The Prisoner's Dilemma of Discounting: Why Every Retailer Promotes and Nobody Wins

TL;DR: The permanent sale is not a failure of discipline. It is an equilibrium. Two retailers facing the same switchers are in a prisoner's dilemma: promoting is each firm's dominant move, and both end up worse off than if both had held price. Repetition can rescue restraint, but only when the shadow of the future is long enough, and the folk theorem (Friedman 1971; Fudenberg and Maskin 1986) makes the condition precise: cooperation survives only if the discount factor clears a threshold set by the one-period gain from cutting. Varian's model of sales (1980) shows the deeper point, that randomized, intermittent discounting is itself the equilibrium once a market mixes informed price-shoppers with uninformed loyalists, and Lal (1990) shows national brands alternate promotions specifically to hold off private label. When a firm tries to quit unilaterally, the game punishes it: Procter and Gamble lost share during value pricing (Ailawadi, Lehmann and Neslin 2001), and J.C. Penney's net sales fell roughly a quarter in fiscal 2012. The exit is real, but it runs through the parameters, detection speed, discount factors, punishment credibility, and the informed-consumer share, not through willpower.

A note on the named cases. Procter and Gamble, J.C. Penney, and Alibaba appear as documented public episodes; their figures come from company filings, the peer-reviewed studies cited, and contemporaneous reporting. Quantitative figures attributed to advisory work are anonymized order-of-magnitude observations from partner operators, not from any named company.

---

The Sale Nobody Chose

Sit in enough quarterly pricing reviews and a particular kind of meeting starts to repeat. A category manager has the margin math in front of her. She knows the weekly promotion is expensive, that most of the volume it moves would have sold anyway, that the discount trains the base to wait for the next one. She can recite the case against it better than anyone in the room. Then she runs the promotion again, because the alternative, being the only chain at full price the week a rival goes to forty percent off, is worse than the disease.

Nobody in that room wants the sale. Each firm would prefer a world where everyone held price and kept the margin. The individual best response, given what the rival is about to do, is to promote anyway. The outcome is a market where discounting is constant, deep, and universal, and where every participant privately regards it as value destruction. That gap, between what each firm wants collectively and what each firm does individually, is not a lapse of nerve. The gap is the signature of a specific game.

The claim of this essay is that persistent promotion is an equilibrium, not a marketing failure, and that treating it as a failure is why so many attempts to fix it fail. A firm that believes its discounting is a discipline problem will reach for discipline: a memo, a margin target, a resolution to promote less next year. A firm that understands discounting as an equilibrium will reach for the parameters that hold the equilibrium in place, because those are the only things that move it. The distinction is not academic. The discount-engineering view of promotions treats the problem inside one firm's profit-and-loss, and I have argued that case at length in a companion essay on when promotions destroy lifetime value. This essay is about the other half of the problem, the one that survives even after a single firm has done its own math perfectly: the rival is still there, and the rival gets a vote.

The One-Shot Dilemma, and Why It Predicts Too Much

Start with the simplest version, two retailers selling substitutable goods to a shared pool of shoppers, each choosing once whether to hold price or run a promotion. The structure is the most famous object in the field.

Write the payoffs as a profit index, a number that encodes the ordering of outcomes rather than a measured margin. If both firms hold price, each earns a comfortable 10. If both promote, they split the same demand at lower prices and each earns 6. If one promotes while the other holds, the promoter takes share and earns 12 while the holdout, stranded at full price, earns 4.

Read the table from either firm's seat. If the rival holds, promoting earns 12 instead of 10, so you promote. If the rival promotes, promoting earns 6 instead of 4, so you promote. Whatever the rival does, promoting pays more. That property has a name, and it is the strongest prediction available in the theory.

Because promoting dominates for both, the only equilibrium is the bottom-right cell, mutual promotion, at 6 apiece. The tragedy is that both firms can see the top-left cell paying 10 apiece and neither can reach it. The ordering that produces this trap is precise. Writing the temptation payoff as T, the reward for mutual restraint as R, the punishment for mutual promotion as P, and the sucker payoff as S, a promotion game is a genuine dilemma when

$$T > R > P > S \quad\text{and}\quad 2R > T + S.$$

The first inequality makes promotion dominant; the second makes mutual restraint efficient, so the firms are not merely unlucky, they are trapped. In our numbers 12 > 10 > 6 > 4, and 2(10) = 20 exceeds 12 + 4 = 16. The Nash equilibrium is a place neither firm would choose if it could choose for both.

