Your Google Search Console shows stable rankings. Your Google Analytics shows a traffic cliff. The answer is architectural — and it’s not going to fix itself.
Google’s AI Overviews (AIOs), powered by Gemini, now appear for a growing share of queries and answer the user’s question before they have any reason to click through. Position-1 click-through rate has collapsed from 7.3% to 1.6% when an AI Overview is present — a 78% drop in two years (Ahrefs, December 2025). Across 300,000+ keywords, organic CTR falls 61% on average whenever an AIO appears.
This article explains the architectural mechanism behind those numbers — written for a technical audience that wants the why, not just the what. We’ll cover query fan-out, Gemini’s citation selection logic, the Jaccard similarity finding, and why the featured snippet playbook no longer applies. All of this connects to the broader AI search zero-click crisis.
What Happens Inside Google When You Search — The Query Fan-Out Architecture
When you submit a search, Google’s AI system doesn’t return a single ranked list. It decomposes your query.
The process — documented in three Google patents (US20240289407A1, US12158907B1, US11663201B2) — works like this. A single user query is broken into 8–12 parallel sub-queries. These fire simultaneously across multiple retrieval surfaces: the Web index, Knowledge Graph, Shopping Graph, Maps, and YouTube. Each sub-query generates its own results. Passages are then extracted via semantic chunking — documents broken into candidate answer fragments, each evaluated against the relevant sub-query. Gemini synthesises the retrieved passages into a single AI Overview, selecting which passages to cite based on relevance scoring, not traditional ranking order.
💡 Retrieval-Augmented Generation (RAG) is an architecture where a language model retrieves relevant content from external sources before generating its response — the model reads before it writes, which is why passage structure in your content directly affects whether it gets cited.
Two details are worth paying attention to here. First, the fan-out sub-queries are not stable — only 27% are consistent across repeated searches of the same head query (Surfer SEO). Second, this isn’t unique to AI Overviews: AI Mode uses the same mechanism with far more extensive fan-out.
Think of it as a research team sent out to answer a question. They split up, each searches a different related angle, and they bring back the most useful passages. The final answer draws from all of those angles — not from whoever ranked first for the question as originally typed. That decoupling is the mechanism that breaks the old ranking-to-traffic chain.
For the zero-click statistics that quantify this impact at scale, see Article 1.
Why Position 1 No Longer Guarantees Traffic When an AI Overview Is Present
When an AI Overview is present, position-1 CTR falls from 7.3% to 1.6% — a 78% collapse. And the drop isn’t confined to position 1. Ahrefs’ data shows position 2 down 50.8%, position 5 down 32.6%, position 10 down 19.4%. The effect cascades across the whole SERP.
The mechanism is straightforward. The AIO satisfies query intent on-page. The user gets a synthesised answer with cited sources visible inline. There’s no reason to click. Your page keeps its ranking impression in Search Console — but the click never happens.
This is structurally different from the featured snippet era. Featured snippets gave users one extractable answer and they often clicked through for more context. AIOs deliver complete synthesised answers. Your organic result shifts from primary destination to fallback content for users who want to go substantially deeper.
AI Overviews now appear for approximately 13–25% of all queries (Q1 2026). For pure informational queries, saturation is near-total. And the effect is spreading: Semrush shows transactional AI Overview presence grew from 1.98% to 13.94% between 2024 and 2025 — funnel-stage traffic is now exposed.
How Gemini Selects Sources for Citation — and Why Most Top-10 Results Are Ignored
Citation selection is not a ranking operation.
Gemini doesn’t look at the top organic result and cite it. It evaluates passages retrieved from fan-out sub-query SERPs — which may not include the head-query SERP at all — and selects based on passage-level relevance scoring. A URL that ranks number one for the head query may not appear in any of the 8–12 sub-query result sets. It’s simply not in the retrieval pool.
Ahrefs’ February 2026 analysis of 863,000 keyword SERPs confirms this: only 37.9% of URLs cited in AI Overviews appear in the top-10 organic results for the same query. That was 76% in July 2025. The overlap has nearly halved in seven months.
