Google AI Mode vs. AI Overviews: 88% Accept the Shortlist, 50% Comparison-Shop

A large-scale cursor-tracking study of tens of thousands of searches found that users behave like two entirely different shoppers depending on which of Google’s AI surfaces they land on. In AI Mode, 88 percent of users accept the AI’s shortlist as-is and 74 percent click the top-ranked result before moving on. In AI Overviews, those same users slow down, scroll backward roughly 50 percent of the time, and comparison-shop directly on the results page. The split has real consequences for how publishers and brands think about visibility in AI-driven search.

How do AI Mode and AI Overviews change user behavior?

Google’s two AI surfaces reward opposite behaviors, and the study makes the difference concrete rather than theoretical.

AI Mode is a closed loop. The model assembles a shortlist, the user trusts it, and the click goes to the top-ranked option. There is little browsing and little room to persuade after the fact. With 88 percent of users accepting the shortlist as presented, the entire contest is about being on the list in the first place. That is a pure visibility problem at the model layer: you are either the authority the model pulls in, or you are invisible.

AI Overviews work the opposite way. Researchers describe the pattern as a “Netflix browse.” Users scroll backward roughly 50 percent of the time to reread and validate options before committing, and the comparison happens on the results page itself, before any click-through. That makes AI Overviews a differentiation problem: your value proposition has to survive a side-by-side look, because the evaluation now occurs before the visit, not after it.

The key numbers from the study

• 88 percent of AI Mode users accept the AI shortlist as-is; 74 percent pick the number-one ranked item.
• Roughly 50 percent backward-scroll rate inside AI Overviews, the “Netflix browse” validation pattern.
• A separate Meltwater study found LinkedIn is now the number-two source for all AI search responses, behind only YouTube.
• 100 percent of cited content used bulleted or numbered lists; 92 percent used clear H2/H3 headings.
• 75 percent named specific companies or tools; 67 percent included hard numbers and data.
• 50 percent used comparison frameworks; 33 percent included how-to or decision guides.
• 35 percent of LinkedIn citations came from accounts with fewer than 10,000 followers.

What actually gets cited by AI assistants?

The Meltwater data closes the loop on how content gets pulled into AI answers in the first place, and the finding is striking: citation is driven by formatting, not follower count. The content AI assistants surface is defined by structure, bulleted lists, clear headings, named tools, and hard data, far more than by audience size. That 35 percent of LinkedIn citations came from accounts under 10,000 followers means smaller publishers and individual subject-matter experts have a genuine shot at citation that traditional search rankings rarely offered them.

The practical implication is that structure has become a technical concern, not just a stylistic one. Content written as dense, unstructured prose is far less likely to be cited than the same information broken into lists, headings, named entities, and concrete figures. The blueprint here is unusually specific, and it is measurable.

How is Google reshaping its AI surfaces?

Google is adding new real estate inside these AI surfaces. Preferred sources lets users hand-pick trusted brands to highlight in AI responses. A perspectives carousel surfaces timely articles and discussions. Expanded highly-cited labels reward original reporting and proprietary data. The signal across all three is that original data is becoming a durable advantage rather than a nice-to-have: if you are doing original reporting and data collection, Google is moving to make sure you get credit and visibility for it.

The paid and competitive landscape is shifting in parallel. OpenAI is rolling out pay-per-conversion ads inside ChatGPT, where purchases, bookings, and lead forms complete without leaving the chat. Privacy-first search is gaining ground, with DuckDuckGo reporting roughly a 30 percent jump in app installs following Google I/O. And on the measurement side, Google Ads will begin deleting hourly, daily, and weekly reporting data older than 37 months starting in June 2026, with standard Display campaigns needing migration to Demand Gen by January 2027. None of these are search-ranking changes per se, but together they signal that the entire discovery and measurement stack is being repoured at once.

What should publishers do differently?

The useful conclusion is to stop treating “AI search” as a single target. The two surfaces reward different work. Optimizing for AI Mode means earning authority signals, consistent, structured, expert content the model trusts enough to shortlist, because there is no persuading after the shortlist is set. Optimizing for AI Overviews means making differentiation legible at a glance, because the comparison happens before the click.

Across both, the citation blueprint is the cheapest lever available: format content with lists, real headings, named tools, and hard numbers. It is a concrete, testable checklist rather than a vague aspiration, which is rare in search optimization. The open question for any publisher is simply whether their existing content already follows it, or whether it reads as the kind of unstructured prose these models now routinely pass over.

The bigger picture

The headline is not that AI changed search. It is that AI fragmented search into surfaces that reward opposite behaviors, while the content feeding those surfaces is increasingly structured, data-rich material that smaller players can produce. The advantage over the next year will go to those who format for citation, publish original data worth citing, and understand that being on the shortlist and standing out in a comparison are now two separate problems requiring two separate strategies.

FAQ

What is the main difference between AI Mode and AI Overviews?

AI Mode is a closed loop where 88 percent of users accept the shortlist as-is and 74 percent click the top result. AI Overviews trigger a “Netflix browse” pattern, with users scrolling backward roughly 50 percent of the time to compare options on the results page before clicking.

What kind of content gets cited most often by AI assistants?

According to Meltwater, 100 percent of cited content used bulleted or numbered lists, 92 percent used clear H2/H3 headings, 75 percent named specific companies or tools, and 67 percent included hard numbers and data. Structure and specifics, not follower count, drive citation.

Can smaller publishers get cited in AI search results?

Yes. 35 percent of LinkedIn citations came from accounts with fewer than 10,000 followers, showing that structured, data-rich content from smaller publishers and individual subject-matter experts can be pulled into AI responses.