Talking does not help. Two managers could meet, shake hands, and promise to hold price, and the promise would change nothing, because when the moment comes each still earns more by promoting regardless of what the other swore. A promise that it is in your interest to break is not a commitment, it is a wish. The one-shot model matters here for a first reason: it explains why exhortation fails. The model also breaks here, because it predicts too much. If defection were truly dominant every period, we would see a constant low price, not the intermittent, deep, oddly rhythmic discounting that real categories actually show. The permanent forty-percent-off sale that resets every Thursday is not what a static dilemma predicts. Something in the repetition is doing work.

The Shadow of the Future: Tit-for-Tat and the Folk Theorem

Retailers do not meet once. They face each other every week, indefinitely, with no known last round. Repetition is not a detail; it changes the solution set entirely, because a firm that undercuts today can be punished tomorrow, and the fear of tomorrow can make restraint rational today.

The classic demonstration is Robert Axelrod's, whose computer tournaments pitted strategies for the repeated prisoner's dilemma against one another. Axelrod and Hamilton (1981, Science 211(4489), 1390-1396) and then Axelrod's book The Evolution of Cooperation (1984) reported that the winner was almost embarrassingly simple: tit-for-tat, which cooperates on the first move and thereafter copies whatever the rival did last. Tit-for-tat never wins a single exchange, since it never defects first, yet it accumulates the highest total because it elicits cooperation from anything capable of it and punishes exploitation immediately. The lesson practitioners took, correctly, is that reciprocity plus a fast, legible response can hold a truce that no contract enforces.

The power to hold a rival in check can rest on the power to bind yourself first.

, Thomas Schelling, The Strategy of Conflict (1960), paraphrased

The general result behind Axelrod's tournaments circulated as folklore among theorists before anyone published it, which is how it got its name.

Friedman (1971, Review of Economic Studies 38(1), 1-12) gave the first rigorous version for supergames, and Fudenberg and Maskin (1986, Econometrica 54(3), 533-554) generalized it to the discounted case. The content is double-edged, and both edges matter for a pricing team. On one edge, repetition rescues cooperation: restraint can be an equilibrium. On the other, almost any outcome can be an equilibrium, so the theorem tells you that a truce is sustainable without telling you whether it will hold. Sustainability, not selection, is what the math delivers.

The discount-factor threshold

The most useful thing the folk theorem gives an operator is a number. Take the simplest punishment, a grim trigger: both firms hold price until someone promotes, and any promotion is answered by promoting forever after. Holding forever pays π^c each period, worth π^c/(1-δ) today, where δ is the discount factor, the weight a firm puts on next period's profit relative to this one. Defecting once grabs the one-period temptation π^d, then collapses into the punishment π^p forever. Restraint is rational exactly when the patient path is worth at least as much as the grab-and-collapse path:

$$\delta \;\geq\; \frac{\pi^d - \pi^c}{\pi^d - \pi^p}.$$

The intuition comes before the algebra. The numerator is what you gain by cheating this week. The denominator is what you gain by cheating plus what you lose once the war starts. Cooperation holds when the future you are risking outweighs the one week you would steal. Plug in the payoffs from Table 1, with π^c = 10, π^d = 12, and π^p = 6, and the threshold is δ* = (12-10)/(12-6) = 1/3. Two firms that value next quarter at least a third as much as this one can, in principle, hold price forever with nothing but the credible promise of retaliation. Let the discount factor slip below a third and the truce is no longer an equilibrium; promotion returns not because anyone lost discipline but because the arithmetic flipped.

The theory earns its keep here, because every term in that fraction maps to something a real business can see and sometimes move. A larger one-period temptation π^d, deeper carts, higher basket values, a rival with slack capacity, raises the threshold and makes truces fragile. A harsher credible punishment π^p lowers it. And δ itself, the shadow of the future, is not a constant of nature. Quarter-end quota pressure, a private-equity owner demanding this year's number, high customer-acquisition costs that make every sale feel urgent: each shortens the effective horizon and pushes δ toward the cliff. Stigler (1964, Journal of Political Economy 72(1), 44-61) added the parameter the grim-trigger story hides, detection. A truce is only as strong as the speed with which a cut is noticed, because a defection nobody sees is a defection nobody can punish. Opaque, negotiated, or promotional prices lengthen the detection lag and quietly raise the bar for cooperation.