The Knowledge Graph functions as a prerequisite filter. It’s Google’s structured entity database — recording and resolving canonical identities for people, organisations, places, and concepts. Pages referencing entities that are verified and resolved in the Knowledge Graph are more likely to pass the initial retrieval threshold. Pages referencing unresolved entities get filtered out earlier.
Two citation patterns confirm that the logic is decoupled from traditional ranking. YouTube accounts for 18.2% of AI Overview citations among pages not in the top 100 organic results — because video is retrieved via a separate fan-out surface. Reddit accounts for approximately 21% of citations overall, despite lacking traditional domain authority. Both point to a system drawing its citation pool from a far wider retrieval surface than the original query’s top-10.
For what the citation data actually shows about surviving AI Overviews, see Article 7.
The Jaccard Similarity Finding: What Low Overlap Means for Your Content Strategy
Jaccard similarity measures overlap between two sets: intersection divided by union, from 0 (no shared elements) to 1 (identical sets).
Research published in arxiv 2604.27790v1 measured the Jaccard similarity between traditional SERP top-10 rankings and the URLs cited in AI Overviews for the same queries. The average score is below 0.2. In plain terms: if your page ranks in the top 10, there’s an 80%+ probability it won’t appear as a citation in the AI Overview for that same query.
Multiple independent datasets back this up. Ahrefs (863K SERPs) shows 37.9% of AIO citations coming from top-10 results — down from 76% in July 2025. Surfer SEO (173,902 URLs) found that 68% of AI-cited pages were not in the top-10 organic results for the same query.
This isn’t a calibration issue — it’s an accelerating divergence. AI Overviews are relying less on direct search results and more on sources turning up in fan-out query SERPs. The direct implication: optimising for traditional SERP ranking is necessary but no longer sufficient for AI search visibility. The two optimisation targets are structurally diverging, and the gap is widening fast.
Passage-Level Retrievability — Why Content Structure Now Matters More Than Page Rank
The fan-out architecture retrieves at the passage level, not the page level.
Semantic chunking breaks a document into candidate answer fragments. Each is evaluated independently against a specific sub-query. The system isn’t asking “is this a good page?” — it’s asking “does this passage directly answer this specific sub-question?”
A page at position 1 may contain no passage that addresses any of the 8–12 fan-out sub-queries. A page at position 40 may contain several. Wellows research (December 2025) provides the operational data: the optimal passage length for AI Overview extraction is 134–167 words. Cosine similarity above 0.88 between a passage and its target sub-query produces 7.3x higher citation rates.
💡 Cosine similarity measures how closely two pieces of text are aligned in meaning when converted into mathematical vector representations — a score of 0.88 is very high, meaning the passage and sub-query are semantically very close.
Surfer SEO’s analysis of 173,902 URLs found a Spearman correlation of 0.77 between topical coverage — how many fan-out sub-query variants a page addresses — and AIO citation probability. Fan-out coverage, not head-query ranking authority, is the structural predictor.
This rewrites the old content model entirely. Old model: one page, one target query, position zero. New model: one page, multiple structured passages, multiple fan-out sub-queries. Each H2 section should function as a self-contained answer — a declarative opening sentence, a specific factual claim, a data reference — completable in 134–167 words without requiring context from surrounding paragraphs.
The optimisation disciplines that operationalise this are the subject of GEO and AEO — the optimisation disciplines this mechanism created.
AI Overviews vs. AI Mode — Two Points on the Same Architectural Trajectory
AI Overviews and AI Mode are not two different products. They’re two deployments of the same architecture at different intensity levels. Both use query fan-out, RAG passage extraction, and Gemini synthesis. AI Overviews run 8–12 sub-queries and produce a zero-click rate of approximately 83%. AI Mode issues potentially hundreds of sub-queries and produces a 93% zero-click rate (Semrush).
The distinction that matters most for content strategy: AI Mode and AI Overviews cite the same URL only 14% of the time for similar answers (Digital Applied). Despite sharing the same underlying mechanism, they draw from largely separate retrieval pools. A content strategy targeting one cannot assume it’s targeting the other.
