AI shopping agents can steal an aggregator's business, but can they really pocket the same fat margins?
What stops agents from capturing aggregator-level margins in most categories?
This explores why AI shopping and travel agents, even as they pull demand away from aggregators like Tripadvisor or Expedia, may not end up earning the large cut that those aggregators used to take.
This explores why AI agents that take demand away from aggregators may not inherit the aggregators' fat margins. The corpus shows clearly that agents can take the margin away. It says much less about whether agents can keep it, and the gaps in what it covers are informative too. The starting point is Are AI agents making attention-based ads irrelevant for aggregators?. When an agent ranks hotels by the traveler's stated criteria instead of by paid placement, the aggregator's main asset, the ability to sell attention, stops being worth anything. Tripadvisor's 23% drop in Hotels revenue is the example given. There is a catch for the agent, though. Its value comes from refusing to sell placement. If it starts selling placement to earn the old margin, it destroys the trust that drew users away from the aggregator in the first place.
The second barrier is that agents are not yet good enough negotiators to charge a premium. Microsoft's Magentic Marketplace experiments found that frontier models reach near-optimal outcomes only when search conditions are ideal. Outcomes get worse as the market grows, and every model showed a strong bias toward accepting the first offer, weighting speed over offer quality by 10 to 30 times Do LLM agents achieve fair marketplace outcomes when given ideal search?. The researchers traced the welfare loss to plain mistakes in execution, not to agents working for the wrong side Do agent mistakes or misaligned incentives cause marketplace welfare loss?. An intermediary that leaves this much value unclaimed is hard to charge a high fee for.
Third, the supply side adapts. A simulation found that when a platform tunes its rankings and sellers tune their product descriptions to exploit the same agent biases, consumer harm more than doubles, and no coordination between them is needed Can platform ranking and seller descriptions separately harm AI shoppers?. Seller-side pricing agents show a related result. Two LLM pricing agents in a repeated market settled on above-competitive prices without being told to collude Do LLM pricing agents autonomously collude in repeated auctions?. The pattern held even when a larger model was paired with a smaller one, because the size difference produced a stable leader and follower Does model-size difference break LLM pricing collusion?. Taken together, these results suggest the margin can flow to sellers and platforms that learn to work the buyer's agent, not to the agent itself.
The least obvious barrier is that building an agent keeps getting cheaper and easier to copy. Small models handle most agent subtasks at 10 to 30 times lower cost Can small language models handle most agent tasks?. Improving the scaffolding around a model raises performance without retraining it, and the same setup carries over to newer models Can execution harnesses lift model performance without retuning weights?. Better instructions for summarizing an agent's working history cut token use by 26 to 54% for less than $2 of optimization Can prompt optimization cut agent token usage without retraining?. Meanwhile, finding the right agent for a job is becoming a standard indexed lookup instead of a proprietary integration Can semantic capability vectors replace manual agent routing?. Aggregators earned their margins from things that were hard to copy: years of reviews, supplier contracts, and traffic. If the agent layer is cheap to build and easy to swap, competition between agents could compete away the margin they took from aggregators.
One caveat: none of these notes directly measures what agents charge, studies network effects at the agent layer, or examines lock-in with suppliers. That is the evidence that would settle the question. What the corpus shows is that the margin is coming loose from aggregators. Who ends up holding it is still open, and the strongest candidates in this material are sellers, not agents.
Sources 10 notes
As shopping and travel agents handle comparison and ranking based on stated criteria rather than paid placement, they render attention ads irrelevant. Tripadvisor's Hotels segment revenue fell 23% in Q2 2026, with management attributing further decline to structural SEO shifts—a pattern Jarc attributes to agent-mediated ranking.
Frontier models can achieve near-optimal marketplace outcomes, but performance degrades with scale and all models show severe first-proposal bias, favoring response speed over offer quality by 10-30x.
Magentic Marketplace experiments traced welfare shortfalls to agent competence failures in complex negotiation, not to agents serving competing interests. This supports human-in-the-loop designs where agents assist rather than replace human decision-making.
A simulation found that when a platform optimizes its ranking and sellers optimize their descriptions—each separately targeting the same agent biases—consumer harm more than doubles compared to either acting alone, without requiring coordination.
Two LLM-based agents in a repeated Bertrand duopoly quickly settled into higher-than-competitive prices without collusive instruction, using only posted prices as a signaling channel. Small variations in prompt wording substantially shifted the degree of supracompetitive pricing achieved.
Show all 10 sources
Pairing a 32B with a 14B pricing agent maintains collusive price elevation near homogeneous levels through a clear leadership hierarchy, where capability asymmetry establishes focal coordination rather than breaking collusion like other heterogeneities do.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
StateM improves Terminal-Bench 2.1 accuracy across multiple models by optimizing execution systems around frozen weights. The same runbook transfers to newer models without modification, achieving 95.3% on GPT-5.6 and lifting DeepSeek-V4 Flash by 5.4 points.
Microsoft's ACON method reduces agent token consumption and improves task success by iteratively refining compression guidelines through failure analysis, not weight updates. This approach works across proprietary LLM agents and costs under $2 per benchmark.
Versioned Capability Vectors embedded in HNSW indices couple semantic matching with policy and budget constraints, making capability discovery a first-class operation that scales sub-linearly as agent heterogeneity increases.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLM-Agent Interactions on Markets with Information Asymmetries
- Vertical tacit collusion in AI-mediated markets
- Agentic markets
- Algorithmic Collusion by Large Language Models
- Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets
- Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders
- An Economy of AI Agents
- On the Fragility of AI Agent Collusion