When the truce breaks: price wars as punishment

Real cartels and tacit truces do not hold price smoothly and then hold it forever. They are punctuated by episodes of vicious price-cutting that look, from outside, like the cooperation breaking down. The repeated-game literature says the opposite: those episodes can be the cooperation working.

Green and Porter (1984, Econometrica 52(1), 87-100) built the canonical model. Suppose firms cannot observe each other's prices directly and see only their own falling sales, which might mean a rival cheated or might mean demand simply dropped. A truce that punished only confirmed cheating would be unenforceable, since cheating is never confirmed. The equilibrium instead punishes the symptom: if the observed price or quantity falls below a trigger, all firms revert to a price war for a fixed spell, then return to restraint, whether or not anyone actually defected. Price wars, in this account, are not failures of collusion. They are the scheduled punishments that make collusion credible under imperfect monitoring, and they fire on bad demand draws that no one caused.

Rotemberg and Saloner (1986, American Economic Review 76(3), 390-407) turned the intuition in a direction that still surprises people. Ask when the temptation to cheat is largest, and the answer is when demand is high, because a boom fattens the one-period prize π^d from grabbing the whole market today. To keep the truce incentive-compatible through a boom, firms must cut prices during good times, precisely when naive intuition expects them to raise prices. Countercyclical markups, discounts that deepen when demand is strongest, fall straight out of the model. Abreu (1988, Econometrica 56(2), 383-396) completed the toolkit by characterizing the most effective punishments, the stick-and-carrot penal codes that make deviation unprofitable with the least collateral damage, which is the formal reason a credible, sharp, temporary response beats a vague threat of endless war.

Why Retailers Randomize Instead: Varian's Model of Sales

The dilemma-and-folk-theorem story explains restraint and its collapse, but it still models the choice as binary, hold or promote. Look at an actual category and we see something more textured: prices that are neither constantly high nor constantly low but that jump around, deep sales appearing on no fixed schedule, the same item at full price this week and thirty percent off the next. That pattern is not noise on top of an equilibrium. It is the equilibrium, and Hal Varian explained why more than four decades ago.

Varian (1980, American Economic Review 70(4), 651-659) asked how a market clears when consumers differ in one respect: some are informed and will buy only from the cheapest seller, while the rest are uninformed and buy from whoever they happen to face. With that single split, no single price can be an equilibrium. Any store posting a steady price high enough to profit from its uninformed shoppers gets undercut for the informed ones; any store posting a steady low price to win the informed leaves money on the table from its captives. There is no stable pure price. What survives is a mixed strategy: each store randomizes its price according to a distribution, and in equilibrium every price in the support yields the same expected profit.

Let c be unit cost, U the uninformed shoppers split across n stores, I the informed who buy only at the lowest posted price, and F(p) the distribution each store draws from. A store charging p always keeps its U/n captives and wins the informed only if it is the lowest of the n draws, which happens with probability (1-F(p))ⁿ⁻¹. Equilibrium requires the store to be indifferent across its whole price range, so profit at any p equals profit at the top price p̄, where the informed are never won:

$$(p - c)\left[\frac{U}{n} + I\,\bigl(1 - F(p)\bigr)^{\,n-1}\right] \;=\; (\bar p - c)\,\frac{U}{n}.$$

Solve for F and you have the equilibrium pattern of sales. The deep discount is the store gambling for the informed pool; the high price is the store harvesting its captives; the randomization is what keeps rivals from predicting and pre-empting each cut. Sobel (1984, Review of Economic Studies 51(3), 353-368) extended the logic to timing, showing how a stock of high-valuation buyers accumulates between sales and gets cleared by periodic markdowns, which is why clearance events cluster in time rather than smoothing out. Narasimhan (1988, Journal of Business 61(4), 427-449) recast the same forces in the language marketers use, loyal segments versus switchers, and derived the promotional intensity a brand should run as a function of how many switchers are in play.

