The Answer Without Referral Becomes the Web Bargain
A July 2026 arXiv paper measures what happens when AI search answers the user before the user reaches the source. The governance problem is not only attribution. It is whether the web can keep producing public knowledge when the route from query to visit is no longer the default.
The Paper
The paper is Qiaoni Shi, Kai Zhu, and Kai Gu's Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain, arXiv:2607.07652 [cs.CY, econ.GN]. The arXiv API lists version 1 as submitted on July 8, 2026. The PDF title page lists Bocconi University in Milan, Italy, and the PDF metadata reports 74 pages.
The paper belongs beside this site's work on answer engines as front pages, AI search and answer engines, crawler payment controls, AI slop farms, robots.txt, and AI data licensing. Its fresh contribution is empirical: it measures user-side sessions, referrals, and displacement rather than only arguing from publisher complaints or platform design.
What Was Measured
Shi, Zhu, and Gu use URL-level Comscore U.S. desktop clickstream data from October 2024 through July 2025, covering the initial ChatGPT Search rollout. The full Comscore sample contains between 168,467 and 238,315 active U.S. desktop households per month; a balanced sub-panel of 45,386 households appears with sufficient activity across the full window.
The authors compare ChatGPT conversation sessions with traditional search queries from Google, Bing, and Yahoo. They treat a clean outbound referral as a foreground third-party website visit carrying ChatGPT or a search engine in the HTTP referrer, after filtering out self-referential traffic, platform-internal navigation, search-results pages, and machine-initiated requests. The paper is careful that these traffic measures do not prove whether a user was helped; they measure the routing of observable attention.
For displacement, the paper exploits three ChatGPT Search access expansions: paid subscribers on October 31, 2024, free logged-in users on December 16, 2024, and anonymous browsers on February 5, 2025. The preferred design compares each treated cohort with a reweighted control group without pre-expansion ChatGPT or Claude activity.
The Referral Gap
The central number is blunt. In the pooled sample, ChatGPT produces a clean outbound referral in 5.2% of conversation sessions, compared with 31.1% of Google queries. At the household level, 74.4% of 56,578 ChatGPT-active households never produce a clean ChatGPT referral during the ten-month window, compared with 9.6% of Google-querying households.
This does not mean the absorbed sessions were useless. A user may have received a good answer and stopped because the task was complete. That is exactly the economic issue. Traditional search was built around a route: query, ranked links, visit, ad impression, subscription funnel, purchase, citation, or audience relationship. AI search can still cite and link, but it can also satisfy the task before the visit happens. Attribution without a visit does not recreate the same economic event.
Where the Route Breaks
The residual click stream is not simply a smaller version of Google. The authors report that ChatGPT sessions are relatively concentrated in reference and tool-oriented contexts, but click-outs are most likely in developer/technical, academic research, and e-commerce brand contexts. Relative to Google referrals, ChatGPT's residual referrals contain more reference/knowledge, tools/SaaS, academic research, and developer/technical destinations, while containing much less social media and fewer ad-supported websites.
Wider ChatGPT Search access also reduces traditional search queries. The preferred stacked design finds a 9.4% average reduction in weekly Google, Bing, and Yahoo query loads per treated household, growing to 17.0% after twenty weeks. Search-engine referral losses are largest for informational destinations: academic research, reference/knowledge, developer/technical, and news/journalism. Transactional and recreational categories show smaller or statistically indistinguishable referral changes.
Governance Reading
The paper's limit is also its discipline. It observes U.S. desktop behavior, not mobile behavior, consumer surplus, publisher revenue, or long-run content investment. It does not prove that every absorbed answer is harmful, nor that every lost referral would have become revenue. It does show that destination-side traffic records miss a large part of AI-search use because many sessions leave no destination-side trace.
For governance, that changes the unit of negotiation. Publisher licensing, crawler access, source citation, ad-market adjustment, and public-interest search policy cannot be evaluated only by visible referrals. The missing object is the absorbed session: the case where web material helps resolve the task inside the intermediary while the source receives no visit.
The Spiralist concern is that the answer layer can make source labor less visible while making the platform more authoritative. The right receipt is not just a blue link. It is a record of which sources were retrieved, which claims they supported, whether the user clicked through, whether the source was compensated or licensed, and whether categories such as news, reference, and academic research are losing routed attention.
The Receipt
An AI-search referral receipt should name the query or prompt class, retrieval time, source URLs retrieved, source URLs cited, answer claims supported, clicked destinations, unclicked cited sources, crawler identity, license or permission basis, ad or sponsored placement, personalization state, refusal or uncertainty behavior, and whether the session ended inside the intermediary.
The practical rule: if an answer engine can consume the web without routing the user onward, then traffic, licensing, and public-knowledge governance need records for both referrals and non-referrals.
Sources
- Qiaoni Shi, Kai Zhu, and Kai Gu, Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain, arXiv:2607.07652 [cs.CY, econ.GN], submitted July 8, 2026.
- arXiv API record for arXiv:2607.07652, checked for title, authors, subject classes, submission date, and abstract statistics.
- arXiv abstract page for arXiv:2607.07652, checked for authorship, abstract, citation metadata, subject listing, and PDF link.
- arXiv PDF for arXiv:2607.07652, checked for page count, Bocconi affiliation, methods, referral ratios, displacement estimates, category losses, appendix methods, and stated limits.