Wider access to ChatGPT Search was associated with a 9% drop in weekly traditional search queries, rising to a 17% decline after 20 weeks, according to researchers at Bocconi University who analysed Comscore clickstream data from U.S. desktop browsing. Search Engine Journal reported the findings on 13 July. The figure is an association measured in one market on one device type. It is not proof that ChatGPT caused the decline.

What the Bocconi researchers measured

The data comes from Comscore's clickstream panel, which records what people actually do in a desktop browser rather than what they say they do. The researchers compared traditional search activity against wider ChatGPT Search access and found weekly query volume 9% lower. Stretch the window to 20 weeks and the gap widens to 17%.

The compounding is the part worth sitting with. This was not a one-off dip that levelled out. Over the period measured, the gap kept growing.

What the study cannot tell you is why. People with ChatGPT Search access ran fewer traditional searches. Whether the assistant pulled them away, or whether the kind of person who adopts new tools early was always going to search differently, is not something clickstream data settles. Association, not causation. The distinction survives contact with a slide deck only if someone keeps saying it out loud.

ChatGPT sends far fewer people to websites than Google

This is the number that should change how you forecast. ChatGPT directed users to an outside website in 5.2% of sessions. For Google searches the figure was 31.1%, roughly six times higher.

ChatGPT is getting more generous with links. Its monthly referral rate grew from about 2.5% to nearly 6.5% across the measurement period, which is a steep climb as a rate of change. It is still a small number in absolute terms, and it starts from a base that a search engine built around handing you ten links never had to climb from.

So the trade is not one visit lost for one visit gained somewhere else. When a query moves from Google to an assistant, the odds of anyone landing on a website at the end of it fall by a lot.

Academic and reference sites took the sharpest hit

Referrals to academic research fell 32.8%. Reference citations dropped 26.5%. Those are the categories where a competent summary usually is the whole job. Someone checking a definition or a study's headline finding has little reason to click through once a chatbot has handed them a clean answer.

Use that as a lens on your own pages. Content whose value is "here is the fact you asked for" competes head-on with the summary. Content that needs the site itself to deliver anything, like a booking, a quote, a calculator, or a comparison the reader has to work through, sits in a different position.

The caveat the researchers put on their own numbers

They were explicit: they measured a change in observable traffic allocation. Not consumer surplus. Not publisher revenue. Not long-run content production.

That matters once the 9% starts getting quoted in meetings. Traffic moving is what the study found. Whether revenue follows it, whether publishers cut output, whether users end up better or worse served, all of that sits outside what was tested. Anyone using this research to predict a business outcome is adding assumptions the authors declined to make.

What this means for Thai marketers

Start with what the research does not cover: Thailand. The panel is U.S. desktop browsing. Thai search behaviour leans heavily mobile, and nothing here measured it. The 9% is a direction, not a number to paste into a Thai forecast.

The direction still has planning value:

Organic forecasts that assume search volume only ever goes up need a second scenario. Flat or falling query volume while rankings hold steady is now a case worth modelling.
Impressions and clicks can move for reasons that have nothing to do with your site. Diagnosing a traffic dip in 2026 means checking whether the query pool shrank before rewriting the page.
If assistants take a real share of discovery, being cited inside the answer is worth more than sitting next to it. That is the practical case for generative engine optimisation: structuring content so an assistant can quote it accurately and attribute it back.
The referral rate is rising. 2.5% to nearly 6.5% is small money and a real trend. Tracking assistant referrals as their own channel now beats discovering them in two years.

The honest read is that a shrinking pool of traditional queries is not a shrinking pool of demand. The demand is turning up somewhere harder to measure. Optimising for how ChatGPT retrieves and cites sources is the closest thing to a lever on the part you cannot see in Search Console.

Common questions

Does this mean Thai search traffic will fall 9%?

No. The study covers U.S. desktop browsing. There is no Thai figure in it, and inventing one from this data would be guessing.

Does the research prove ChatGPT caused the drop?

No. It found an association between wider ChatGPT Search access and lower traditional query volume. The researchers made no causal claim.

Should we cut SEO investment because of this?

The study does not support that. It found clicks redistributing, and it explicitly did not measure revenue. The better response is to know which of your pages depend on informational clicks and which do not.

Where to go from here

The useful thing to take from Bocconi's work is not the 9%. It is 5.2% against 31.1%. Assistants answer more and refer less, and that gap is the structural change, whatever the exact percentage turns out to be in any given market. If your organic forecast assumes 2027 looks like 2024, that is the assumption to stress test first. Talk to our team about where AI search fits your channel mix.