The first hard number on AI search erosion: what the Bocconi study measures, and why the curve matters more than the percentage
Why this study is different
Since ChatGPT Search gained traction, the debate over its impact on organic traffic has fed mainly on platform-internal correlations, statements from interested parties, and traffic analyses from individual publishers. What was missing was a study with causal design, independent clickstream data, and a long enough time horizon to distinguish adoption effect from structural change.
The paper by Qiaoni Shi, Kai Zhu, and Kai Gu from Bocconi fills that gap with Comscore data: desktop browsing clickstream in the United States between October 2024 and July 2025, with access to ChatGPT Search as the treatment variable. The design lets them compare search behavior before and after exposure, controlling for the adoption curve. The central result, a 9.4% reduction in weekly traditional search queries with an effect that rises to 17% at 20 weeks, is not an observational correlation: it is an estimate with quasi-experimental design built on real behavioral data.
Coverage of the study was picked up by Search Engine Journal on July 13 through Matt G. Southern, but the primary source is the preprint on arXiv, which allows direct examination of the methodology.
The finding that most changes the urgency
The search-reduction numbers are significant, but what really changes the strategic calculation is the shape of the curve: the effect does not stabilize, it deepens. That has a direct implication for any team that has observed moderate organic traffic drops and attributed them to seasonal noise or a Google algorithm adjustment. If part of that audience migrated to AI chat searches and the behavior of those users reinforces over time, the trend will not reverse on its own.
The second finding, that ChatGPT produces outbound clicks in just 5.2% of conversation sessions, complements the first precisely. It is not just that people search less on Google: it is that when they search on ChatGPT, the answer stays in the conversation. The web model that worked for 25 years assumed that the search engine was an intermediary that sent traffic. ChatGPT is designed to answer, not to refer. Shi, Zhu, and Gu call this "answering without referring," and it is the central thesis of the paper: the implicit economic deal of the web, attention in exchange for traffic, is being rewritten at the information-access layer.
In our note on the rise of Answer Engine Marketing we already laid out the underlying mechanics, but at that point the argument rested on product logic and early signals. This study provides the first external number, with clear methodology, that allows the conversation to move beyond intuition.
For whom does this change something this week
This note is worth reading twice for operations with high dependence on organic traffic as their primary acquisition channel: ecommerce with a long SEO history, regional SaaS with a funnel that starts in informational search, fintech with educational content as a lead-generation engine. In those profiles, the risk is not abstract. If 9% of the queries that used to reach Google are no longer generating impressions, part of that audience never reaches the site, no matter how much the ranking improves.
For operations where paid media is already the primary acquisition channel and organic contributes less than 20% of volume, the immediate impact is smaller, though the long-term implication applies equally: if demand migrates and there is no presence in AI answers, SEM will have to compensate with higher spend to capture the same commercial intent that used to arrive on its own. Acquisition cost rises structurally.
For teams with audiences specifically in LATAM, there is a reason for caution that we develop further below: direct extrapolation of this study has geographic limits.
What the study does not answer, and why that matters before drawing conclusions
Three limitations are relevant for reading this data without over-fitting.
The first is geographic: the study uses desktop clickstream in the United States. ChatGPT Search has different adoption rates in LATAM, where access to paid ChatGPT plans is less widespread and mobile search habits, which dominate in the region, may differ. Extrapolating the 9.4% or the 17% to an operation in Argentina, Mexico, or Colombia is an assumption, not a data point. That does not invalidate the direction of the effect, but it does qualify the magnitude.
The second is disaggregation: the paper reports the average effect on traditional searches in general, not on intent segments or industry categories. It is reasonable to assume the impact is asymmetric: informational and comparison searches are more substitutable by AI chat than high-urgency transactional searches. Knowing how much each segment weighs in your own organic base is the work each operation has to do with its own data.
The third is causality: the measured reduction applies to users with broad access to ChatGPT Search, not to the general population. As adoption becomes more widespread, the aggregate effect may be different, either larger or smaller, depending on how new users integrate the tool versus how the study's early adopters did.
These limitations do not nullify the central argument, but they do define the perimeter of what the data proves and what remains inference.
What is worth reviewing in your operation before deciding whether this applies to you
Without claiming to know what any particular team is running, there is a set of analyses worth doing to find out whether this phenomenon is already affecting a specific operation, before changing spend or strategy.
The first step is to disaggregate organic traffic by intent type. If Google Search Console shows impression drops on informational queries (guides, comparisons, "how" or "what is" questions) that are not explained by ranking changes, that can be a signal of migration to AI search, not a loss of position. It is worth separating that segment from transactional organic to see whether the pattern is different.
The second useful analysis is to review traffic from AI search sources directly. In our note on GA4's Source Group and the AI Assistant channel we documented how GA4 now groups ChatGPT, Perplexity, and other AI search sources as a separate channel. Having that reading available lets you measure how much traffic arrives from those sources, even if it is a fraction of total organic.
The third exercise is to assess the potential for AI answer presence in your own categories. That means running representative queries from your ideal customer profile in ChatGPT, Perplexity, and Gemini, and auditing whether your brand or domain appears cited, in what position, and with what framing. In our note on how to measure ads in ChatGPT we mention the Paid Share of Voice framework that Profound began measuring for AI search, which can serve as a starting point for understanding current exposure.
What we do not recommend is moving SEM budget into AI search experiments based solely on the Bocconi study without first understanding how much of your own organic has already migrated. The prior analysis defines whether the problem is urgent or something to monitor this quarter.
Developed with AI. Reviewed by humans. If any data ages poorly, blame the ecosystem, not us.