When Collusion Cuts Prices: The Counterintuitive Economics of Algorithmic Bidding

TL;DR for operators

Marketplace operators usually worry that pricing algorithms learn the oldest trick in commerce: stop undercutting each other and raise prices. That worry is real. But this paper makes a more interesting point: when sellers use algorithms to optimise both product prices and sponsored-ad bids, collusion can move through the cost side before it moves through the price side.¹

The mechanism is simple enough to be dangerous. Sponsored ads are valuable because impatient shoppers mostly inspect the top results. Sellers therefore bid aggressively for visibility. Those bids are not decorative marketing spend; they become a marginal cost of selling. If competing algorithms learn to coordinate on lower bids, they can reduce advertising costs. When enough shoppers have high search costs, the cost saving can dominate the usual incentive to raise prices. The result is the paper’s counterintuitive finding: algorithmic coordination can produce prices below the full-competition benchmark.

For sellers, the message is not “collude and prosper”, because lawyers have hobbies too. The operational lesson is that joint price-and-ad optimisation behaves differently from price-only automation. In categories where top-of-page placement matters and ad bids are expensive, the bid optimisation layer can change the economics of pricing.

For platforms, the obvious fix is not necessarily the right one. Raising ad-auction reserve prices may fail because algorithms can respond by coordinating on even lower effective bids. Adjusting commission rates is more effective in the paper’s simulations, because commissions let the platform recover value from higher transaction volume when lower prices expand demand.

For regulators and strategy teams, the paper is useful because it breaks a lazy binary. Algorithmic collusion is not automatically “higher prices, lower welfare”. The welfare effect depends on which margin the algorithms coordinate on. Price coordination raises consumer costs. Bid coordination can lower seller costs. When both are present, the winner depends on consumer search behaviour.

The boundary is equally important. The cleanest evidence is theoretical and simulation-based. The Amazon evidence is consistent with the mechanism, not a causal proof. Algorithm usage is inferred from price-correlation patterns, not directly observed. The core learning setup uses Q-learning, with robustness checks for richer settings. This is a strong mechanism paper, not a permission slip for pretending every retail AI system is secretly pro-consumer.

The expensive part of advertising is not the ad; it is the pricing feedback loop

A sponsored product slot looks like a marketing instrument. In the model, it behaves like a cost shock.

That distinction does most of the work.

On an e-commerce platform, sellers do two things at once. They set product prices, and they bid for sponsored placement. A top sponsored position gives access to consumers who may not scroll very far. In a market with low search friction, that placement is useful but not decisive. In a market with high search friction, it can be the market. If consumers mostly inspect the first few results, visibility becomes demand.

Now add competition. If both sellers want the sponsored slot, bids rise. If bids rise, the cost of acquiring a buyer rises. If the cost of acquiring a buyer rises, prices rise. This is the standard competitive benchmark in the paper: advertising competition can push prices above the price-only benchmark because sellers pass through part of their ad cost.

That is already a useful reminder. Sponsored placement is often treated as a separate marketing channel, as if the ad team can optimise bids and the pricing team can optimise margins independently. The paper’s model refuses that convenience. A bid is not just a bid. It changes the seller’s marginal economics, which changes the price that maximises profit.

The twist arrives when sellers do not merely compete once with full information. They repeatedly interact through learning algorithms.

In price-only algorithmic competition, previous work has shown a familiar pattern: algorithms can learn to avoid aggressive price competition and converge toward higher-than-competitive prices. That is the “algorithmic collusion” worry. Zhao and Berman keep that concern, then add the missing marketplace layer: sponsored advertising.

Once bids enter the action space, coordination has two possible targets.

This is why the paper’s result is not “collusion is good now”. That would be a lovely headline for someone trying to fail an antitrust exam. The result is narrower and more useful: when algorithms coordinate over both price and advertising bids, the cost-saving channel can beat the price-raising channel in high-search-cost markets.

The consumer who does not scroll is the hidden economic primitive

The paper’s mechanism depends on search costs. Not the vague “friction” that appears in consulting decks like parsley. A specific behaviour: some consumers consider only the sponsored/top product, while others consider more products on the page.

In the simplified theoretical model, a fraction of consumers has high search costs and considers only the product in the sponsored position. The remaining consumers compare all listed products. The parameter governing this split becomes the key dial. When the high-search-cost fraction is low, sponsored placement matters less because consumers still compare across sellers. When that fraction is high, winning visibility means reaching consumers who may never inspect the rival.

