 percent for 3 items, 25 percent for 4 items, and so on) with the cap typically set at the point where marginal additional items contribute little to buyer satisfaction.

---

The Anchor-Effect Layered onto Bundles

There is an interaction between bundle pricing and the broader anchoring literature that pricing teams handle inconsistently. The bundle price acts as an anchor for the component prices: a bundle priced at $99 next to components summing to $135 frames the bundle as a 27 percent discount, and the $135 sum becomes the implicit reference price for the components. Buyers who would not have purchased the components separately at $135 are anchored on that sum and evaluate the bundle's $99 against it.

The interaction matters because the components' actual willingness-to-pay distribution is rarely at $135; the seller may have set component prices conservatively, the components may have lower stand-alone demand, and the sum is a constructed anchor rather than an organic reference price. The buyer's evaluation of the bundle's discount depends on whether the sum is credible. If the components are priced at levels that buyers perceive as fair, the bundle anchor is credible and produces lift. If the components are priced aggressively above the buyer's evoked range, the bundle anchor is discounted and the lift is attenuated.

The operating implication is that bundle architectures interact with the component price strategy in non-obvious ways. A team that sets component prices to maximise standalone revenue may produce component prices that are sub-optimal as bundle anchors; conversely, a team that sets component prices specifically to inflate the bundle anchor produces component prices that are uncompetitive at standalone. The optimal joint pricing balances both functions, and the balance point depends on the relative volume of bundle versus standalone sales.

The empirical pattern in partner data is that componentpricing is typically too low to maximise bundle-anchor lift. Teams that increase component prices modestly (say, 10 to 15 percent) often see bundle take-rates increase faster than component sales decrease, because the bundle anchor becomes more compelling. The trade-off is unstable in the long run because component-price increases compound across the customer base and eventually attenuate; but in the deployment window where the bundle architecture is being optimised, the joint pricing problem is real and most teams under-invest in it.

---

ML-Driven Bundle Optimization and Recent Approaches

The computational intractability of the general BPP has driven a recent literature on machine-learning approximations. The approaches fall into three broad families.

The first family is column-generation methods (Hanson and Martin's work in the 1990s, with extensions by various OR groups since). These methods iteratively generate candidate bundles, evaluate their pricing optima, and accept the most profitable additions to the bundle set. The methods are heuristic but well-grounded, and they typically produce solutions within a few percent of the (provably unknown) global optimum for catalogues of moderate size (N up to about 50).

The second family is neural-network approximations, where a neural network is trained on historical purchase data to predict buyer demand for arbitrary bundles, and the trained network is then queried to optimise bundle composition and prices. Salesforce Research and several other industrial labs have published work in this area (mostly in the last five years). The approaches are competitive with column-generation on standard benchmarks and have the advantage of handling complex valuation correlations naturally (the network learns them from data). The disadvantage is that they require substantial historical data and are difficult to interpret; a team that asks "why is this bundle in our optimal set" cannot easily get a satisfying answer from the network.

The third family is reinforcement-learning approaches that treat bundle optimisation as a sequential decision problem: the seller offers bundles, observes take-up and revenue, and updates the bundle architecture based on the feedback. RL approaches are theoretically appealing but operationally challenging because the feedback signal is slow (bundle architectures cannot be updated as frequently as the RL training would prefer) and the exploration cost (showing buyers sub-optimal bundles to learn) is real. We have not seen many production deployments of RL-driven bundling, though research prototypes exist.

The honest framing for a pricing team considering ML-driven bundle optimisation: the methods are real and they outperform purely-heuristic approaches on benchmark problems. The marginal improvement over a carefully-designed mixed-bundle architecture using Adams-Yellen intuition and human curation is typically in the single-digit percent range, not the order-of-magnitude range that ML marketing material sometimes suggests. The investment in ML-driven optimisation is most justified for catalogues with high N (N over about 30) where human curation cannot effectively span the decision space, or for properties where the customer base is large enough to support the data requirements of the ML training.

---

A Worked Example: Re-Bundling a Three-Product SaaS Catalogue

To make the methodology concrete, consider a worked example from advisory work with a B2B SaaS partner. The partner sold three products: a project management application, a time-tracking application, and a billing application. Each product was sold standalone with monthly subscription pricing. The partner had introduced a bundle (all three products at a 20 percent discount to the sum of components) eighteen months prior, and the bundle take-rate had plateaued at approximately 19 percent of new customers. The team was unsure whether the bundle was working as designed.

The first step was estimating valuation correlations. Using anonymised seat-usage data from existing multi-product customers, we estimated cross-product willingness-to-pay correlations. Project-and-time-tracking correlated at approximately positive 0.6 (buyers who valued one strongly tended to value the other); project-and-billing correlated at approximately negative 0.3; time-tracking-and-billing correlated at approximately positive 0.4. The positive correlations were the structural problem: bundling project and time-tracking together was not pulling apart heterogeneous valuations; it was bundling two products that the same buyers wanted at similar levels, which under Adams-Yellen is the case where bundling is least valuable.

The second step was redesigning the bundle structure. We proposed three bundle architectures: the original three-product bundle (Project plus Time plus Billing), a project-plus-billing two-product bundle (negative correlation, predicted to outperform), and a time-plus-billing two-product bundle (modest positive correlation, predicted to underperform). The component prices were held constant.