AI Mode is currently primarily a US feature; AI Overviews are globally deployed. If you’re outside the US, the optimisation work required for AI Overviews — passage-level structure, entity resolution, topical coverage — is exactly the preparation you need for AI Mode when it arrives.
What Has Replaced Traditional Ranking — The New Citation Logic
The traditional search value exchange was predictable: produce quality content, earn rankings, receive traffic. That chain is architecturally broken — not cyclically depressed. The ranking system still functions. But ranking no longer guarantees the click.
The new citation logic operates on three layers.
Entity Verification: Is the brand or source a resolved entity in the Knowledge Graph? Pages referencing verified, canonical entities pass the initial retrieval threshold more reliably.
Passage-Level Relevance: Does the content contain passages that directly address fan-out sub-queries at the right semantic distance? Cosine similarity above 0.88, passage length 134–167 words, declarative structure.
Topical Coverage: Does the content address enough sub-query variants to appear across the fan-out retrieval pool? The Spearman correlation of 0.77 between topical coverage and citation probability makes this the strongest structural predictor.
These three layers replace “rank well → get traffic” as the working mental model. AIO citations are probabilistic, not deterministic — only 27% of fan-out sub-queries are stable across repeated searches. So AI citation share — how often a brand appears across many query executions — is the correct measurement metric. Tools worth knowing about: Ahrefs Brand Radar, Semrush AI Visibility Toolkit, Profound, Peec AI, and Otterly.
Traditional SEO remains a baseline requirement. For the 75–87% of queries that don’t generate an AI Overview, organic ranking remains the primary traffic driver. The practical approach: maintain SEO as a floor and layer passage-level structuring and entity building — GEO and AEO — as the differentiated layer that determines AI citation outcomes.
This article has explained why the old model is broken. GEO and AEO — the optimisation disciplines this mechanism created covers the optimisation disciplines that replace it. Article 7 examines what the citation data actually shows about surviving AI Overviews. For the full scope of the zero-click shift, the pillar article covers the complete landscape.
Frequently Asked Questions
Why did my organic traffic drop even though my Google rankings haven’t changed?
When an AI Overview is present, position-1 CTR falls from 7.3% to 1.6% without any ranking change. The AIO satisfies user intent on-page — the user reads the synthesised answer and doesn’t click through. Your page still registers an impression in Search Console; the click simply doesn’t happen. Rankings and traffic are now decoupled.
What is query fan-out in plain terms?
Instead of returning one ranked list, Google’s AI decomposes your query into 8–12 related sub-questions, searches each in parallel across the web index, Knowledge Graph, YouTube, and other surfaces, then synthesises the most relevant passages into one AI Overview answer. The specific wording of your original query matters less — the AI rewrites it before retrieval begins.
Is being cited in an AI Overview the same as ranking in the top 10?
No. Only 37.9% of AI Overview citations come from top-10 organic results for the same query (Ahrefs, February 2026) — down from 76% in July 2025. Citations are drawn from fan-out sub-query SERPs, not the head-query SERP. The two lists are largely non-overlapping, and increasingly so.
What does “passage-level retrievability” mean in practice?
AI search systems break your page into passage-sized chunks and evaluate each against a specific sub-query. Each H2 should open with a direct declarative sentence, followed by specific data, completable in under 170 words without requiring surrounding context. Cosine similarity above 0.88 between a passage and its target sub-query produces 7.3x higher citation rates (Wellows, December 2025).
Is traditional SEO still worth doing if AI Overviews are taking the clicks?
Yes — it’s a baseline requirement, not a sufficient strategy on its own. For the 75–87% of queries that don’t generate an AI Overview, organic ranking remains the primary traffic driver, and organic signals (E-E-A-T, topical authority, site health) remain relevant inputs to citation selection. The practical approach: maintain SEO as a floor and layer passage-level content structuring and entity building (GEO/AEO) as the differentiated layer that determines AI citation outcomes.