Under full competition, higher search costs make sponsored placement more valuable. That pushes sellers to bid more aggressively. Higher bids raise advertising expense. Higher advertising expense pushes equilibrium prices upward. The competitive benchmark therefore gives a clean result: prices with advertising are at least as high as prices without advertising, and they rise as the fraction of impatient consumers increases.

So far, nothing counterintuitive. If attention is scarce, sellers pay more for it. If sellers pay more for it, buyers eventually feel it.

But collusion changes the bid logic. If two sellers could coordinate directly, they would not want to burn money fighting over sponsored placement. In the fully collusive benchmark, they minimise bids and split the market through the platform’s ranking randomness and demand structure. Lower bids reduce costs. With high search costs, the outside option becomes more threatening because many consumers evaluate only one product before deciding whether to buy at all. In that setting, the monopolist-like collusive seller group has a reason to lower prices to keep platform demand from leaking away.

This creates the crossover result. At low search costs, collusive prices remain above competitive prices. At high search costs, collusive prices can fall below competitive prices because ad-cost savings and outside-option pressure dominate the normal collusive incentive to raise margins.

The paper states this as a formal proposition: there exists a search-cost threshold beyond which the fully collusive price is lower than the fully competitive price. The intuition is cleaner than the notation. When attention is expensive, competition makes sellers overpay for visibility. Coordination stops the bidding war. Once the bidding war stops, prices can come down.

What the Q-learning agents actually learn

The main simulation uses tabular Q-learning in a repeated two-seller environment. Each seller chooses from a discretised set of prices and bids. In the baseline setup, there are 15 possible prices and 10 possible bids. Sellers observe competitor prices, but not competitor bids, which mirrors the practical asymmetry of e-commerce marketplaces: prices are public, ad bids are mostly not.

Each simulation runs until the sellers’ induced strategies remain stable for 100,000 periods, or until a maximum runtime is reached. The paper repeats each parameter setting 1,000 times and averages the results. The search-cost parameter is swept from zero to one in increments of 0.01. This is not a toy one-off run dressed up in a lab coat. It is still a simulation, but it is trying to map the mechanism across the relevant parameter space.

The results follow the mechanism almost too neatly.

When search costs are zero, the model resembles the price-only literature. Q-learning produces prices above the Nash-Bertrand competitive price. The paper reports a market-share-weighted Q-learning equilibrium price of about 1.7 versus a Nash-Bertrand price of roughly 1.47 in that case. That is the familiar collusion concern.

As search costs rise, the pattern changes. Q-learning prices first remain above the competitive benchmark, then fall below it. The bid results explain why: at high search costs, the algorithms learn to coordinate on lower bids than full competition would produce. Lower bids reduce advertising costs. Lower advertising costs make lower prices profitable.

The simulation is therefore main evidence for the behavioural mechanism, not merely a numerical illustration. It shows that independent learning agents, without explicit communication, can discover the cost-saving route.

Still, the paper is careful not to claim that algorithms magically choose benevolence. The agents maximise seller profit. Consumers benefit only because the profit-maximising route runs through lower advertising costs and expanded demand. The consumer surplus result is a consequence, not an objective. That distinction matters, because it tells operators where the result is likely to fail: if lower bids do not expand demand, or if the platform design prevents bid savings from flowing into prices, the nice part of the story evaporates.

The crossover result prevents the simulation from becoming anecdote

A simulation can show that something happens. A mechanism paper must explain why it should happen outside one calibrated example.

The paper’s theoretical crossover does that work. It compares the full-competition benchmark with the fully collusive benchmark. The fully collusive case is not presented as a realistic or legal market arrangement. It is a bound. If algorithmic outcomes tend to lie between competitive and fully collusive outcomes, then understanding where the collusive price crosses the competitive price helps explain when learning algorithms might generate lower-than-competitive prices.

The important result is not that Q-learning perfectly implements collusion. It does not. The important result is that the direction of “more coordinated” changes with search costs.

At low search costs:

At high search costs:

That inequality is the paper in miniature. Collusion usually means higher prices because sellers restrict price competition. Here, collusion over bids reduces a cost that full competition inflates. When search costs make sponsored placement extremely valuable, competitive bidding becomes expensive enough that the coordinated low-bid outcome can support lower consumer prices.