The parameter that governs everything in Varian's world is the informed share, the fraction I/(U+I) of buyers who shop on price. Push it toward zero and the market can support high, stable prices; push it toward one and the market collapses toward marginal-cost competition punctuated by savage sales. Most of what retailers call promotional strategy is, underneath, a fight over that fraction: who gets to treat a customer as a captive and who is forced to treat them as a switcher.

National brands, private label, and the defensive promotion

Varian's shoppers are symmetric. Real categories are not, and the most important asymmetry in modern retail is the national brand facing the store's own private label. Here promotion acquires a second rationale that has nothing to do with stealing from a symmetric rival and everything to do with defense.

Lal (1990, Marketing Science 9(3), 247-262) modeled national brands as promoting not to fight each other but to keep switchers from defecting permanently to private label. In his account, two national brands tacitly alternate their promotions, so that in any given week one of them is on deal and the price-sensitive switcher always has a discounted national brand to reach for, rather than being driven to the cheaper store brand. The promotions look like rivalry between the national brands; functionally they are a coordinated defense of the whole national-brand tier against encroachment from below. Rao (1991, Marketing Science 10(2), 131-144) worked out the asymmetric-duopoly case directly, showing how a stronger brand and a weaker one settle into different promotional intensities, with the results depending on the relative sizes of their loyal followings. The practical reading is that a national brand's discount is often a rent paid to keep a customer inside the branded part of the shelf, and that the arrival of a credible private label raises the equilibrium promotional intensity for everyone above it.

What the Promotion Actually Buys: The Long-Run Evidence

The models say discounting is an equilibrium. They do not say it is a good one, and the empirical literature on what promotions actually purchase, over horizons longer than a single quarter, is where the equilibrium reveals its cost.

The most sobering result is that promotions rarely grow the category they run in. Nijs, Dekimpe, Steenkamp and Hanssens (2001, Marketing Science 20(1), 1-22) studied 560 consumer categories across four years of Dutch national data and found the contemporaneous kick of a price promotion is large, an average category-demand elasticity around 2.21, and almost entirely temporary. The bump dissipated over a dust-settling window averaging roughly ten weeks, and a permanent lift in category demand appeared in only a small minority of categories, on the order of two to four percent. For most categories the long-run effect of all that promotional activity on total demand is statistically indistinguishable from zero. The discounting moves purchases around in time and across brands; it does not, in aggregate, make people consume more toothpaste.

Pauwels, Hanssens and Siddarth (2002, Journal of Marketing Research 39(4), 421-439) decomposed the short-run response into its parts, category incidence, brand choice, and purchase quantity, and reached the same destination from the demand side: the effects are real but transient, with no permanent component surviving once the promotional dynamics play out over a couple of months. The reason the short-run pull is so strong, and therefore so tempting, is visible in the meta-analytic record on price sensitivity. Tellis (1988, Journal of Marketing Research 25(4), 331-341) pooled 367 brand-sales estimates and found a mean price elasticity around −1.76; Bijmolt, van Heerde and Pieters (2005, Journal of Marketing Research 42(2), 141-156) pooled 1,851 estimates from 81 studies and found a mean near −2.62. Demand responds sharply to price in the short run, which is exactly what makes the one-period defection payoff π^d so large, and exactly why the truce is so hard to hold.

The chart puts three different demand constructs side by side, so we should read it carefully: Tellis and Bijmolt measure how a brand's own sales respond to its price, while Nijs measures how a whole category's demand responds, and the constructs are not interchangeable. What they agree on is direction and rough magnitude. Price moves quantity a lot in the short run, and that short-run responsiveness is the engine of the whole promotional equilibrium. The tragedy is that the same elasticity that makes a promotion look like a triumph this week is the reason the category as a whole gains nothing from a decade of them.

Two Firms That Tried to Leave the Game

If constant promotion is an equilibrium, the interesting question is what happens when a firm decides it has had enough and simply stops. Game theory predicts trouble: changing your own strategy without changing the game leaves you playing a different move in the same payoff matrix, which is the definition of the sucker's cell. Two documented episodes show the prediction with unusual clarity.

