ChatGPT Answers, But It Rarely Refers: Inside the 45,000-Household Study Rewriting Search Economics
A Bocconi clickstream study finds ChatGPT refers out in 5.2% of sessions vs Google's 31.1% — and cuts traditional search 17% in twenty weeks.
A new Bocconi University study is the first causal, panel-level proof that AI search does not merely divert clicks — it withholds them. Across a 45,386-household clickstream panel, ChatGPT produced an outbound referral in only 5.2% of sessions, against Google's 31.1%. Wider ChatGPT Search access cut traditional search 9.4% immediately and 17% within twenty weeks. The web's traffic-for-content bargain is being rewritten in real time.
The paper — Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain, by Qiaoni Shi, Kai Zhu, and Kai Gu, posted to arXiv on July 8, 2026 and first reported by PPC Land — matters because it moves the conversation past anecdote. Publishers have complained about AI traffic loss for two years. This is the first work to measure the displacement with URL-level Comscore data and a clean quasi-experiment: staggered ChatGPT Search access expansions used as the shock.
Key takeaways
- ChatGPT refers out in 5.2% of sessions; Google, 31.1%. The referral is now the exception, not the rule.
- Wider ChatGPT Search access cut traditional search 9.4% immediately and 17% after twenty weeks — a causal, habituating decline.
- Academic (32.8%) and reference (26.5%) sites lose most: AI is defunding the substrate it depends on.
- Independent Google-side data from Pew Research shows the same direction — the click is eroding across every engine.
The number that reframes everything: 5.2%
Most coverage of AI and search argues about quality of referral traffic. Shi, Zhu, and Gu reframe the argument around existence of referral traffic. In their panel, 74.4% of ChatGPT-using households never generated a single outbound referral across ten months. Google's equivalent zero-referral rate was 9.6%.
That gap is the whole story. Google's model, even in its AI Overviews era, still assumes the click is the product — the search engine's job is to route you somewhere. ChatGPT's model assumes the answer is the product, and the citation is a courtesy. Logged-in message sends referred out just 1.2% of the time in the July 2025 data. The referral is not a scaled-down version of Google's stream. It is a structurally different, and far thinner, thing.
This is what the zero-click answer pattern looks like once it becomes the default rather than the exception. The engine resolves the query in place, and the source that supplied the knowledge gets neither the visit nor, in most cases, a visible name.
Who bleeds most: the sites AI depends on
The category breakdown is the sharpest finding, and it should worry anyone running a high-authority destination:
| Category | Traditional search decline |
|---|---|
| Academic research | 32.8% |
| Reference / knowledge sites | 26.5% |
| Developer / technical resources | 15.1% |
| News / journalism | 13.4% |
The pattern is not random. The steepest losses land on exactly the sources large language models lean on to answer well — academic papers, reference material, technical documentation. AI search is quietly defunding the substrate it is built on. That is not a stable equilibrium, and it is the real policy question buried under the traffic numbers.
For operators, the tactical read is blunt: if your category sits high on that list, the referral channel you optimized for a decade is shrinking by design, not by accident. No amount of on-page tuning restores a click the interface never intends to send.
Google's own data points the same direction
This is not a ChatGPT-only phenomenon; it is the direction of the whole market. A Pew Research Center analysis of real browsing data found Google users clicked a traditional result link 8% of the time when an AI summary was present, versus 15% when it was not — and just 1% clicked a link inside the summary itself. Search Engine Land's read of the same data framed it as a structural erosion of the click, not a temporary dip. Two different engines, two different datasets, one direction: the answer is replacing the referral.
The measurement trap most teams are in
Here is the trap. If your dashboard still equates visibility with sessions, you are measuring a channel that is structurally contracting — and you will read the decline as your own failure rather than a market shift. The study's contribution is to prove the shift is causal: expanded access caused the search decline, deepening over twenty weeks as behavior habituated.
The correct response is not to chase the vanishing click. It is to move up a layer — to measure whether you are the cited source inside the answer, whether the machine names you, and how often. That is the AI-search attribution and measurement gap in one sentence: the value is now created at the moment of citation, but almost no one instruments that moment.
This is the discipline Jaxon Parrott named Machine Relations — the shift from human-mediated search to machine-mediated discovery, where being the answer beats ranking for it. The Bocconi data is, in effect, the empirical case for why the frame is necessary: when 94.8% of ChatGPT sessions end without a click, "rank and get the visit" stops describing the game.
What actually works when referrals collapse
The counterintuitive move is to optimize for the citation you may never get credited for. Earned, third-party authority is what AI engines pull into answers, and it compounds independently of whether the click lands. AuthorityTech, which operates a results-only earned-media model across 50+ tier-one publications, has argued this from the citation side: the durable asset is being the source the machine trusts, not the URL it occasionally links.
Practically, three moves follow from the study:
- Instrument citations, not just clicks. Track how often engines name or quote you, across ChatGPT, Perplexity, Gemini, and AI Overviews — because that is where the 94.8% of "lost" attention actually resolves.
- Invest where the machine sources, not where the human clicked. If academic and reference sites are down 30%+ but still feed the answers, the leverage is in being inside those answers.
- Treat referral traffic as a lagging, shrinking metric. Useful for now, misleading as a north star.
The Bocconi paper will be cited for its headline 17% number. The more important line is the 5.2%. It says the answer economy has already arrived, and it does not run on clicks. The full arXiv preprint lays out the identification strategy for anyone who wants to pressure-test the causal claim.
FAQ
Does this mean SEO is dead?
No — but the payoff is moving. Traditional SEO still captures the 31.1% of Google sessions that refer out. The study shows that share is structurally shrinking as AI search access widens, so the growth is in being the cited source inside AI answers, not only in ranking for the click.
What is the difference between the 9.4% and 17% figures?
9.4% is the immediate drop in traditional search queries after ChatGPT Search access expanded. 17% is the deeper decline measured twenty weeks later, as users habituated to answering in place rather than searching out.
Why do academic and reference sites lose the most traffic?
Because they are exactly what large language models draw on to produce accurate answers. The engine ingests the knowledge and resolves the query in place, so the high-authority source supplies the substance without receiving the visit.
Want to know whether AI engines are citing you or your competitors? Run a free AI visibility audit to see where your brand shows up inside the answers.