A study published July 8, 2026 by researchers at Bocconi University found that expanding access to ChatGPT Search reduced traditional search engine queries by 9.4% on average, with the decline deepening to 17.0% after twenty weeks of exposure, based on desktop clickstream data covering more than 45,000 United States households. According to the paper, titled "Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain," the authors are Qiaoni Shi, Kai Zhu, and Kai Gu, and the work was posted to the arXiv preprint server under the Computers and Society category.

The research draws on URL-level Comscore desktop clickstream records spanning October 2024 through July 2025, a window chosen because it captures the staggered rollout of ChatGPT Search to different tiers of OpenAI's user base. Rather than relying on survey responses or platform-reported figures, the authors reconstructed individual browsing sessions from raw page loads, HTTP referrers, and timestamps, then compared how often a ChatGPT conversation or a Google query ended with a visit to an outside website.

A different kind of search denominator

Much of the existing public commentary on AI search traffic has measured click-through rates against a results-page impression or a bounded browsing session. The Bocconi team instead anchored its measurement to what it calls an information-seeking occasion: one ChatGPT conversation session, treated as a single unit regardless of how many messages it contained, compared against one Google search query.

Using that framing, the study found that ChatGPT produced a clean outbound referral in only 5.2% of conversation sessions, compared with 31.1% of Google queries pooled across the ten-month sample. The gap was even starker at the household level. Among the 56,578 households that used ChatGPT during the panel window, 74.4% never generated a single clean referral to an outside website across the entire ten months. For Google, the equivalent zero-referral share among 291,747 querying households was 9.6%.

The researchers tested whether this gap could be explained by differences between the kinds of households that gravitate toward each platform, or by broader shifts happening across particular calendar weeks. Restricting the analysis to household-weeks in which the same household used both ChatGPT and Google, and adding fixed effects for household identity and calendar week, moved the estimated gap only marginally, from a raw 31.5 percentage-point difference to 29.0 points. The persistence of the gap under this stricter test indicates that the difference is not primarily driven by who uses each service, but by how each service behaves once a person is using it.

Where the residual clicks go

The clicks that did leave ChatGPT were not simply a smaller version of what Google sent out. When the authors classified destination websites by content type and monetization model, ChatGPT's referral pool showed a heavier concentration in reference and knowledge sites, developer and technical resources, academic research destinations, and general-purpose software tools, while showing markedly less traffic to social media platforms and to e-commerce marketplaces than Google's referral stream did over the same window.

On monetization, the tilt was pronounced. ChatGPT's residual referrals favored nonprofit and public-interest websites, freemium software products, and subscription-supported destinations, while sending proportionally less traffic to advertising-supported websites than Google did. Because advertising-supported sites depend on routed visits to generate ad impressions, this composition detail carries direct implications for how publishers that rely on programmatic or direct-sold display inventory might experience the shift toward conversational search.

The paper also examined how concentrated each platform's referral traffic was across destination websites. Measured in aggregate across the entire sample, Google's referral pool proved between 1.87 and 3.47 times more concentrated than ChatGPT's, depending on which cutoff was applied to the destination universe, with Google's clicks converging heavily on a small number of large platforms. Within an individual household's own week-to-week activity, however, that difference mostly disappeared. Among lightly active households, the concentration of ChatGPT referrals and Google referrals was statistically indistinguishable, and only among the heaviest-referring households did ChatGPT's traffic become more concentrated than Google's. The authors interpret this as evidence that ChatGPT's apparent diversity at the aggregate level is a pattern of different households reaching different niche destinations, rather than any single household spreading its own attention more evenly.

The access-expansion test

To move beyond correlation, the researchers exploited three dates on which OpenAI widened who could use ChatGPT Search. According to OpenAI's own rollout announcement, referenced in the paper, access opened to paid subscribers on October 31, 2024, to free logged-in users on December 16, 2024, and to users browsing without an account on February 5, 2025. Because these expansion dates were set by OpenAI's product rollout schedule rather than by any individual household's decision to adopt the feature, the authors argue the timing offers a cleaner test of cause and effect than simply comparing early adopters against everyone else.

