PATRICK COLLISON: Internet businesses that monetize demand routing built on short attention spans may see the nature and value of that routing change substantially

The Gist

Patrick Collison says many internet companies make money by charging sellers to be shown first, because shoppers often click the first thing they see. That system has real costs, though it is not all waste. Agents that read every option change that, so the business of routing shoppers may change a lot, in ways no one can yet predict. This packet steelmans the argument Patrick Collison makes in this X post, in its best light, for analysis; it is not an endorsement of his conclusions.

Conclusion

Because much internet revenue comes from routing demand toward paid placements shown to inattentive users, and because personal agents evaluate options differently from hurried people, the nature and value of demand routing may change substantially, although the direction and size of the change remain uncertain.

Premises

1. Many internet businesses earn revenue by routing demand through sponsored placement and ranking, in settings where many users choose among the first options they see.
2. Amazon's advertising revenue was about $68.6 billion in 2025, most of it from sponsored product listings shown alongside search results.
3. A study of Amazon search results across four marketplaces found that about 30 percent of results were sponsored, that a sponsored result usually appeared before the top organic result, and that top sponsored results were often costlier than the top organic results.
4. Paid placement is not pure waste: willingness to pay can signal quality or novelty, and one structural study of Amazon found that removing ads would help consumers and sellers if commissions stayed fixed but would harm them if the platform reset commissions optimally.
5. Agents read options differently from hurried people: in one controlled audit, AI shopping agents penalized sponsored tags, although they still showed strong position effects and favored platform endorsements.

Assumptions

• In August 2026 the FTC and 22 states sued Amazon, alleging hidden surcharges in its ad auctions; these are allegations that Amazon denies, and they concern advertisers rather than shoppers directly.
• Routing may move into the agents themselves: ChatGPT began showing labeled ads in February 2026 and has expanded them to more countries, with OpenAI stating that ads do not influence its answers, a commitment that outside parties have not independently audited.
• The post's admitted uncertainty is preserved: it predicts a large change, not its direction.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization (P2–P5 supporting P1 and the conclusion) — Amazon marketplace data (P2, P3, P4) carries the claim about 'many internet businesses', and one controlled audit (P5) stands in for how agents evaluate options generally. Search, social, travel and app-store routing may depend differently on user inattention, and agent behavior varies by model, prompt and version. The hedged wording limits the damage.
• Missing bridge premise (enthymeme) (P5 to conclusion) — Moving from 'agents read options differently' to 'routing value may change substantially' requires an unstated premise that agents will mediate a meaningful share of purchase and discovery decisions. No adoption evidence or time horizon is given, so a key link remains implicit.
• Unsupported causal attribution (inattention) (P1, title and conclusion framing) — The framing that routing revenue is 'built on short attention spans' is asserted rather than shown. Position effects can also reflect rational satisficing under search costs, trust in a platform's ranking, or genuine relevance. P4 partly concedes this. The premises show prominent paid placement, not that its effectiveness derives from inattention.
• Unfalsifiable hedging (low informational content) (Conclusion and A3) — Predicting 'substantial change' without direction, size or timeframe protects the claim from refutation, because nearly any outcome can be read as confirming it. This is intellectually honest, but it limits the claim's testability and usefulness for decisions.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Visibility is scarce, so someone will always pay to be found, whether the evaluator is a human or an agent. Agents are probeable and biased toward position and platform endorsement (P5), and agent providers have incentives to sell placement (A2). The toll booth simply moves upstream, so the real change is concentration of the routing layer, not a change in the nature or value of routing.
• Premise 5 (High impact) — The audit does not show that deployed agents will behave this way. Results may be artifacts of particular prompts, model versions or lab conditions, and sellers can optimize for agents (agent-focused SEO, disguised sponsorship, mimicked endorsements). Any tag penalty could vanish or be gamed away.
• Conclusion (High impact) — The premises establish no adoption rate for agent-mediated purchasing. If most shopping remains human-browsed, or platforms such as Amazon restrict third-party agents, the agent-versus-human difference never reaches the monetized surface and little changes.
• Premise 1 (Medium impact) — Choosing among the first options can be rational satisficing under search costs or trust in the platform's ranking, not a cognitive deficit. If so, the 'short attention spans' foundation is weaker, and agents may inherit the same defaults, making routing more stable than the title implies.
• Premise 4 (Medium impact) — If agents are right to penalize sponsorship, paid placement is mostly an exploit of human inattention, which conflicts with P4's point that willingness to pay signals quality. If agents are wrong to penalize it, the difference from humans is an error, not an improvement. The argument does not say which.
• Premise 3 (Low impact) — Sponsored results being pricier and placed first may reflect advertiser selection (higher-margin sellers advertise more) rather than exploitation of inattention, so the data fit several mechanisms.

