Vertical tacit collusion in AI-mediated markets

AI shopping agents are being deployed to hundreds of millions of consumers, creating a new intermediary between platforms, sellers, and buyers. We identify a novel market failure: vertical tacit collusion, where platforms controlling rankings and sellers controlling product descriptions independently learn to exploit documented AI cognitive biases. Using multi-agent simulation calibrated to empirical measurements of large language model biases, we show that joint exploitation produces consumer harm more than double what would occur if strategies were independent. This superadditive harm arises because platform ranking determines which products occupy bias-triggering positions while seller manipulation determines conversion rates. Unlike horizontal algorithmic collusion, vertical tacit collusion requires no coordination and evades antitrust detection because harm emerges from aligned incentives rather than agreement. Our findings identify an urgent regulatory gap as AI shopping agents reach mainstream adoption.

Introduction. Delegation pervades economic life, from consumers relying on financial advisors to voters entrusting representatives to principals contracting with agents throughout organizations. When the intermediary to whom decisions are delegated has exploitable biases, parties at different levels of the value chain share a common target, and each benefits from exploitation regardless of what others do. This convergence of incentives can produce coordinated harm without any communication or agreement. We study this phenomenon in AI-mediated consumer markets, where platforms control the display environment through which products are presented and sellers control the product-level inputs that AI agents process. Major technology firms are deploying AI shopping agents at scale. ChatGPT’s shopping features are available to over 700 million weekly users, Amazon reports 250 million customers have used Rufus, and Perplexity enables direct purchases through AI interfaces [1–3]. These developments mark a shift in how consumers interact with digital marketplaces. Rather than browsing and comparing products directly, consumers increasingly delegate purchase decisions to AI intermediaries that evaluate options and make or strongly influence final choices by making recommendations. This delegation creates a novel market structure with three distinct classes of strategic actors. The platform controls the information architecture through which the AI agent encounters products, including ranking algorithms, endorsement badges like “Amazon’s Choice” or “Best Seller” [4, 5], and the strategic placement of comparison options. Sellers control the product-level inputs that the AI agent processes, including descriptions optimized for algorithmic interpretation, pricing anchors, and keyword strategies. The AI agent serves as the consumer’s delegate, evaluating options through processes that incorporate both legitimate quality-price tradeoffs and systematic biases. The platform economics literature [6–8] has analyzed two-sided markets where platforms balance buyer and seller interests, but we introduce a key modification: when AI agents mediate consumer decisions, the platform’s incentive to serve consumers honestly is attenuated because exploitation of the AI intermediary becomes profitable. Recent research has documented extensive cognitive-like biases in the large language models that power these shopping agents, and the evidence is both broad and deep. These biases persist across model families, generations, and interface modalities, suggesting they reflect fundamental properties of how these systems process information rather than quirks of particular implementations. Position effects are particularly pronounced and well-documented. Liu et al.[9] demonstrated that LLMs exhibit strong position biases in retrieval tasks, with accuracy differing by 15-25 percentage points between first and middle positions in context windows. Guo et al.[10] confirmed these serial position effects across GPT, Llama-2, and T5 model families, while Shi et al.[11] documented similar patterns in GPT-3.5, GPT-4, Claude-3, and Gemini-Pro. Most directly relevant to our setting, the ACES benchmark evaluated AI shopping agents making actual product choices and found substantial asymmetries: Claude Sonnet 4 selected products from the top half of listings 77% of the time versus 23% for the bottom half, with GPT-4.1 and Gemini 2.5 Flash exhibiting similar patterns, and these biases persisted even when visual interfaces were replaced with text-only APIs [12]. The Magentic Marketplace study found that AI shopping agents exhibit severe first-proposal bias, creating 10-30x advantages for response speed over quality [13]. Beyond position, LLMs exhibit anchoring biases where initial reference points systematically shift subsequent value judgments [14], susceptibility to framing effects, confirmation bias, and keyword manipulation [15], and decoy effects consistent with the 20-30% choice shifts documented in human decision-making [16–18]. Broader evaluations have found that LLMs exhibit cognitive biases in 40% of tested scenarios across diverse model families [19]. Critically, these biases are far more exploitable than analogous human cognitive biases. Human populations are heterogeneous: different individuals have different susceptibilities, responses depend on context and attention, and strategies that sway some consumers fail on others. However, AI agents exhibit biases uniformly and consistently. The same input reliably produces the same biased output across millions of interactions. When consumers delegate to shared AI systems, this homogeneity concentrates what was a diffuse attack surface into a single predictable target. This consistency creates systematically exploitable vulnerabilities. Because AI responses are predictable rather than noisy, exploitation requires no prior knowledge of AI psychology or adversarial intent.