Pooling all three expansions and comparing newly eligible households against a control group with no prior ChatGPT or Claude activity, the study found that traditional search queries - covering Google, Bing, and Yahoo - fell by 3.14 queries per household per week immediately following expanded access, a 9.4% decline relative to the pre-expansion average of 33.51 weekly queries. That effect deepened as more time passed: by the twentieth week after access opened, the decline had grown to 17.0% of the pre-expansion baseline. A narrower comparison, restricted to the December 2024 expansion and pitting newly-eligible logged-in users against not-yet-eligible anonymous users who were already familiar with the ChatGPT interface, produced a smaller but directionally consistent decline of 8.2% after twenty weeks.

The category breakdown of that displacement is where the paper's implications for publishers sharpen. Search-engine referral visits fell most steeply for academic research destinations, down 32.8%, followed by reference and knowledge sites at 26.5%, developer and technical resources at 15.1%, and news and journalism destinations at 13.4%. Transactional and entertainment-oriented categories showed smaller and, in several cases, statistically indistinguishable changes. The authors note that total visits to these informational categories also declined, not just search-engine referrals specifically, suggesting the lost traffic was not simply picked up by another channel the study could observe.

What the paper does and does not claim

The authors are explicit about the boundaries of their findings. The measurements describe observable routing patterns in United States desktop browsing, not a judgment about whether ChatGPT provides more or less value to the person using it. A ChatGPT session that ends without a click could reflect a fully satisfied information need, an abandoned task, or a visit whose referrer information happened to be stripped before the researchers could observe it. The paper draws a similarly careful line around causality in the destination-composition results, describing the pattern as consistent with intent selection - meaning households may simply be bringing different kinds of questions to ChatGPT in the first place - as well as with retention driven by the interface itself.

The study situates its 5.2% referral figure alongside a range of previously published industry estimates, several of which have circulated in marketing trade coverage. Public benchmarks using comparable per-query or bounded-session denominators have generally placed ChatGPT's outbound click rate between roughly 3% and 7%, a band that comfortably contains the paper's pooled estimate. The authors also built a supplementary check using message-level referral data, available only for July 2025 once OpenAI's newer send-tracking endpoint reached sufficient coverage, and found an even lower referral rate of 1.2% for logged-in message sends - an order of magnitude below the search-engine benchmark.

The paper also documents a supply-side pattern relevant to ongoing disputes between publishers and AI companies over content access. Of 3,844 classified domains for which the researchers obtained robots.txt scrapes, 79% blocked at least one major AI crawler from their training-time content. Yet those same blocking domains received higher average ChatGPT referral counts, 6.48 referrals per domain, than non-blocking domains, at 4.48. The explanation the authors offer is a distinction between two separate mechanisms: robots.txt rules govern whether a site's content can be scraped for model training, while runtime referrals depend on live retrieval systems that operate independently of those training-time restrictions. Opting out of AI training, in other words, does not appear to reduce a publisher's odds of receiving a citation-driven click at the moment a user is searching.

Context for a fast-moving traffic debate

The Bocconi paper arrives amid a broader, ongoing reassessment of how AI-mediated search is redistributing web traffic, a topic PPC Land has tracked closely as figures from multiple analytics vendors have circulated throughout 2025 and 2026. Research from Ahrefs found that AI Overviews correlate with a 58% reduction in click-through rates for top-ranking pages as of December 2025, nearly double the 34.5% decline the same firm measured in April 2025. A separate randomized field experiment, described in coverage of the first causal study of AI Overviews' effect on clicks, found a 39.8% reduction in outbound publisher clicks when the feature was shown to a sample of 1,065 desktop Chrome users, alongside a 34.5% rise in zero-click searches.

On the ChatGPT side specifically, the low single-digit referral rate the Bocconi authors measured is broadly consistent with prior trade reporting. Ahrefs data covering 44,421 websites found ChatGPT capturing just 0.19% of total web traffic share as of August 2025, against Google's 41.9%, even as ChatGPT's month-over-month growth rate outpaced Google's several times over. A more recent, February 2026 comparison found that ChatGPT sent roughly 190 times less traffic than Google despite handling an estimated 12% of Google's query volume. Citation-driven volatility has also proven substantial on a shorter timescale: coverage of a mid-2025 shift found that ChatGPT referral traffic dropped 52% within weeks after OpenAI adjusted its source-weighting algorithm, with Reddit and Wikipedia absorbing much of the redirected citation share.

