If you’re optimising for position one on the assumption that gets you into Google’s AI Overviews, the data says you’re wrong about four times out of five. Two independent studies — one peer-reviewed, one from a benchmarks firm — arrive at the same finding: the sources cited in AI responses and the pages ranking in traditional organic search are largely different populations.
Two numbers anchor everything in this article: a Jaccard similarity of 0.17 between AI Overview citations and top-10 organic results (arxiv 2604.27790v1), and a 17% AIO-to-top-10 overlap figure from Conductor AEO/GEO Benchmarks. This is an honest evidence synthesis — confidence levels are stated explicitly and you’ll find calibrated uncertainty here, not advocacy. This article is the evidence base for the AI search zero-click crisis — part of a broader series covering the full scope of the shift.
What the Research Actually Shows — And Where the Evidence Is Thin
The research base on AI citation predictors is thin relative to the volume of practitioner claims. Before using any finding to drive strategy, it helps to know what kind of claim you’re actually looking at.
Three sources with disclosed methodologies anchor the evidence hierarchy. The arxiv study (2604.27790v1) analysed 11,500 queries from the ORCAS dataset — real-user queries from Bing search logs — measuring source overlap between SERP results and AI Overview citations, with data collected in December 2025. The Ahrefs study sampled 300,000 keywords using Google Search Console data to measure CTR per keyword position. The Conductor AEO/GEO Benchmarks are large-scale but not independently peer-reviewed.
💡 Rank-Biased Overlap (RBO) is a ranking similarity metric that weights top positions more heavily than Jaccard — useful when citation order matters, not just which sources appear.
What the Ahrefs study actually measured: CTR differential at position one with and without an AI Overview present. The 0.073 → 0.016 CTR collapse is the result. What it does not measure: whether those ranked pages were cited in the AI Overview — which is a different question entirely.
Most other practitioner guidance on “what gets you cited” is pattern-matching on observed outcomes, not causal research. Where that applies in this article, it will be stated.
The Jaccard Finding — Why Your Rankings and Your AI Citations Are Different Lists
Jaccard similarity measures set overlap on a scale from 0 to 1. A score of 0 means two sets share nothing. A score of 1 means they’re identical.
Here’s a concrete example. Google returns 10 URLs in organic search for a query. An AI Overview cites 10 sources for the same query. If Jaccard similarity is 0.17, roughly two sources appear in both lists. The other eight in the AI Overview come from somewhere else entirely.
The arxiv study found average Jaccard similarity of 0.17 for real-user ORCAS queries. The authors state it plainly: “On average, only 18% of the sources returned by either the AIO or traditional SERP will be retrieved by both search engines.” Conductor’s finding from a different methodology lands in the same place — only 17% of AIO-cited URLs come from pages ranking in the traditional top 10.
Ahrefs’ longitudinal analysis of 863,000 keywords adds the trend line: AIO citations from Google’s top-10 pages dropped from 76% in mid-2024 to 38% by early 2026. For AI assistants like ChatGPT, Gemini, and Copilot, only 12% of cited links rank in the top 10. The trend is moving away from overlap, not towards it.
The practical implication: optimising exclusively for traditional SERP ranking addresses approximately 17% of the AI Overview citation pool. The architectural mechanism that explains these citation patterns is covered in detail in the AI Overviews mechanism article.
What Content Characteristics Does the Data Show Predict AI Citation?
Not all citation predictors are equally supported by evidence. Here is the hierarchy with confidence levels based on data quality, not marketing appeal.
Entity clarity — Strongly evidenced (arxiv 2604.27790v1). Generative search engines are less likely to cite sources that lack consistent entity signals — even high-traffic, well-ranked sites. Inconsistencies reduce AI citation accuracy by 30–40%. Content with 15 or more connected named entities shows 4.8x higher selection probability.
Passage-level content structure — Strongly evidenced (multiple GEO studies). AI retrieval systems extract specific passages, not entire pages. Q&A format sections show a +25% citation impact; clear H2/H3 hierarchy shows +40% citation likelihood. The mechanism is straightforward — the system extracts a passage that answers a query, it doesn’t summarise a document.
Schema markup — Moderately evidenced (AIVO, Lumar, arxiv). Pages with three or more schema types are 13% more likely to be cited. Organisation, Product, Person, and FAQ schemas are most consistently associated with AI citation. The evidence is correlational, not causal.
