Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain

AI Search Rewrites the Web's Economic Bargain

Traditional search engines like Google act as a bridge. They send users to websites to find answers. AI search, like ChatGPT, often gives the answer directly in the chat. This means users do not click through to the original website. This study shows that AI search is significantly reducing the amount of traffic sent to websites. This effect is especially strong for informational and research sites.

The vanishing referral bargain

For two decades, the internet has operated on a relatively stable economic bargain. When a user expresses an information need through a search engine, the engine acts as an intermediary. It ranks and monetizes the query. Ultimately, it routes the user to a destination website. This "routing" is the lifeblood of the web. The search engine captures value at the moment of the query. Meanwhile, the content producer captures value through the subsequent visit. That visit can be turned into ad impressions, subscriptions, or sales.

The core question investigated by Shi, Zhu, and Gu is whether AI search breaks this cycle. Does an interface that synthesizes answers from the web actually fulfill the user's need inside the chat? Or does it still serve as a gateway to the open web? If the latter is not true, the fundamental link between search, traffic, and content production may be unraveling.

A shift from routing to resolution

The conventional model of digital intermediation assumes that discovery must culminate in a visit. As illustrated in, traditional search architectures focus on routing attention through a click.

The intermediary's job is to point the user toward the right destination.

However, the authors argue that AI search introduces a paradigm shift. Instead of simply returning ranked links, an AI interface can retrieve information from various websites. It then synthesizes this into a coherent response. This satisfies the user's need entirely within the chat window. This moves the economy from "routed visits" to "residual referrals." In this new model, a click is no longer the primary goal. It is merely an optional byproduct of the conversation. If the AI can answer the question, the incentive to visit the source vanishes.

Measuring the absorption of intent

To test this shift, the researchers analyzed URL-level Comscore U.S. desktop clickstream data from October 2024 through July 2025. This dataset allowed them to track the same households across different platforms. This provided a rare view of how individual users switch between traditional search and AI chat.

The study employed a "stacked difference-in-differences" design (a statistical method used to estimate causal effects by comparing groups across different time periods). The authors exploited three specific expansions of ChatGPT Search access. These rollouts went first to paid subscribers, then free logged-in users, and finally anonymous browsers. This helped estimate how wider access to AI search causes the displacement of traditional search. They also used a GPT-4o-based classifier to categorize thousands of destination domains. This ensured they could see exactly who was losing traffic.

The researchers did not just look at raw clicks. They looked at "information-seeking occasions." For Google, this is a search query. For ChatGPT, this is a conversation session. By comparing these units within the same household and week, they isolated the effect of the interface itself. This removed the bias of different users having different personal habits.

High absorption and specialized residuals

The results reveal a massive divergence in how these platforms handle user intent. The authors report that ChatGPT produces an outbound referral in only 5.2% of conversation sessions. This stands in stark contrast to Google’s 31.1% referral ratio for queries .

This means most ChatGPT interactions end without a click. In fact, the study finds that 74.4% of ChatGPT-active households never produce a single clean referral during the observation period [Table 1].

Crucially, the traffic that does escape ChatGPT is not a representative sample of the web. The authors find that ChatGPT's residual referrals are highly selective. While Google serves a broad spectrum of the web, ChatGPT's clicks skew toward specialized destinations. These include academic research, developer/technical sites, and tools/SaaS providers .

Furthermore, the traffic avoids much of the ad-supported web. It favors instead subscription-based or nonprofit/public models .

Perhaps most significantly, the study quantifies the "displacement" effect. The authors report that wider access to ChatGPT Search reduces traditional search queries by 9.4% on average. This loss grows to 17.0% after twenty weeks of exposure . This loss is not felt equally. The decline in search-engine referrals is most severe for informational categories. Academic research referrals declined by 32.8%. Reference/knowledge sites saw a 26.5% decline .

The cost of answering without referring

The findings suggest a profound reallocation of digital attention. If these patterns generalize, the economic foundation of the "open web" faces a structural threat. The very content that makes AI search useful is high-quality research, news, and technical documentation. This is the same content most likely to be "absorbed" by the model. This leaves the original creators with diminished traffic and fewer opportunities to monetize their work.

There are two immediate implications for the digital economy. First, for publishers, the "attribution" provided by a citation may not be enough. It may not replace the economic value of a visit. If the AI satisfies the need without a click, traditional mechanisms for building subscriber relationships are bypassed. Second, for advertisers, the contraction of traditional search queries suggests a shrinking inventory of routed attention. This could force a shift in how brands reach consumers.

The paper does not explore the long-term impact on content production itself. Specifically, it does not look at whether creators will continue to produce high-quality information if the "referral bargain" no longer holds. A critical next step for researchers would be to investigate the supply side. They must determine if the decline in routed traffic leads to a measurable decrease in the volume of new digital content.

Figures from the paper