Google's AI Overviews don't work like the ten blue links sitting below them. They run a parallel citation selection process that pulls from the organic results but filters them through a completely different set of priorities. If you've been treating AIO visibility as an extension of your existing rankings, the data suggests you're only partly right.
SE Ranking's large-scale research found that roughly 85% of AI Overview citations link to at least one domain from the top 10 organic results. That sounds reassuring until you realize: 15% of citations pull exclusively from outside the top 10. And even within that 85%, the system isn't just grabbing the highest-ranked page.
It's evaluating which source can best explain the topic.
Retrieval First, Then Selection
The distinction that matters is between retrieval eligibility and citation selection.
Retrieval is stage one. Google identifies candidate pages, heavily weighted toward top-10 organic results. If you don't rank, you're unlikely to enter the candidate pool. This is where traditional SEO still does its job.
Stage two is where things get interesting.
From that candidate set, AI Overviews apply a second filter that evaluates content structure, E-E-A-T signals, and explanatory quality. The pages that get cited are the ones that explain the answer, not necessarily the ones that rank best for the query.
SE Ranking's follow-up study found that featured snippets and AI Overviews match sources 61.79% of the time when they appear together. That 38% mismatch tells you something: AIO selection is a genuinely different evaluation, not just a reshuffle of existing SERP features.
Here's the part that should make you reconsider assumptions about Google's AI systems converging on the same outputs. AI Overviews and Google's AI Mode cite the same URLs only 13.7% of the time, despite reaching 86% semantic similarity in their answers. Two systems, same company, nearly identical conclusions, almost entirely different source selections.
Where the Citations Actually Go
The content type distribution is where the real actionable signal lives. Analysis of roughly 8,000 AI citations found that 82.5% link to deeply nested pages rather than homepages. Blog and editorial content made up about 38% of citations, news around 23%, and expert review platforms like NerdWallet, Consumer Reports, and Investopedia captured roughly 9%. Wikipedia, despite its reputation as the default AI source, accounted for somewhere between 0.6% and 7.8% depending on the platform.
That Wikipedia stat is worth sitting with. ChatGPT leans on Wikipedia at a 27% rate. Google AI Overviews favor blogs at roughly 46% and community content at about 4%.
These are fundamentally different citation philosophies. Google's system favors diverse source types over a single authoritative reference; it's building multi-source validation rather than deferring to one trusted encyclopedia.
YouTube as Validation Hub
One of the most counterintuitive findings: YouTube holds a 29.5% citation share in AI Overviews, according to the SE Ranking data. For health queries, YouTube gets cited more frequently than major medical institutions.
This isn't traditional authority at work. It's what the researchers describe as demonstration value. AI Overviews appear to combine textual sources (Reddit for user opinions, Amazon for product specs) with YouTube for visual verification. The system selects based on the type of proof a query demands, not domain authority alone.
The category breakdowns make this concrete. YouTube captures 78% of citations for electronics queries but just 9% for groceries. How-to queries saw a 651% surge in citations. The system is matching source type to query intent with surprising specificity.
Our read: this pattern looks like a "validation hub" architecture. AI Overviews aren't just finding an answer; they're assembling a portfolio of evidence types. Text for facts, video for demonstration, community discussion for real-world experience. If your content only provides one evidence type, you're competing for a fraction of the citation slots.
YMYL: Where Safety Filters Override Everything
The starkest pattern in the data involves YMYL (Your Money or Your Life) categories.
AI Overview appearance rates vary enormously by topic sensitivity: food and beverage queries trigger AIOs 33.2% of the time, while insurance sits at 8.03% and news/politics at just 5.34%.
For healthcare, legal, and finance queries, the data shows an average of 3.12+ pre-click links per AIO, higher than other categories. More links means more scrutiny, more source diversity, and a higher bar for any single page to clear.
The practical implication: authoritative pages in sensitive categories aren't getting skipped because they lack authority. They're getting skipped because they lack the specific structural signals AI Overviews use for extraction. A hospital's health page with dense paragraph text and no structured data is less extractable than a YouTube walkthrough with clear step-by-step formatting, even if the hospital page is objectively more trustworthy.
Format Signals That Win Across Categories
The structural patterns that correlate with citation selection are remarkably consistent across niches.
Lists, tables, FAQs, and short paragraphs all appear disproportionately in cited sources. Pages with question-based headings get matched to queries more easily. Content organized as self-contained answer blocks (40–60 words that directly address a specific question) fits what retrieval systems are optimized to extract.
The data on page depth reinforces this: 82.5% of citations go to nested pages, not homepages.
Your about page and your homepage aren't getting cited. The blog post with a clear question in the H2 and a structured answer in the first paragraph is.
Long-tail queries also show higher AIO trigger rates. Queries with 8+ words produce AIOs 32.11% of the time versus 12.03% for single-word queries. More specific questions create more opportunity for structured, explanatory content to earn citations.
The Platform Divergence Problem
One more finding worth flagging: each AI platform runs completely different citation logic. ChatGPT concentrates 47.9% of its citations on Wikipedia. Perplexity draws 46.7% from Reddit. Google AI Overviews maintain a balanced mix of professional, social, and editorial sources. Optimizing for one platform's citation preferences may actively work against you on another.
The SE Ranking data on post-rollout changes adds another dimension: maximum pre-click links per AIO decreased 72.86%, dropping from 70 to 19. Google is tightening the citation count, which means fewer slots and higher competition for each one.
For practitioners, the signal is clear: AIO citation isn't a ranking problem. It's a content structure and source-type matching problem. The pages that win are the ones that make it easy for the system to extract a clean, verifiable, well-formatted answer to a specific question. Rank gets you into the candidate pool. Structure and format determine whether you get cited out of it.