The category-level erosion the Bocconi researchers documented for academic, reference, and technical destinations echoes patterns publishers have reported independently. Chartbeat data cited in reporting on small publisher traffic lossesfound that outlets receiving between 1,000 and 10,000 daily page views lost 60% of their search referral traffic over two years, while ChatGPT referrals remained under 1% of total publisher page views across the same period - a gap consistent with the Bocconi finding that AI search absorbs many informational queries without generating a comparable volume of compensating clicks elsewhere.

The paper's finding that robots.txt blocking does not reduce runtime referral volume is relevant to a wider argument publishers and news organizations have made about compensation for AI training data. Independent publishers filed a formal antitrust complaint with the European Commission and the UK Competition and Markets Authority in mid-2025 over AI-related traffic declines, and the News Media Alliance has separately argued that AI systems draw on published content while returning comparatively little economic value in return. The Bocconi paper's data point - that domains actively refusing AI crawler access still receive more, not fewer, referral clicks than domains that permit crawling - adds a specific empirical wrinkle to that argument, since it suggests the training-access dispute and the runtime-traffic dispute may not move together in the way some advocacy framing has assumed.

Measurement infrastructure has also evolved considerably since the Bocconi study's data window closed in July 2025. Google Analytics added a dedicated AI Assistant channel in May 2026 specifically to capture sessions referred by ChatGPT, Gemini, and Claude, addressing a longstanding gap in which such visits were often misclassified as direct traffic. That update followed research finding that a majority of AI-influenced traffic never appears as a direct AI referral at all: Similarweb's panel-based clickstream study found that 55.9% of AI-influenced site visits arrived through a subsequent branded search query rather than a direct click from within the AI conversation, a pattern the Bocconi paper's own methodology, built around observing whole conversation sessions rather than isolated click events, was partly designed to capture more completely.

For advertisers and search marketers specifically, the practical implication centers on where query volume and routed attention are concentrated. The Bocconi data indicate that the categories losing the most traditional search traffic under wider ChatGPT access are research, reference, and how-to queries, while transactional and recreational search categories showed smaller changes. That distinction matters for paid search budget allocation, since it suggests informational, top-of-funnel query volume may erode faster than the commercial, bottom-of-funnel queries that carry more direct advertising value. The study's authors describe this as the routing margin through which AI search is reallocating attention, distinct from any claim about overall advertiser return on investment.

The paper's authors caution that their conclusions rest on United States desktop behavior observed through Comscore's household panel and do not extend to mobile usage, international markets, or measures of publisher revenue and long-run content investment. They describe the central open question for future research as whether content producers will continue creating high-quality informational material at the same scale if AI search continues to draw on that material while returning a shrinking share of visits in exchange.

Timeline

  • October 31, 2024: OpenAI opened ChatGPT Search to paid subscribers, the first of three access expansions the study analyzes.
  • December 16, 2024: OpenAI extended ChatGPT Search to all free logged-in users globally.
  • February 5, 2025: OpenAI extended ChatGPT Search access to anonymous, non-logged-in browsers.
  • October 2024 through July 2025: The ten-month Comscore clickstream panel window analyzed in the study.
  • July 2025: Message-level referral-tracking data first reached sufficient coverage to support the study's finer-grained cross-check.
  • July 8, 2026: The paper "Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain" was submitted to arXiv.

Summary

Who: Researchers Qiaoni Shi, Kai Zhu, and Kai Gu of Bocconi University in Milan, studying United States households in a Comscore desktop clickstream panel.

What: A study finding that ChatGPT produces an outbound referral click in only 5.2% of conversation sessions, compared with 31.1% for Google search queries, and that expanding access to ChatGPT Search reduces traditional search engine queries by 9.4% on average, rising to 17.0% after twenty weeks, with the largest declines concentrated in academic, reference, and technical search categories.

When: The underlying clickstream data covers October 2024 through July 2025; the paper was submitted to arXiv on July 8, 2026.

Where: The study is based on United States desktop household browsing behavior recorded through Comscore's panel.

Why: The findings matter to publishers, advertisers, and search marketers because they quantify, for the first time using household-level clickstream data rather than platform-reported estimates, how much of the information-seeking activity moving to ChatGPT is being resolved without generating the referral visits that have historically funded content production, digital advertising, and publisher revenue on the open web.