An AI Overview is not one search. Google breaks your question into several related sub-queries, runs them all against its Search index, pulls passages from the pages that come back, and has a model write one answer grounded in those passages. That fan-out step is why the panel cites pages you never ranked for, and why the same query gives you a different panel two days later.
You have seen these panels for two years now, and you have read the headlines about lost clicks. What follows is the mechanism underneath: what Google documents about its own pipeline, what independent studies measure about citations and traffic, and where the numbers genuinely conflict. Where a figure is a vendor’s tracker rather than a disclosure from Google, it is labeled as such.
How does Google build an AI Overview, step by step?
Google runs a five-stage pipeline: it expands your query into related sub-queries (fan-out), retrieves candidate pages from the core Search index for each one, selects the passages that answer best, has a model synthesize those passages into one response with links, and shows the panel only when its systems judge it adds something to classic results.
The critical detail is that stage two uses the ordinary Search index. Google’s AI features documentation is explicit that pages need to be indexed and eligible to appear in Google Search with a snippet before they can show up as supporting links. There is no separate AI index to get into. What changes is the question being asked of that index.
- Query fan-out. Google generates a set of related queries alongside yours to cover the subtopics your question implies.
- Retrieval. Each of those queries runs against the Search index and returns its own candidate set of pages.
- Passage selection. The system picks specific passages from those pages, not whole documents. A page can be cited for two sentences buried in the middle of it.
- Grounded synthesis. A model writes one answer from the selected passages and carries links back to the sources it leaned on.
- Display decision. Google states that “AI Overviews are only shown when our systems determine that it is additive to classic Search, and as such, often don’t trigger.”
What is query fan-out, and why does it matter so much?
Query fan-out is Google generating extra searches around your question and answering from all of them at once. Google’s AI optimization guide defines it as “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user’s query.”
Google’s own worked example: for “how to fix a lawn that’s full of weeds”, the fan-out queries “might include ‘best herbicides for lawns’, ‘remove weeds without chemicals’, and ‘how to prevent weeds in lawn’.”
Read that example again from a site owner’s point of view. You may hold position 3 for the original query and never be cited, because the passage that made the answer came from a page ranking for “remove weeds without chemicals” — a query you never targeted. Conversely, you can be cited in a panel for a query where you do not rank at all.
The data has moved sharply in that direction. Ahrefs’ July 2025 analysis of 1.9 million citations found roughly 76% of pages cited in AI Overviews also ranked in the top 10 for the same query. Their March 2026 update, covering 863,000 SERPs and 4 million AI Overview URLs, put that figure at 38%. Ahrefs attributes the drop partly to wider fan-out and partly to their own improved citation parsing, so treat it as a strong directional signal rather than a precise year-over-year delta. An earlier Semrush study of 200,000 AI Overviews found the overlap lower still — 20 to 26% of AI Overview links matched the top-10 organic results, and the #1-ranking page showed up in only 46% of desktop panels. Different methods, different snapshots in time, same direction: ranking well helps you get cited, but it stopped being the deciding factor.
Why does the same query return a different panel each time?
Because the panel is regenerated, not stored. Ahrefs tracked 43,000+ keywords with at least 16 recorded AI Overviews each over a month in late 2025 and found a 70% chance the overview changes between one observation and the next, with content persisting about 2.15 days on average.
The citations churn hardest. In that study only 54.5% of URLs carried over between consecutive responses, meaning nearly half the cited sources were new each time. Yet the substance held steady: average cosine similarity between consecutive answers was 0.95, which the authors read as Google “continuously rephrasing a stable, underlying consensus drawn from their sources.”
That split has a direct practical consequence. Checking whether you appear in an AI Overview once tells you almost nothing — you may have caught a run where your URL was in the 45% that rotated out. Any honest measurement uses a fixed prompt set, repeated runs and a score over time, which is the method we lay out in how to measure AI visibility.
How is AI Mode different from AI Overviews?
AI Overviews is a panel that appears above your normal results. AI Mode is a separate conversational surface inside Search, powered by Gemini, where the whole session is the AI answer and you can ask follow-ups. Both use query fan-out, but they draw on sources differently.
Google expanded AI Mode on 7 October 2025 to more than 35 new languages and over 40 new countries and territories, bringing it to more than 200 countries and territories in total. It began as a US, English-only experience, which is why English-language markets have had a longer run of real-world data on it than most others.
| Aspect | Classic results | Featured snippet | AI Overviews | AI Mode |
|---|---|---|---|---|
| Answer format | Ranked list of links | One page quoted verbatim | Synthesized summary above results | Conversational, multi-turn |
| Sources used | 10 separate links | 1 page | Pew: 88% of summaries cited 3+ sources | Broader pool; no independently published count yet |
| Overlap with top 10 | n/a | High, it is a ranked result | Semrush: 20–26% (Sept 2024) · Ahrefs: 38% (Mar 2026) | Not yet independently measured |
| Selection method | Ranking | Best-matching passage | Fan-out plus grounded synthesis | Fan-out across a multi-turn session |
The two AI Overviews figures above come from the Semrush and Ahrefs studies already cited in this article. What is missing is an equally rigorous, source-linked study of AI Mode’s own citation and domain-diversity pattern — it only reached its full 200-plus-country rollout in October 2025, so independent data is still catching up. Treat any precise AI Mode percentage you see quoted elsewhere with the same skepticism this article applies to the prevalence numbers below.
