ChatGPT is eating informational search. Not commercial.

The Bocconi paper landed yesterday and it deserves more attention than it's getting. Researchers pulled Comscore clickstream data covering the period OpenAI expanded ChatGPT Search access — paid subscribers in October 2024, free users in December, anonymous users by February 2025 — and measured what actually happened to traditional search behaviour as access widened.

The headline number is a 9.4% weekly drop in traditional search queries after broader access, rising to 17% after twenty weeks. That's the part everyone's quoting. It's not the interesting part.

The interesting part is which searches disappeared, and where the click-throughs that remain are going. Because taken together, those two things describe a specific type of web being hollowed out — and it's not the one most SEO conversations are focused on.

Informational search is the category being eaten

The drop isn't evenly distributed. Transactional and recreational searches barely moved. Informational search collapsed. Academic research referrals fell 32.8%. Reference categories dropped 26.5%.

Informational SEO as a discipline is being restructured. Transactional SEO isn't. That distinction hasn't made it into most planning documents yet.

That's a targeted extraction, not a general decline. People still go to Google when they want to buy something, book something, or find a local business. They stop going to Google when they want to understand something. The "let me look this up" query — the one that used to load a SERP full of ten blue links, three ads, and a knowledge panel — is the query being satisfied inside the chat window instead.

If your site's traffic model was built on informational intent — how-to content, explainer posts, glossaries, the kind of top-of-funnel material every content strategy deck from 2018 to 2023 recommended — you're the category being taxed. If your site sells things or takes bookings, you're comparatively insulated for now.

Informational SEO as a discipline is being restructured. Transactional SEO isn't. That distinction hasn't made it into most planning documents yet.

The 5.2% figure is the one that should worry publishers

ChatGPT sends users to external sites in 5.2% of sessions. Google does it in 31.1%. Both numbers are the point.

The 5.2% tells you what everyone already suspected — that AI answers absorb most of the intent they satisfy, and referral traffic from LLMs is a fraction of what search delivered. That's not new information; it's confirmation of what publishers have been reporting for a year.

The 31.1% is more interesting. Google's referral rate is roughly six times ChatGPT's, and Google's referral traffic is heavily concentrated on a small number of dominant destinations — YouTube, Reddit, Wikipedia. ChatGPT's outbound traffic, when it happens, goes somewhere different: reference sites, tools, SaaS platforms, academic and developer resources. Ad-supported sites make up 27.6 percentage points less of the referral mix.

Read that carefully. It's not just that ChatGPT sends less traffic. It's that when it does send traffic, it routes it away from ad-supported publishing and towards subscription, freemium, and non-profit destinations. The ad-supported open web is being disintermediated by ChatGPT in a way that the subscription and tools web isn't.

Which means the business model most affected by this shift isn't SEO. It's ad-supported content publishing that depended on informational search traffic to sell display impressions. That model has been under pressure for a decade. This accelerates the timeline.

The Google response is already in the data

The other piece of yesterday's news — Ginny Marvin's Ads Decoded clarifications on how ads become eligible for AI Search — reads differently in this context.

Google's position is unambiguous: eligibility in AI Overviews and AI Mode requires Broad Match or keywordless targeting through AI Max, Performance Max, Shopping, and Dynamic Search Ads, plus Smart Bidding. Nothing has changed, Marvin said, and that's exactly the point. The AI Search product is being built on top of the ad-buying behaviours Google wants to consolidate around anyway. If you're not running the AI-powered campaign types, your ads are less competitive in the surfaces Google is actively expanding.

Sit that alongside the Bocconi data. Informational search — the queries where AI answers most effectively substitute for clicks — is what's declining. Transactional and commercial intent, where the ad revenue actually lives, is largely intact. Google is defending the surface that matters to its business by making AI Search ads dependent on the campaign types where its automation is strongest.

That's a coherent strategy. Whether it's a good one for advertisers is a separate question. But the pattern is clear: Google is willing to cede informational query volume to ChatGPT if it can hold onto commercial query volume with an AI-native ads product that only functions properly when advertisers hand over targeting control.

What this actually means for the reader

If you run a business that gets traffic from commercial and transactional search, the sky is not falling. Your queries are the ones Google is actively defending, and the Bocconi data suggests users still turn to search engines for those intents. Your priorities remain what they were: technical hygiene, structured data, brand signals, the boring work.

If you run an ad-supported informational site — or your content strategy is anchored in informational top-of-funnel content — the picture is different. Not catastrophic, but genuinely different. The informational web that Google monetised via referrals for two decades is being partially reabsorbed into chat interfaces. Some of that traffic will come back through AI citations, but at a much lower rate and with a very different economic profile.

The strategic implication isn't "abandon informational content." It's that informational content now has to justify itself as a brand-building or authority-building asset, not a traffic-generating one. If a piece exists solely to capture informational search intent and monetise the visit via display, its economics have quietly stopped working. If it exists to build citation share, brand recall, and trust that eventually converts through direct or branded search — the numbers might still add up.

The other implication, which almost nobody is stating plainly: the mid-market SEO agency model built on selling informational content production to service businesses is now selling a product with a shrinking ROI curve. That reckoning is coming whether the industry wants to talk about it or not.

The measurement problem gets worse from here

Here's the awkward bit. Everything above is inference from one paper, using U.S. desktop Comscore data, covering a specific access-expansion window. The authors are careful to hedge — they measure a change in observable traffic allocation, not welfare, not publisher revenue, not long-term content production incentives. Mobile isn't captured. International isn't captured. The paid subscription and API-driven ChatGPT usage isn't visible in clickstream data.

Which is where we always end up. The AI search measurement problem isn't that we lack data — it's that the data we have is partial, lagging, and drawn from surfaces that don't reflect how most people actually use these tools. The Bocconi study is one of the better empirical looks at this shift so far. It's still a snapshot from one country on one device type over one window of time.

If you're making budget decisions from a single dataset, you're making them wrong. If you're refusing to make decisions because no dataset is complete, you're also making them wrong. The honest position is: informational search demand is being partially absorbed by ChatGPT, at a rate that's material, in a pattern that's uneven, and the people most exposed are the ones who spent the last five years optimising for the exact query types that are now being satisfied inside the model. Plan accordingly.

That's the loop. And the industry built a lot of it.