AI Overviews make users scroll back up twice as often

What happened

Per Search Engine Journal, an analysis of 846,000 Google searches found that users who see an AI Overview scroll back up the results page nearly twice as often as users on a classic results page. The study, published by Eric Van Buskirk, tracked actual session behaviour rather than self-reported survey data.

The headline finding is simple. AI Overviews do not end the search. They restart it. Users read the AI summary, scroll down to scan organic listings, then scroll back up to reconcile what the AI told them with what the blue links promise. That reconciliation loop is new behaviour, and it has direct consequences for how titles and descriptions need to be written.

Search Engine Journal frames this as a writing problem for SEOs. We think it is a brand visibility problem first, and a copywriting problem second.

Why it matters for your brand

The back-scroll is a moment of doubt. The user has just been told something authoritative by Google's AI, and now they are looking at ten organic listings to decide whether to trust it, verify it, or go deeper. Whichever brand owns the listings that resolve that doubt wins the click. Whichever brand sits in position four with a generic title loses, even if it ranks well.

For financial services brands, this is the difference between a retail investor accepting an AI Overview's summary of a fund's risk profile and clicking through to the actual prospectus on your domain. If your listing says "Fund factsheet | [Firm name]" you have given the doubting user no reason to verify. If it says "Q3 2024 holdings, fees, and 5-year performance: official factsheet" you have promised resolution. Asset managers, custodians, and ratings agencies all sit in a category where users specifically distrust AI summaries on regulated information. That distrust is your opening.

For multilateral and policy institutions, the back-scroll behaviour is even more consequential. When a user asks about climate finance flows, vaccine coverage, or sanctions regimes, the AI Overview will pull a number from somewhere. If that number came from a think tank rather than from UNDP, WHO, or the World Bank, the back-scroll is when the user looks for the primary source. Your listing has to signal "this is the original data, not a commentary on it." Multilateral brands have spent years failing to do this in metadata. Most UN agency title tags still read like internal document codes. That has to change.

For major industrial groups, the implication is sharper. B2B procurement researchers do not trust AI Overviews on technical specifications, compliance claims, or sustainability metrics. They scroll back up. If you are Holcim, Siemens, or ArcelorMittal, your organic listing is now the verification layer for whatever the AI just said about your category. Treat the meta description as a fact-check, not a pitch.

For philanthropic and policy institutions, the back-scroll is where attribution gets fought over. Foundations have a chronic problem: their research gets summarised by AI, cited without naming the foundation, and the click goes to a secondary aggregator. Owning the organic listing that says "Original research: [methodology, sample size, date]" is now the recovery mechanism for credit the AI Overview did not give you.

Content strategy needs to absorb this. Title tags and meta descriptions have been treated as a 2010s discipline, neglected in favour of AI-era plays around llms.txt and Reddit citation farming. The 846,000-session data says that classic SERP real estate matters more in the AI era, not less, because the back-scroll has made it the verification layer. Every page that ranks for a query likely to trigger an AI Overview needs its metadata rewritten with one question in mind: what would make a doubting user pick this listing over the AI's answer?

The signal in context

This study sits alongside a growing body of behavioural evidence that AI Overviews are not, as feared, ending the click economy. They are reshaping it. Pew Research found earlier this year that users on AI Overview pages click less, but the clicks they do make are higher intent. Now Van Buskirk's session data adds the mechanism: users hesitate, reread, and verify before clicking. The funnel has not collapsed; it has gained a new step.

For enterprise brands, this changes the priority order. The dominant question of the past eighteen months has been "how do we get cited inside the AI Overview." That matters, but it is no longer sufficient. The parallel question is "when the user back-scrolls after reading the AI Overview, does our listing resolve their doubt." Brands that solve both win twice in the same session. Brands that solve only the citation problem get summarised and abandoned. Brands that solve neither become a footnote in someone else's AI answer.