Topical authority over domain authority — Moderately evidenced (ZipTie/Conductor). Domain Authority explains less than 4% of AI citation variance (r²=0.032). Topical authority — depth of coverage within a subject area — shows a correlation of r=0.41. Pages ranking sixth to tenth with strong topical authority are cited 2.3x more often than pages ranking first with weak topical authority.
Google-Extended blocking reduces citations — Strongly evidenced (arxiv). Blocking the Google-Extended crawler statistically reduces AIO citation rate. Playwire data shows publishers who blocked AI crawlers experienced a 23.1% monthly traffic decline in early 2026 with no reduction in AI citations.
First-party data, author markup — Practitioner hypothesis. The Princeton/Georgia Tech GEO study found adding statistics improved citation visibility by 41% on Perplexity. Real effect sizes, but not yet verified across other platforms.
What the data does not support: a clean checklist that guarantees AI citation. Citation is probabilistic and query-dependent. What GEO and AEO practice should look like based on this evidence is covered in the AEO and GEO disciplines article.
The Conversion Data — How Valuable Is an AI Citation, Exactly?
The measurement range here is enormous, and the endpoints reflect genuinely different things.
AIVO (tryaivo.com) reports AI search visitors convert 23x better than traditional organic visitors. Pew Research‘s tracking of 900 US adults across 68,879 searches found users clicked AI Overview links less than 1% of the time. These figures aren’t contradictory — AIVO’s 23x is self-reported attribution from a self-selected, high-intent population; Pew measured observed behaviour across a general population performing ordinary searches. Neither is wrong. They answer different questions.
The Ahrefs data provides the most reliable signal on click volume: position-one CTR collapsed from 0.073 to 0.016 when an AI Overview is present — a 78% reduction based on 300,000 keywords.
For B2B tech businesses, Semrush‘s 10M+ keyword study found AI Overview saturation highest in Computers & Electronics at 17.92%. At that saturation level, whether you’re cited or not has real stakes regardless of per-citation click rate.
The honest summary: the evidence supports “AI citations matter” without supporting “AI citations produce X conversions for my specific business.” How agent search ranking diverges from human search citation is explored in the AI Search for Agents article.
Why Your GA4 Dashboard Is Not Showing You the Full Picture
Attribution decay is the mechanism that makes AI search damage invisible in standard analytics. GA4 strips attribution from visits that arrive via AI-answered queries because no referral click occurs. The RADM (Revenue Attribution Decay Model) framework from DigitalApplied estimates 35–52% of branded-query attribution is lost before the click stage even begins.
The framework identifies three decay stages. Pre-click decay: zero-click SERPs answer the query before any click is possible — the content contributed to a decision, GA4 records nothing. Click-path decay: a prospect asks ChatGPT about vendors, gets an answer naming your product, then types your URL directly — GA4 records a direct visit, the AI citation is invisible. Post-click decay: users who previously googled your brand name now just ask the AI directly.
RADM’s sample worksheet shows an Organic + AI Citations channel with a reported ROAS of 1.2x and a decay-adjusted ROAS of 5.4x — model outputs, not guarantees, but the direction is consistent. Google-Agent adds a further gap: it ignores robots.txt and doesn’t appear in analytics as an identifiable source.
Before you cut content investment on the basis of declining GA4 traffic, the AI search zero-click crisis context explains why that number may be misleading you.
AI Citation Monitoring in Practice — How to Build Your Own Data
Only 16% of brands systematically track AI search performance. Without monitoring you can’t distinguish between two completely different problems: “our content is not being cited” (a content problem) and “we are being cited but can’t measure the pipeline contribution” (an attribution problem). Both require different responses.
The rotating-prompt harness. Identify 20–50 queries relevant to your topic clusters. Run them against Google AI Overviews, ChatGPT, Perplexity, and Gemini weekly — citation patterns can shift materially within a month. Record citation frequency, citation position, and citation share by platform.
Primary tools. Profound tracks citation frequency by prompt and query volume with platform-level breakdowns. Peec AI provides AI visibility tracking across platforms as a category alternative.
Platform prioritisation. Google AI Overviews and AI Mode are the highest-priority citation surfaces by query volume. ChatGPT search is second: 1.2 billion referral clicks in Q4 2025. Perplexity is smaller but cited for high-intent traffic. Citation logic differs across platforms — Google favours brand-managed websites; ChatGPT favours encyclopedic depth; Perplexity is community-weighted — with only 10–15% overlap in citations across platforms. For most B2B businesses, Google AI Overviews is the right place to start.