Suggested Improvements

• Adoption bridge — Add a premise or estimate of the share of discovery and purchase decisions likely to be agent-mediated, by category and time horizon, and tie the 'substantial' qualifier to a revenue-at-risk range. This is the missing link between documented agent behavior and a change in routing economics. It would also give the claim a timeframe.
• Falsifiability — Add at least one conditional, testable sub-prediction, e.g. how sponsored click share or ad revenue per query changes as agent traffic passes a given share, or the conditions under which routing rents would migrate into agents. It would turn a near-unfalsifiable claim into one that evidence can confirm or refute.
• Evidence base for agents — Cite replications or additional audits across models, categories and time, ideally including deployed commercial agents and not only controlled setups. P5 carries the mechanism of change but rests on a single audit of a fast-moving target.
• Breadth beyond Amazon — Include routing evidence from search advertising, travel aggregators, app stores and local services, and distinguish commission-based from ad-based models. The thesis concerns many internet businesses, but the evidence is largely Amazon-specific.
• Mechanism of inattention — Provide evidence separating inattention from rational delegation to rankings, such as click elasticity by rank under attention or time-pressure manipulations. It would support the 'built on short attention spans' framing, which is currently asserted and partly undercut by P4.
• Agent economics and incentives — Analyze who funds and controls agents (labs, retailers, browsers, payment firms) and how seller adaptation and platform access restrictions would respond. The direction of change depends on these incentives, and A2 suggests routing may relocate rather than shrink.

Scenario Tests

• Agent adoption for ordinary purchases stays marginal for several years, and most shopping remains human-browsed. (Challenges) — The conclusion would be true only trivially. Existing routing economics would stay largely intact, which shows how much rides on the unstated adoption premise.
• Large platforms and AI providers sell placement inside agent answers (as with ChatGPT's labeled ads in A2), and sellers learn to optimize for agent heuristics. (Supports) — The nature of routing changes (new gatekeepers, less visible and harder-to-audit formats), and its value may persist or grow. This is consistent with the 'substantial change, uncertain direction' claim, though it is a relocation of routing more than its erosion.
• Agents widely adopted and sponsored-tag penalties persist across models, compressing willingness to pay for visible placement while increasing price comparison. (Supports) — The value of attention-based routing falls substantially, and platforms shift toward commissions or fees for agent access, which fits the P4 commission-reset dynamic.
• Subsequent audits show that the sponsored-tag penalty disappears or reverses with new model versions or commercial deals between agent providers and retailers. (Challenges) — P5 would no longer supply a mechanism for change, leaving the claim resting on adoption and A2 alone.
• Major marketplaces block or throttle third-party shopping agents and keep routing inside their own interfaces. (Challenges) — The human-facing, attention-based model continues to govern the monetized surface, so the agent-versus-human difference has limited economic effect.
• An independent audit finds that ads in agent answers measurably influence recommendations despite the stated commitment. (Neutral) — It confirms that routing can migrate into agents, changing who captures rents and how trust operates, but it does not settle whether total routing value rises or falls. The claim of change survives while its direction stays open.

Coherence & Relevance

The argument is coherent as a cumulative, hedged case for possibility. P1–P3 establish that paid routing is large and structured around position, P4 explains why the direction is unclear, P5 and A2 offer a mechanism and a countervailing relocation scenario, and the conclusion is scoped to match. A1 is tangential, since it concerns advertiser-side auction practices and not the agent mechanism. The central gap is the absence of any premise on how widely agents will be adopted. The conclusion is well calibrated and honest, but its safety comes from being weak: it is supported as a plausibility and gives little guidance on what will actually happen.

• Many internet businesses earn revenue by routing demand through sponsored placement and ranking, in settings where many users choose among the first options they see. (Strong) — Establishes the foundation of the argument, but the attention-span mechanism is asserted rather than shown, and 'many' is unquantified.
• Amazon's advertising revenue was about $68.6 billion in 2025, most of it from sponsored product listings shown alongside search results. (Moderate) — Shows scale and stakes but not fragility. It is a necessary condition for a large change, not evidence that one will occur, and is consistent with entrenchment.
• A study of Amazon search results across four marketplaces found that about 30 percent of results were sponsored, that a sponsored result usually appeared before the top organic result, and that top sponsored results were often costlier than the top organic results. (Moderate) — Documents the current structure of paid routing. It says nothing about agent behavior or future adoption, and it generalizes from Amazon to the wider category.
• Paid placement is not pure waste: willingness to pay can signal quality or novelty, and one structural study of Amazon found that removing ads would help consumers and sellers if commissions stayed fixed but would harm them if the platform reset commissions optimally. (Strong) — Directly supports the conclusion's refusal to specify a direction and guards against a naive 'agents will kill ads' story. It is tied to one model and does not address agent-mediated settings.
• Agents read options differently from hurried people: in one controlled audit, AI shopping agents penalized sponsored tags, although they still showed strong position effects and favored platform endorsements. (Strong) — The only premise on the mechanism of change, but it is a single snapshot with mixed findings and no adoption premise linking agent behavior to market-wide routing.

View this argument on LogicFirst.ai