The Algorithm Isn't Colluding. It's Correcting Your Overconfidence.

Based on the research ofPhilip Hanspach, Geza Sapi, and Marcel Wieting, "Algorithms in the Marketplace: An Empirical Analysis of Automated Pricing in E-Commerce," Information Economics and Policy, 2024

Philip Hanspach, Geza Sapi, and Marcel Wieting studied roughly two months of pricing data for about 2,800 products on Bol.com, the largest online marketplace in the Netherlands and Belgium. Their question was simple and timely: does letting a repricing algorithm set your price push that price up, as years of antitrust commentary have warned, or down. Averaged across the marketplace, they find no robust evidence that algorithmic pricing is associated with higher Buy Box prices at all. The more striking result sits inside products with no competing seller. There, the presence of an algorithmic seller is associated with prices more than 20% lower than human-set prices for similar products. Some patterns in the data are consistent with tacit algorithmic collusion in narrow pockets of the marketplace, but the authors describe this as a fringe phenomenon rather than the rule. The headline fear and the empirical floor point in opposite directions.

Why a Human Manager Overprices a Monopoly Listing

The mechanism the paper points to is not exotic. A manager setting the price on a product nobody else sells has no live competitor quote to anchor against, no price war to react to, and every incentive to protect margin rather than test the market. Pricing research has documented managerial overconfidence in demand forecasts for decades. It shows up as sticky list prices, infrequent repricing, and a bias toward round, comfortable numbers that reflect what a manager believes the product is worth rather than what the market would actually bear. An algorithm has none of that attachment. It re-prices on a schedule, reads recent sales and competitor signals however thin they are, and has no ego invested in yesterday's number. On Bol.com's monopoly listings, correcting that gap between belief and demand is what produces the more than 20% drop. The algorithm is not predatory. It is simply the first actor in the pricing chain willing to admit the old price was too high.

Some pricing patterns look like tacit collusion. The authors call it a fringe phenomenon, not the dominant effect.

The Collusion Fear Has a Real Paper Behind It Too

The worry that pricing software could learn to coordinate is not invented by journalists. The OECD's Competition Division devoted a 2017 background paper, "Algorithms and Collusion: Competition Policy in the Digital Age," to exactly this question, asking whether self-learning pricing algorithms could multiply the forms of tacit collusion antitrust law already struggles to catch, while also acknowledging that dynamic pricing algorithms can improve market efficiency. That concern has empirical support in at least one well-documented market. Stephanie Assad, Robert Clark, Daniel Ershov, and Lei Xu studied Germany's retail gasoline stations after algorithmic pricing software became widely available there in 2017, and found that adoption raised margins, but only at stations that already faced competitors, not at monopoly stations, and in duopoly and triopoly markets margins rose only when every station in the market adopted the software together. An earlier working-paper version of that study put the duopoly effect at a 28% margin increase when both competing stations adopted, against no change when only one did.

Put the two papers side by side and the picture sharpens rather than contradicts itself. Collusion through pricing algorithms is a real, demonstrated phenomenon, but it needed a specific structural condition in Germany: a handful of rivals in the same local market, and most or all of them running the same kind of software at the same time. Bol.com's marketplace is a different shape entirely, thousands of products, far more sellers, far less of the tight local oligopoly structure that made the German gas stations vulnerable. In that setting, the dominant effect the data shows is the overconfidence correction, not the collusion risk the OECD paper and the gasoline study flagged. The lesson is not that algorithmic pricing is always benign. It is that the two outcomes depend on market structure, and most e-commerce marketplaces do not look like a four-station fuel oligopoly in a small German town.

How Repricing Software Actually Behaves on a Live Marketplace

Amazon offers a useful, verifiable look at how an algorithmic seller behaves day to day. Its built-in Automate Pricing tool, free to sellers on a Professional selling plan, lets a seller create rules that compare their price against the Featured Offer or the lowest price on or off Amazon, then adjusts automatically, with price updates typically processed in under 15 minutes. A seller sets a minimum and, optionally, a maximum, and the tool does the rest around the clock. Nothing in that design rewards holding a price up. It rewards winning the Featured Offer slot, which on a marketplace with many sellers usually means moving toward the competitive price, not away from it. That is the same incentive structure Hanspach, Sapi, and Wieting describe for Bol.com's algorithmic sellers in monopoly listings, pressure toward the number the market will actually bear rather than the number a manager hoped it would bear.

Airline revenue management offers a longer and different history, worth knowing precisely because it predates the current debate and shows algorithmic pricing was never purely about collusion risk. American Airlines, under CEO Robert Crandall, built one of the first systematic yield management systems in the mid-1980s, using its SABRE reservation infrastructure to track bookings by fare class and adjust prices as a direct competitive response to the low-cost entrant People Express, which lacked the technology to price as finely. That episode illustrates the same core logic running in the opposite direction from monopoly correction: dynamic pricing as a competitive weapon against a rival, not a shield against competition. Both airline revenue management and Amazon's repricer show that "algorithmic pricing" is not one behavior with one effect. It is a tool that amplifies whatever competitive pressure already exists in the market it is deployed into.

What This Means for Anyone Setting Pricing Policy

The practical error is treating "algorithmic pricing" as a single policy question with a single right answer. The Bol.com data says the dominant effect on a large, fragmented marketplace is correction of human overpricing, concentrated exactly where a manager had the least market feedback to rely on: listings with no competitor. The German gasoline data says the collusion risk the OECD flagged in 2017 is real but conditional, appearing only where a small number of rivals adopt the same technology inside the same tight local market. Amazon's repricer and American Airlines' decades-old yield management system both show the same tool pushing prices toward whatever the live competitive signal demands, upward against a weak competitor, downward against a strong one, with no inherent direction of its own.

For an operator, the question worth asking is not "should we allow algorithmic pricing." It is "what does our market structure do to whatever an algorithm is told to optimize." A monopoly listing with an overconfident manager and a tight oligopoly with a handful of coordinated adopters will respond to the identical software in opposite directions. Policymakers drafting the next round of algorithmic pricing rules would do well to read the structural condition in the German study as carefully as the headline number, and marketplace operators would do well to remember that the single biggest price distortion their pricing algorithm might fix is not a rival's software. It is a manager's own confidence in a number nobody ever tested.