What This Means for Content Investment Decisions
Ranking optimisation (traditional SEO) and citation optimisation (GEO/AEO) are partially overlapping but fundamentally different disciplines. A content team treating them as the same thing is addressing 17% of the AI citation pool and leaving the rest to factors it isn’t optimising for.
💡 GEO (Generative Engine Optimisation) and AEO (Answer Engine Optimisation) are the emerging disciplines focused on making content retrievable and citable by AI systems — distinct from traditional SEO.
The evidence-supported investments in priority order: entity clarity (Knowledge Graph resolution), passage-level structure (declarative opening sentences, self-contained sections), structured data (Organisation, Product, Person, FAQ schemas), and topical coverage completeness — the r=0.41 correlation is the strongest single quantitative predictor in the data.
The attribution problem is as urgent as the citation problem. Run RADM Step 1 — baseline funnel segmentation — before making any content investment decisions based on GA4 traffic figures alone.
Set up a rotating-prompt harness for 20 target queries. Check your robots.txt for Google-Extended blocking. Implement schema markup on your top-performing content. Then build the citation frequency dataset that lets you test whether changes are working — because the published research cannot answer that for your specific business. That is the evidence base for how to respond to the zero-click shift.
Frequently Asked Questions
Does ranking at position 1 in Google mean my content will appear in AI Overviews?
No. The arxiv ORCAS study found Jaccard similarity of 0.17 between AIO citations and top-10 organic results. Conductor shows only 17% of AIO citations come from URLs in the traditional top 10. Ahrefs confirms the share dropped from 76% in mid-2024 to 38% in early 2026. Position 1 is not a reliable predictor.
What does “Jaccard similarity below 0.2” mean in plain language?
A score of 0.17 means roughly 17% of items are shared between two sets. For any given query, approximately four out of five sources cited in an AI Overview do not appear in the top-10 organic results. The two lists operate on different selection logic.
Is blocking Google-Extended a safe option if I don’t want my content used in AI Overviews?
No. The arxiv study found blocking significantly reduces AIO citation rate even though Google may still access content through other crawlers. Playwire data shows publishers who blocked AI crawlers experienced a 23.1% monthly traffic decline in early 2026 with no reduction in AI citations.
Why are the AIVO 23x conversion figure and the Pew Research 1% CTR figure so different?
They measure different things. AIVO’s 23x is self-reported by visitors who said they arrived from AI search — a self-selected, intent-filtered population. Pew Research measured observed click behaviour across 68,879 ordinary searches by 900 US adults. Different populations, different funnel stages. Neither is wrong.
What is the RADM framework and why does it matter?
RADM (Revenue Attribution Decay Model) is a seven-step framework from DigitalApplied that quantifies how much pipeline credit is lost when AI search answers queries before a trackable click occurs. It identifies pre-click, click-path, and post-click decay stages. Its sample worksheet shows an Organic + AI Citations channel with a reported ROAS of 1.2x and a decay-adjusted ROAS of 5.4x — which explains why content contribution is systematically undercounted in standard dashboards.
Which AI platform should I prioritise for citation visibility — Google, ChatGPT, or Perplexity?
Google AI Overviews and AI Mode are highest-priority by query volume. ChatGPT search is second (1.2 billion referral clicks in Q4 2025). Perplexity is smaller but high-intent. Citation logic differs fundamentally across platforms with only 10–15% overlap. Monitoring all three using Profound or Peec AI is more reliable than guessing.
What schema markup is most strongly associated with AI citation?
Organisation, Product, Person, and FAQ schemas are most consistently named in GEO/AEO practitioner research. Pages with three or more schema types are 13% more likely to be cited. The arxiv entity-resolution finding provides the strongest mechanistic rationale: if the AI can’t identify the entity behind content, citation becomes less probable.
Does E-E-A-T affect AI citation probability?
E-E-A-T appears to function as a binary eligibility filter: 96% of AI Overview citations come from sources with strong E-E-A-T signals. Among eligible pages, E-E-A-T does not appear to be a gradient boost. Its direct effect on AI citation beyond the gatekeeper function is not empirically verified.
How do I know if my content is being cited in AI Overviews right now?
Search Google for your target queries with AI Overviews active and check whether your domain appears in the citation sources. For systematic monitoring, Profound and Peec AI automate this across multiple queries and platforms.
Does Google-Agent show up in my analytics?
No. Google-Agent is a user-triggered fetcher launched in March 2026 that ignores robots.txt and does not appear as an identifiable referral source in GA4 or CRM attribution pipelines. It is one component of the attribution decay problem the RADM framework is designed to surface.