Method. We study a stylized AI-mediated marketplace designed to capture the essential strategic interactions while remaining analytically tractable. The market consists of n = 6 sellers, one platform, and one AI shopping agent representing a population of consumers. This market size reflects typical consideration set sizes in e-commerce [44] while enabling comprehensive robustness analysis across seller counts from 4 to 36. Our simulation models the selection stage of AI shopping, where the agent evaluates a consideration set of products and chooses which to recommend. This is architecturally consistent with deployed systems: Amazon Rufus retrieves products via retrievalaugmented generation then reasons over them [42]; ChatGPT Shopping searches across retailers then synthesizes recommendations [1]; Google AI Mode queries its Shopping Graph of 50 billion listings then generates suggestions [43]. The ACES benchmark [12] adopted the same approach, presenting AI agents with 8 products and measuring choice behavior. We present 6 products, representing a realistic consideration set size after platform retrieval. The platform’s control over ranking (through bid weight w) captures how algorithms determine product visibility; sellers’ control over descriptions (through manipulation intensity m) captures how product listings influence AI interpretation. Each simulation runs for T = 20,000 rounds. We measure outcomes over the final 40% of rounds to ensure convergence, following standard practice in the algorithmic collusion literature. Results are averaged across 100 independent trials with different random seeds. The relatively simple structure is deliberate: following Calvano et al.[21], we seek to identify whether vertical tacit collusion can emerge in a minimal setting, establishing the mechanism before introducing complications.

Agents and Strategies

Discussion. We have demonstrated that AI-mediated markets enable a previously unrecognized form of anti-competitive behavior. When platforms and sellers independently optimize their strategies in markets where AI agents mediate consumer decisions, their profitmaximizing behaviors converge on joint exploitation of AI vulnerabilities without any communication between them. This exploitation emerges through learning rather than intent: platforms discover that certain ranking configurations increase revenue, sellers discover that certain description patterns increase win rates, and neither requires knowledge of AI bias literature or adversarial purpose. The result is vertical tacit collusion, a form of coordination that arises not from agreement but from aligned incentives around a shared vulnerability. The gatekeeper mechanism distinguishes vertical tacit collusion from horizontal algorithmic collusion. In horizontal settings, competing sellers face a coordination problem: each benefits from industry-wide high prices but has individual incentive to undercut. Our vertical setting presents no such coordination problem because platform and sellers are not competitors. They occupy different positions in the value chain with nonoverlapping instruments, and their incentives are naturally aligned around a common exploitable target. No reward-punishment scheme is needed to sustain cooperation because there is no temptation to defect. The welfare implications extend beyond consumer harm to encompass substantial deadweight loss. The platform captures modest rents while destroying considerably more total value, imposing negative externalities on sellers who see profits fall even as they participate in exploitation. This inefficiency arises because platform and seller strategies, while individually profitable, are socially wasteful when combined. Resources flow to manipulation and bid optimization rather than quality improvement, and the information architecture that should help consumers find good products instead systematically misleads them. These dynamics pose immediate challenges for competition policy. Current antitrust frameworks focus on detecting coordination among competitors, with the US Sherman Act and EU competition law prohibiting agreements in restraint of trade and enforcement requiring evidence of communication or facilitating practices [24, 25]. Vertical tacit collusion evades this framework entirely because the platform optimizes its ranking algorithm while sellers optimize their product descriptions, with neither party communicating with the other or even knowing what the other is doing. The harm emerges from aligned incentives rather than coordinated action. The regulatory challenge is compounded by the legitimate purposes that exploitation strategies serve in isolation. Platforms have valid reasons to incorporate seller payments into ranking decisions since advertising revenue supports platform development, and sellers have valid reasons to optimize product descriptions since compelling presentations can benefit consumers. The problem arises not from any single practice but from how practices interact when deployed against AI agents with exploitable biases. Recent research demonstrates that while LLMs exhibit cognitive biases, these biases often manifest differently than in humans, and models may be irrational in ways that diverge from studied human heuristics [36]. This unpredictability compounds the detection challenge because regulators cannot simply apply human-calibrated bias tests to AI agents when the exploitation surfaces may differ from those documented in behavioral economics. Several policy responses merit consideration.