Does an AI Overview actually cost you traffic?
Yes, and the best available measurement is behavioral rather than modeled. The Pew Research Center tracked 900 US adults’ real browsing across 68,879 Google searches in March 2025 and found users clicked a traditional search result on 8% of visits where an AI summary appeared, against 15% where none did.
Two further numbers from that study matter more than the headline. Clicks on links inside the summary itself accounted for just 1% of visits, so being cited is not a traffic substitute for ranking. And users ended their browsing session on 26% of pages with an AI summary against 16% without, which is the answer-satisfied behavior the panel is designed to produce.
Ahrefs approached it from search data instead. Comparing 150,000 keywords with AI Overviews against 150,000 informational keywords without, they reported a 58% lower average clickthrough rate for the top-ranking page when an AI Overview is present, up from 34.5% in their April 2025 study. Their own caveat deserves repeating: CTR fell for non-AI-Overview keywords too, from 0.076 to 0.039 between December 2023 and December 2025, and they use that broader decline as a control. It is a correlation with a stated confounder, not a clean causal estimate.
Google’s counter-position, in its AI features documentation, is that “when people click from search results pages with AI Overviews, these clicks are higher quality (meaning, users are more likely to spend more time on the site).” That is plausible and unfalsifiable from outside — Google has not published the underlying data. Fewer, better-qualified clicks is a real phenomenon; it is also exactly what you would say if volume were down. Watch your own analytics for session depth and conversion rate, not just sessions, and decide for your own site.
Which queries trigger an AI Overview, and which do not?
Panels cluster on informational, comparative and multi-step questions, and thin out on navigational and clearly transactional ones. Google states the panel shows only when its systems judge it additive to classic Search and “often” does not trigger at all. “How do I choose an SEO agency” invites synthesis; “flaminjoe studio contact” does not.
Prevalence numbers are where you should be most skeptical of anything you read. In Pew’s panel, 18% of all Google searches produced an AI summary in March 2025. Commercial trackers publish figures ranging from the teens to well above 50% for the same broad period, because each one uses a different keyword set, country, device mix and detection method. There is no single correct percentage, and any article quoting one without naming the tracker and its sample is repeating folklore.
What does this change for your pages?
Fan-out means you are no longer optimizing one page for one query. You are trying to be the best available passage for a cluster of related sub-questions, any of which Google might generate on its way to an answer. Cover the adjacent questions properly instead of writing one page that gestures at all of them.
Practically, that means headings phrased as the questions people actually ask, a direct 40-to-60-word answer under each one before you expand, and concrete facts — prices, timelines, numbers, named specifics — that a model can lift and attribute without guessing. Vague positioning copy is not merely weak here; it is unquotable, so it is invisible. The full checklist sits in how to optimize your website for AI search, and if the acronyms around this are still fuzzy, GEO vs AEO vs SEO clears them up in a table.
One caution against over-engineering. Google’s optimization guide states plainly that you do not need to create new machine-readable files, AI text files or Markdown to appear in Google Search, that there is “no requirement to break your content into tiny pieces for AI to better understand it,” and that “structured data isn’t required for generative AI search.” That guidance covers Google’s own surfaces. ChatGPT, Perplexity and Claude run their own crawls and weight signals their own way, which is the gap AI visibility vs SEO is about.
If you want to know where you actually stand across all four systems before changing anything, that is what our AI visibility audit measures: 39 criteria across 6 areas, repeatable measurement of what ChatGPT, Gemini, Perplexity and Claude say about you, and an expert reading of the results. You get at least 5 concrete fixes with instructions, or your money back. Prefer something broader and free first? The 3-minute marketing audit scores your whole marketing, or you can talk it through in a free 20-minute consultation.
FAQ
Can I opt out of AI Overviews without losing normal rankings?
Not cleanly. Google names nosnippet, data-nosnippet, max-snippet and noindex as the controls, and describes them as limiting “the information shown from your pages in Search” — Search as a whole, not AI features alone. Google-Extended is a separate control covering training and grounding in some of Google’s other systems, so it does not remove you from AI Overviews, which run on the core Search index. For most businesses, suppressing snippets sitewide to avoid a panel is a bad trade.
Why am I cited for queries I do not rank for? Query fan-out. Google runs additional related searches on the way to an answer, and a page that ranks for one of those sub-queries can supply the passage that gets cited. Ahrefs’ March 2026 analysis found only 38% of AI Overview citations came from pages ranking in the top 10 for the original query.
Does being cited in an AI Overview send traffic? Barely, on its own. Pew found clicks on links inside AI summaries accounted for just 1% of visits in their March 2025 panel. Treat citation as a brand-visibility and trust signal rather than a traffic channel, and keep the underlying ranking, because that is still where the clicks come from.
Should I optimize for AI Mode separately from AI Overviews? No separate content is needed. Google’s documentation states that a page must be “indexed and eligible to be shown in Google Search with a snippet” to be a supporting link in either AI Overviews or AI Mode — both run on the same core Search index and the same fan-out mechanism. AI Mode is the newer, conversational surface, and it only reached its full 200-plus-country rollout in October 2025, so independent studies of its citation pattern are still thin. Until that data exists, covering a topic in depth is the safer bet for both surfaces rather than optimizing for one specifically.