This also explains why the article needs a mechanism-first reading. A summary-first reading would produce a sentence like “algorithms can lower prices through collusion.” Accurate, perhaps. Also dangerously underspecified. The actual claim is conditional on the interaction among search costs, sponsored placement, bid costs, outside-option demand, and learning dynamics. Yes, that is less catchy. It is also how economics avoids becoming vibes with Greek letters.

The appendix mostly asks whether the mechanism survives being made less tidy

The robustness checks are not a second thesis. They ask whether the same mechanism survives when the model relaxes some clean assumptions.

The strongest robustness logic is conceptual. The mechanism does not require the exact two-segment search model. It requires three ingredients: sponsored placement affects consideration, bids are costly, and learning agents explore strategies that can discover lower-bid coordination. Change the modelling wrapper, and the result can move. Remove one of those ingredients, and the mechanism weakens.

This is also why the paper’s generalisation claim should be read carefully. The authors argue that the result should extend to learning algorithms with exploration, because exploration allows agents to drift away from competitive equilibria and discover non-competitive outcomes. That is plausible. It is not the same as proving that every black-box retail optimiser, every auction format, and every platform ranking system will reproduce the result. A useful paper does not need to solve the whole internet.

The platform response is a revenue-mix problem, not an ad-auction reflex

Platforms earn from both advertising and transactions. That matters because lower bids hurt ad revenue but lower prices can expand demand and increase commission revenue.

The paper studies two direct platform responses: changing commission rates and changing ad-auction reserve prices.

The reserve-price result is the more surprising one. A platform that sees sellers coordinating on low bids might try to raise the auction reserve price. In a simple auction story, that sounds reasonable. In this model, it may not work. When the reserve price rises above the Q-learning bid level, the algorithms can coordinate on lower bids below the new reserve, causing the sponsored slot to be allocated without producing the hoped-for ad revenue. The platform has moved the line; the agents walk around it. Very elegant. Mildly annoying, as most platform governance is.

Commission changes perform better in the paper’s simulations. Because commissions tax sales rather than clicks, they allow the platform to recover value from expanded transaction volume. The platform can raise commission rates and still preserve some of the beneficial lower-price outcomes for sellers and consumers relative to full competition.

The business lesson is not “raise commissions”. It is sharper: when seller algorithms reduce ad costs and stimulate demand, platform monetisation should be evaluated across the full revenue stack. A platform that stares only at advertising revenue may fight the symptom and damage the mechanism that increases total marketplace activity.

For marketplace operators, the practical dashboard should separate at least four movements:

This is the part of the paper that speaks most directly to business design. Algorithmic behaviour is not just a competition-policy risk; it is a monetisation-design problem. If the platform changes the wrong lever, it may reduce welfare, seller profit, and its own long-run surplus in one tidy act of managerial confidence.

The Amazon evidence is consistent with the mechanism, not a causal victory lap

The empirical section looks for real-world evidence on Amazon. The dataset covers April 2024 to April 2025, 2,382 highly searched keywords across product categories, and over 2 million products. The authors scraped first-page search results every three hours, generating more than 19,000 requests per day, and combined those observations with data from Jungle Scout and Keepa.

There are two empirical tasks.

First, estimate consumer search costs by keyword market. The paper models how far consumers continue through ranked results, using variation in product positions and sales. Categories differ substantially. “Clothing, Shoes & Jewelry,” “Pet Supplies,” and “Beauty & Personal Care” show consumers tending to stop later in the results, while “Office Products,” “Sports & Outdoors,” and “Tools & Home Improvement” show consumers searching fewer products before purchase. The interpretation is intuitive: if later results are unlikely to reveal much new differentiation, shoppers stop earlier.

Second, infer algorithm usage. The authors do not observe which sellers use pricing algorithms. They construct an algorithm usage index based on correlation between a product’s price vector and the market average price, following prior work. The idea is that algorithmic repricers often react quickly to competitors, creating correlated price movements. With a threshold of 0.5, the mean inferred share of sellers using algorithms across keyword markets is 31.1%, with a standard deviation of 12%.

Then comes the key empirical pattern. Markets with low search costs show the familiar relationship: higher algorithm usage correlates with higher prices. Markets with high search costs show the opposite: higher algorithm usage correlates with lower prices. In regression results controlling for category fixed effects, the interaction between high algorithm usage and high search cost is significantly negative. In the appendix regression table, the interaction is reported as -22.41 without category fixed effects and -26.83 with category fixed effects, both significant at the 1% level.

That is meaningful. It is not definitive.