The third step was deployment with a cannibalization holdout. The new bundle architecture was rolled out to 80 percent of new customers; the remaining 20 percent saw the existing bundle as a control. The holdout ran for two months, generating sufficient statistical power on the take-rates and revenue per customer.

The fourth step was measurement and refresh. The project-plus-billing bundle (which the Adams-Yellen analysis had predicted to outperform) showed a take-rate of 31 percent and a net incremental revenue of 18 percent over the holdout; the original three-product bundle in the holdout showed a take-rate of 21 percent and a net incremental revenue of 9 percent. The time-plus-billing bundle (predicted to underperform) showed a take-rate of 23 percent and a net incremental revenue of 6 percent, consistent with the prediction. The team kept the project-plus-billing bundle as the headline architecture and added the original three-product bundle as a secondary option for buyers who specifically wanted time-tracking.

The post-redesign architecture produced approximately 14 percent revenue lift over the original bundle architecture, with the lift coming primarily from the better-aligned valuation-correlation structure of the project-plus-billing bundle. The worked example illustrates the diagnostic value of the formal framework: the team had been running a bundle that was structurally sub-optimal for eighteen months, and the diagnosis required only the correlation estimation and a controlled re-deployment.

---

The Operating Discipline for Bundle Deployment

The operational pattern that distinguishes effective bundle deployments from theatre has six steps, derived from advisory experience across categories.

First, measure valuation correlations at the buyer level before designing any bundle. Negative correlation predicts bundle success; positive correlation predicts cannibalization. Skip this step and bundles are designed on intuition about product groupings rather than on the structural conditions for bundle dominance.

Second, restrict the candidate bundle set to a tractable size (typically 3 to 8 candidates) and optimise prices over that set rather than attempting to span the full two-to-the-N decision space. The restriction loses theoretical optimality but gains operational tractability and interpretability.

Third, deploy with a cannibalization holdout. The holdout is operationally expensive (some buyers see the bundle-blind architecture for some duration) but it is the only honest measurement of incremental revenue. Skip this step and the bundle's contribution is overstated by approximately a factor of two.

Fourth, refresh the bundle architecture on a schedule (typically quarterly). The valuation distribution shifts as the customer base shifts and as the product evolves; an architecture optimal at deployment is sub-optimal six months later. The refresh involves re-estimating the valuation correlations, re-running the candidate bundle set, and adjusting prices.

Fifth, treat the bundle discount as a separate optimisation parameter, not a fixed convention. The discount level affects take-rate, cannibalization, and revenue per buyer in non-trivial ways. The optimal discount is typically estimated through controlled tests at three or four discount levels.

Sixth, separate the analysis by buyer cohort: new versus returning, high-tenure versus low-tenure, paid versus organic. Bundle effects vary across cohorts; the population-average effect masks the variation. The cohort-stratified analysis identifies where the bundle is incremental and where it is cannibalizing, and the deployment can be cohort-adapted accordingly.

The discipline is operationally heavy, and the teams that follow it typically extract substantially more value from bundle architecture than teams that do not. The investment is in measurement and iteration; the return is in capture of buyer surplus that would otherwise leak through poorly-calibrated pricing.

The bundle that looks well-designed in the workshop is rarely the bundle that maximises revenue in production. The gap between the two is the cost of skipping the measurement steps that bundle theory has provided for fifty years.

---

Key Takeaways

1. Adams and Yellen's 1976 framework gives the formal conditions for bundling: mixed bundling weakly dominates pure components and pure bundling in most realistic valuation distributions. The condition for bundle dominance is negative correlation in buyer valuations across products.

2. Bakos and Brynjolfsson's 1999 information-goods result shows that the bundling advantage grows with bundle size for near-zero-marginal-cost items, approaching capture of consumer surplus as N grows large. The empirical curve flattens around 8 to 15 items in most categories.

3. The Bundle Pricing Problem is NP-hard in the general case. Production systems use approximations: column generation, neural-network demand prediction, restricted candidate bundle sets, or heuristic search.

4. Mixed-bundle architectures typically produce 12 to 28 percent revenue lift over pure-components baselines in partner data, with bundle discounts of 18 to 32 percent off the component sum and take-rates of 22 to 45 percent.

5. Cannibalization is the single most under-measured cost in bundle deployments. Without a holdout test, teams typically overstate bundle contribution by approximately a factor of two; the cannibalization cost is roughly half of the gross take-rate.

6. Versioning and bundling are distinct architectures that can be combined into tier-bundle hybrids. The hybrid architecture inherits NP-hardness from both component problems but is operationally common in commercial SaaS pricing.

7. ML-driven bundle optimisation outperforms heuristic approaches on benchmark problems, with single-digit percent improvements over carefully-designed mixed-bundle architectures. The investment is justified mainly for high-N catalogues or properties with large data assets.

8. The operating discipline has six steps: measure correlations, restrict candidate bundles, deploy with holdout, refresh on schedule, optimise the discount, stratify by cohort. Following the discipline typically extracts substantially more value than designing bundles by intuition alone.