Most warnings about algorithmic collusion tell the same story: sellers using AI to set prices end up coordinating—without explicit communication—to keep prices higher than competition would allow. This is what regulators fear: supra-competitive prices, reduced consumer welfare, and harder-to-detect anti-competitive behavior.
A new study, however, flips the narrative on its head. By analyzing multi-dimensional decision-making—where reinforcement learning (RL) agents set both prices and advertising bids on a platform like Amazon—the authors uncover a surprising outcome: in markets with high consumer search costs, algorithmic “collusion” can lower prices below competitive benchmarks.
Why Search Costs Change the Game
The core mechanism hinges on advertising costs. In a one-dimensional, pricing-only world, collusion benefits sellers by keeping prices high. But when ad auctions enter the picture:
- High search costs mean a large share of consumers only view top-ranked products.
- Winning those positions requires ad bids, which are costs to sellers.
- RL agents, optimizing jointly over price and bid, learn that mutually lowering bids reduces costs more than raising prices increases profit.
- Lower costs → lower optimal prices → higher demand.
The result? A tacit agreement—emerging purely from learning dynamics—to keep bids low and prices attractive.
From Theory to Simulation
The authors start with a duopoly model featuring both sponsored and organic search positions, then compare:
- Pricing-only Nash-Bertrand competition
- Pricing + bidding competition
- Multi-Agent Q-learning with varying shares (θ) of high-search-cost consumers
In classic Calvano et al. (2020)-style settings (θ = 0), RL agents still produce above-competitive prices. But as θ rises, the equilibrium shifts—and at high θ, Q-learning prices drop below competitive levels.
Evidence from Amazon
To test real-world plausibility, the researchers scraped:
- 1,918 keywords across categories
- 2M+ products, millions of daily observations
They estimated search costs via structural models of ranking and demand. Many categories—Clothing, Beauty, Pet Supplies—show high search costs.
Next, they built an algorithm usage index from high-frequency price correlations. The key empirical finding:
Search Costs | Algorithm Usage | Price Effect |
---|---|---|
Low | High | Prices ↑ |
High | High | Prices ↓ |
This negative interaction matches the simulation prediction: algorithms in high-search-cost markets often collude in a way that benefits consumers.
Platform Strategy Implications
For platform managers, the takeaway is nuanced:
- Commission rates: Raising them can recoup lost ad revenue while keeping consumer-friendly prices.
- Ad reserve prices: Risky—raising them above the RL equilibrium bid can trigger even lower bids and lost revenue.
- Information restrictions: Limiting bid data disclosure has minimal effect on outcomes.
From a regulatory standpoint, the message is clear: algorithmic collusion isn’t universally harmful. In multi-dimensional settings, especially where search costs are high, it can increase total surplus.
Why This Matters
Most policy debates treat collusion as a binary evil. This research shows the context matters:
- Dimension of decision-making (pricing-only vs. pricing+bidding)
- Consumer search behavior
- Platform revenue structure
For sellers, the findings suggest that in certain niches, adopting RL for joint price/bid optimization can boost profits and consumer goodwill. For platforms, it reframes the question from “how to stop collusion” to “when and where might it help?”
Cognaptus: Automate the Present, Incubate the Future