The paper says the empirical analysis is correlational and cross-sectional. That matters because both search costs and algorithm adoption could be related to unobserved market features. The algorithm usage index is also an inference from pricing patterns, not a vendor log. A seller with correlated prices might be using an algorithm; it might also be manually following competitors, using rule-based repricing, or reacting to common shocks. The authors control for some alternatives, but not all possible ones.

So the right reading is: the Amazon evidence is aligned with the model’s prediction. It gives the mechanism external texture. It does not independently prove that algorithmic bid coordination caused lower prices on Amazon.

That is still valuable. In applied platform economics, “consistent with a precise mechanism across a large dataset” is not nothing. It is just not a courtroom confession.

What Cognaptus infers for operators

The paper directly shows that, in a model of e-commerce sponsored search, multi-agent learning over prices and bids can produce lower-than-competitive prices when consumer search costs are high. It also directly shows that this happens through lower coordinated ad bids, and that platform commission adjustments can be more effective than reserve-price adjustments in the simulated environment.

Cognaptus infers three operational lessons.

First, do not evaluate pricing automation without ad automation. A seller tool that jointly optimises price and sponsored placement is economically different from a repricer alone. If procurement, growth, and pricing teams buy these systems separately, they may misread the combined margin effect.

Second, search-cost segmentation belongs in marketplace governance. A category where shoppers compare twenty products should not be governed like a category where shoppers buy from the first plausible listing. Algorithmic pricing risk changes with attention depth.

Third, platform monetisation should be stress-tested under algorithmic seller behaviour. Reserve prices, commission rates, ad-ranking rules, and disclosure policies are not independent knobs. They form an incentive system that learning agents can adapt to.

The uncomfortable lesson is that “more algorithmic collusion” is not a complete diagnosis. A serious operator needs to ask: collusion over what? Price? Bids? Inventory? Keywords? Fulfilment promises? Once algorithms coordinate across multiple decision dimensions, welfare analysis becomes multi-margin. Annoying, yes. Also reality.

Where this result should not be overused

There are five boundaries worth keeping.

First, the result depends on advertising bids acting as seller costs. If the relevant coordinated variable does not reduce marginal cost, the pro-consumer pathway weakens.

Second, the lower-price outcome depends on high consumer search costs. If shoppers compare widely, sponsored placement has less power, and price-side collusion can dominate.

Third, the main behavioural evidence uses Q-learning. The paper argues that exploration-based algorithms may exhibit similar movement away from competitive outcomes, but that is not a universal theorem for all AI pricing systems.

Fourth, the empirical evidence is not causal. The Amazon results are consistent with the theory and impressive in scale, but they rely on inferred algorithm usage and cross-sectional variation.

Fifth, the model focuses on pricing and advertising. Real marketplaces also involve fulfilment, reviews, private-label competition, keyword targeting, inventory constraints, delivery promises, and platform self-preferencing. Any one of those can add a new strategic margin. The authors explicitly point to platform-owned products and keyword targeting as future directions. Sensibly, they do not try to swallow the whole marketplace in one paper. That would be heroic, and therefore probably wrong.

These limitations do not weaken the central insight. They locate it. The paper is strongest where sponsored placement is costly, consumer attention is shallow, sellers use adaptive optimisation, and platform revenue comes from both ads and commissions.

The useful conclusion is not that collusion became good

The paper’s best contribution is not the provocative outcome. It is the mechanism that makes the outcome possible.

In one-dimensional price competition, algorithmic coordination points toward higher prices. Add sponsored advertising, and coordination gains a second route: lower bids. Add high consumer search costs, and sponsored placement becomes valuable enough that competitive bidding inflates seller costs. Add an outside option, and sellers cannot simply raise prices without losing demand. Under those conditions, algorithms that coordinate on lower bids can generate lower prices and higher consumer surplus.

That is not a moral defence of algorithmic collusion. It is a warning against analysing platform AI with a one-margin brain.

For sellers, the paper says joint optimisation can change the economics of customer acquisition and pricing. For platforms, it says auction design and commission design must be evaluated together. For regulators, it says the harm story needs to inspect the coordinated variable before naming the outcome.

The lazy version of the debate asks whether pricing algorithms are good or bad. The useful version asks which margin they learn to coordinate on, which costs they reduce, which prices they move, and which consumers actually see the result.

That is less convenient. Convenient theories are often just expensive mistakes with better typography.

Cognaptus: Automate the Present, Incubate the Future.