The first is Procter and Gamble's value-pricing program. Beginning in 1991, P&G cut list prices, slashed coupons and trade deals, and raised advertising across a broad swath of its portfolio, an attempt to move its brands toward everyday low prices and away from the promotional treadmill. Ailawadi, Lehmann and Neslin (2001, Journal of Marketing 65(1), 44-61) assembled data on 24 categories in which P&G held significant share, covering 1990 to 1996, and traced how consumers and competitors responded. Their finding, stated plainly, is that the strategy cost P&G market share, and that the loss was attributable to the very cuts in coupons and deals that defined the program, because promotions in their data worked mainly by pulling in new triers rather than by holding existing ones. Competitors and private label, still promoting, absorbed the switchers P&G stopped courting. The company retained enough loyalists that value pricing may have improved profitability even as share fell, which is the honest complication, but the share loss was real and it was caused by unilaterally leaving the promotional game while rivals kept playing.

The second episode is starker, because it happened faster and in full public view. In early 2012 J.C. Penney, under a new chief executive who had come from building Apple's stores, replaced its century-old promotional calendar with a "Fair and Square" everyday-low-price scheme: no more hundreds of sales a year, no more coupons, just consistent low prices the customer could trust. The logic was impeccable and the customers hated it. Stripped of the ritual of the sale, and of the reference prices that made a marked-down item feel like a win, shoppers left. Comparable-store sales fell 18.9 percent in the first quarter of fiscal 2012 and worsened through the year to a decline of 31.7 percent in the fourth quarter, ending down 25.2 percent for the year. Total net sales fell from $17.26 billion in fiscal 2011 to $12.99 billion in fiscal 2012, roughly a quarter of the company's revenue gone in twelve months. The chief executive departed in April 2013, and the promotions came back.

Both cases carry the same moral, and it is not that everyday low pricing is a bad idea. The moral is that a firm cannot exit a promotional equilibrium by changing only its own move. P&G and J.C. Penney each altered their strategy while leaving the game, the rivals, the customers' trained expectations, the reference prices, the informed-consumer share, exactly as it was. The market did to them precisely what the payoff matrix says it does to a firm that holds price alone: it handed them the sucker's payoff and handed the switchers to whoever kept promoting.

Changing the Game, Not Your Strategy

The way out of a bad equilibrium is never to play the losing move more bravely. The way out is to change the parameters that put the equilibrium there. Every lever below is something a single firm can pull on its own, without a word to a competitor, because the one thing this essay is not recommending is coordination. Explicit agreement among rivals to hold prices or restrict promotions is per se illegal under United States antitrust law and its analogues elsewhere, and the courts have extended their scrutiny to so-called facilitating practices, the price-signaling devices and information exchanges that let rivals reach a tacit understanding without a handshake. Everything that follows is unilateral, lawful, and aimed at the four numbers from the folk-theorem condition.

Shift the informed-consumer share

Varian's model says the market's price competitiveness is governed by the fraction of buyers who shop purely on price. A firm that can convert price-shoppers into identified, loyal, habit-based customers is shrinking the informed share it faces, and shrinking that share is the most durable way to soften the equilibrium. This is the strategic purpose of loyalty programs, memberships, and customer-relationship data that the profit-and-loss rarely credits correctly: their value is not the redemption economics of the points, it is the migration of a customer from the switcher pool, where price is the only variable, into the captive pool, where it is one variable among many. A membership that makes a shopper default to you unless something goes wrong has removed that shopper from the auction Varian describes.

Commit credibly to a price

The one-shot dilemma fails because promises are not credible. The repair is to make the promise self-enforcing, to take an action that removes your own option to break it, so the rival believes it. Everyday low pricing, done right, is exactly such a commitment: a public, durable, hard-to-reverse pledge that turns your price into a fixed feature of the market rather than a move the rival must anticipate. A price-match guarantee is a subtler commitment still, because it converts any rival's cut into an automatic response and so removes the rival's incentive to cut in the first place. The mechanics of credible commitment, why binding your own hands can strengthen your position, deserve their own treatment, and I have given it in a companion essay on commitment devices and credible threats; the point here is only that a commitment which is easy to reverse is not a commitment, and the market will test it. J.C. Penney's everyday-low-price pledge was credible; what it lacked was any change to the customer's trained expectations or the rivals' behavior, so credibility alone stranded it in the sucker's cell.

Coordinate on the calendar, not on the price

There is a lawful cousin of coordination that the whole market already practices, and it is worth naming because it shows the equilibrium can be reshaped rather than only endured. Black Friday, Cyber Monday, and Alibaba's 11.11 Singles' Day are focal points in Schelling's sense: dates so salient that every firm expects every other firm to promote then, so the promotion concentrates onto a shared moment instead of smearing across the whole year. Schelling (1960, The Strategy of Conflict) showed that players who cannot communicate still coordinate on whatever option stands out, and a calendar date that everyone already knows is the purest such option. Concentrating discounting onto a focal event does not require anyone to agree to anything; it requires only that the date be common knowledge, which the culture supplies for free.

The escalation of Singles' Day gross merchandise volume, from RMB 35 billion in 2013 to RMB 268.4 billion in 2019, shows the double edge of a focal point. Concentrating promotion onto one date can discipline the rest of the calendar, since a shopper who knows the big sale is coming in November has less reason to demand a discount in July. It can also create an arms race on the focal day itself, as every firm pours ever-deeper discounts into the moment everyone is watching. Whether a retailer-led calendar cools the everyday equilibrium or simply relocates the war depends on whether the focal event substitutes for routine promotion or merely adds to it, which is an empirical question a well-run operator should be measuring rather than assuming.

The deepest reason the promotional trap is so durable is that it is not one firm's problem to solve. A profit-and-loss can be made perfect inside four walls and the sale will still be there next Thursday, because the sale is a property of the game, not of any player. The same fact is also the good news: games have parameters, and parameters can be moved. The firm that stops asking "how do we have the discipline to promote less" and starts asking "which of the four numbers can we move, and in what order" is the firm that occasionally gets out. The others keep running the sale nobody in the room ever wanted, and calling it a failure of will, when it was an equilibrium all along.

Key Takeaways

1. Persistent promotion is a Nash equilibrium, not a discipline failure. Two retailers sharing switchers face a prisoner's dilemma in which promoting is the dominant move, so the market settles at mutual promotion (6 apiece in Table 1) even though mutual restraint would pay each more (10 apiece). Exhortation cannot move an equilibrium; only its parameters can.

2. Repetition can rescue restraint, but only above a threshold. The folk theorem (Friedman 1971; Fudenberg and Maskin 1986) makes cooperation sustainable when the discount factor clears δ* = (π^d-π^c)/(π^d-π^p), which for the Table 1 payoffs is one-third. Quarter-end pressure, high acquisition costs, and slow detection push the discount factor toward the cliff and break the truce.

3. Intermittent, unpredictable discounting is itself the equilibrium. Varian (1980) shows that once a market mixes informed price-shoppers with uninformed loyalists, no single price is stable and firms must randomize, which is why deep, irregular sales appear in every category and why internal resolve never eliminates them.

4. National brands promote to defend the branded tier against private label. Lal (1990) shows two national brands can alternate promotions so a price-sensitive switcher always has a discounted national option, keeping them from defecting to store brands; the arrival of a credible private label raises equilibrium promotional intensity for everyone above it.

5. Periodic price wars are consistent with tacit cooperation, not proof against it. Green and Porter (1984) show wars triggered by observable demand slumps are the punishments that keep restraint credible under imperfect monitoring, and Rotemberg and Saloner (1986) show the temptation to cut is largest in booms, so markups can move countercyclically.

6. Promotions rarely grow the category. Nijs and colleagues (2001) found the long-run effect of price promotions on total category demand was indistinguishable from zero in the large majority of 560 categories, with a permanent lift in only about two to four percent, even though the short-run elasticity averaged 2.21. The discounting mostly redistributes and pulls demand forward.

7. You cannot exit the game by changing only your own move. P&G lost share during value pricing (Ailawadi, Lehmann and Neslin 2001) and J.C. Penney's net sales fell roughly a quarter in fiscal 2012, because both cut promotion while leaving the informed share, the reference prices, and the rivals untouched. The exit runs through the four parameters, in sequence, and reducing promotion is the last step, not the first.

Further Reading

• Discount Engineering: When Promotions Destroy LTV, the single-firm complement to this essay: how to tell an incremental discount from a value-destroying one inside your own profit-and-loss, once the competitive game is set aside.
• Anchor Pricing and Its Limits: When the Reference Stops Working, why stripping out the sale also strips out the reference price that made a markdown feel like a win, the demand-side mechanism behind the J.C. Penney collapse.
• Platform Cannibalization Dynamics, a game-theoretic model of marketplace versus first-party sales, for the same equilibrium logic applied to a platform choosing whether to compete with its own sellers.