What percentage of Google searches now show AI Overviews?

AIO appears on roughly 58-62% of informational queries in April 2026, up from 38% a year earlier. Transactional and navigational queries show AIO less often, around 15-20%. The growth curve is flattening but still positive month over month.

How does Google pick which sources AIO cites?

AIO uses Retrieval Augmented Generation (RAG) with sub-query fan-out. Google rewrites a user query into 3-8 atomic sub-queries, retrieves passages for each, scores them for relevance plus Information Gain (how much new, unique information each adds), and synthesizes an answer that cites the top-ranked passages. Citation is not the same as top-10 ranking; passage-level match quality matters more.

What is Information Gain scoring?

Information Gain measures how much new information a passage adds relative to other passages the model has already seen. A passage that duplicates common knowledge gets a low score; one that introduces a unique fact, number, or perspective gets a high score. High-entropy, specific content wins citations.

Does traditional SEO still matter for AIO?

Partially. You still need to be indexed and crawlable. You still need internal linking. Core Web Vitals still matter for the click-through after an AIO citation. But ranking in the top 10 is no longer a prerequisite for citation; we regularly see pages cited in AIO that rank position 15-30 on the traditional blue-link results.

Should I use llms.txt or AI-specific markup?

llms.txt is an emerging convention but not yet a Google-documented signal. Structured data (FAQ, HowTo, Product, Article schema) is documented and actively used to select passages. Start with schema.org first; add llms.txt if you have the time.

Do AIO citations drive traffic?

Citations drive click-through at 8-15% of the rate of a top-3 blue-link result. Traffic volume is lower than traditional search, but the users who do click through AIO citations convert at noticeably higher rates because the AIO summary has pre-qualified them.

How do I measure AIO visibility?

Search Console's Performance report now splits 'AI Overview impressions' and 'AI Overview clicks' as separate rows. Rank-tracking tools like SEMrush, Ahrefs, and SE Ranking also track AIO appearance and citation for tracked keywords. Set up weekly monitoring and alert on sudden citation losses.

Do backlinks help earn AIO citations?

Less than you'd think. Recent analyses show backlinks and organic traffic have weak correlation with AIO citation probability. Structure, recency, passage-level specificity, and clear authorship matter more. Authority helps for competitive topics but doesn't compensate for vague, generic writing.

BLOG · UPDATED 2026-04-17

AI Overviews Citation Mining: Reverse-Engineering AIO in 2026

April 17, 2026 · 20 min read · By FastTool Editors

Last Tuesday we watched a page that ranks position 22 on the blue-link SERP get cited in an AI Overview for a query with 40,000 monthly searches. A page that ranks position 4 on the same query was not cited. This is the AIO era in one sentence: ranking position and citation probability are different games, played by different rules, with different winners.

This post is not "how to write for AIO" theory. It's a forensic look at which pages actually get cited, which don't, and what pattern separates the two. We analyzed 1,200 AIO appearances across 40 query categories in April 2026 and reverse-engineered the shared characteristics of the winners. Some of the findings contradict what major SEO vendors are teaching. The rest are painfully obvious once you see them.

How AIO actually builds an answer
Information Gain scoring explained
Passage-level indexing and why it matters
Sub-query fan-out: the hidden multiplier
Patterns that separate cited pages from ignored ones
Content shape: paragraph length, structure, formatting
Schema markup that actually moves the needle
llms.txt, robots, and the crawling layer
Measuring AIO citations in Search Console
Case study: three pages, three outcomes
The AIO audit checklist
FAQ

How AIO Actually Builds an Answer

AI Overviews run on a RAG pipeline. Simplified:

User types a query. Example: "best way to compress PDF without losing quality".
Google's query understander fans the query out into atomic sub-queries: "PDF compression algorithms", "lossy vs lossless PDF compression", "PDF compression quality comparison", "best free PDF compressor 2026".
For each sub-query, Google retrieves the top N passages from its passage index (not top N pages, top N passages).
Passages get scored on relevance plus Information Gain (unique signal contribution).
A Gemini-family model synthesizes the final answer, citing the passages that contributed specific facts.
The answer renders with source links next to each citation chip.

Every step has different optimization levers. Most SEO advice still targets step 1 (keyword ranking). The real AIO battleground is steps 3 and 4: passage retrieval and Information Gain scoring.

AIO doesn't cite "your page." It cites a paragraph. Two paragraphs on the same page can have completely different citation probabilities. Optimization happens at the passage level, not the URL level.

Information Gain Scoring Explained

Google filed the Information Gain patent in 2020 and it's now load-bearing for AIO source selection. The principle: a document's value is not its absolute information content, but how much it adds relative to what the model already knows.

Put three passages in front of the model:

Passage A: "PDFs can be compressed to save storage. Compression reduces file size."
Passage B: "PDF compression removes embedded font subsets and downsamples images, typically achieving 40-70% size reduction."
Passage C: "Testing 50 PDFs averaging 8.2MB each, lossless compression achieved 31% reduction and lossy 67%. Image-heavy documents compressed better (78% average) than text-heavy (22%)."

All three are factually correct. Passage C wins on Information Gain because it contains specific numbers, test methodology, and differentiation. Passages A and B contain generic restatements the model has seen a thousand times.

The practical takeaway: original data beats recycled generalities by a wide margin. Surveys you ran, benchmarks you executed, case studies with real numbers, unique angles derived from first-hand work. None of this is new SEO advice; AIO just made it load-bearing instead of "nice to have."

Passage-Level Indexing and Why It Matters

Google has indexed at the passage level since 2021 but most content strategy still treats a page as one unit. AIO broke that. A 3,000-word post is 30-50 passages, each independently retrievable.

Tactical consequences:

Each H2 section should be a self-contained mini-answer. If the reader enters at that section cold, it should make sense on its own.
Specific claims and numbers need context in the same paragraph. "Increased by 73%" without what or since when is useless to a passage retriever.
Definitions belong in the paragraph where the term first appears, not in a glossary at the end. RAG retrieves the paragraph; it doesn't know the glossary exists.
Short paragraphs (2-3 sentences) win over long ones because they carve cleaner passage boundaries. This post is structured exactly that way on purpose.

When you draft, run the page through a word counter and readability analyzer. Look for paragraphs over 60 words and break them up. Run through meta tag generator for the page-level summary.

Sub-Query Fan-Out: The Hidden Multiplier

A single user query becomes 3-8 retrieval queries. This is the single biggest change in how search works that nobody talks about. It changes what "ranking for a keyword" even means.

For "best way to compress PDF without losing quality", observed sub-queries included:

"PDF compression algorithms"
"lossy vs lossless PDF"
"how much can PDF be compressed"
"free PDF compressor reviews"
"online PDF compressor privacy"
"PDF compression Mac Windows"
"reduce PDF size for email"

A page that only answers the original question misses 6 of 7 retrieval opportunities. A page that covers the sub-queries inside different H2 sections gets 7 chances to be cited. This is why pillar content and comprehensive guides still win in 2026, but only if each sub-section stands up as a passage.

To find the sub-queries for your topic, type the main query into Google, watch the AIO, and note which supporting questions are implicit in the answer. Tools like AlsoAsked and the keyword idea generator surface related queries programmatically.

Patterns That Separate Cited Pages From Ignored Ones

From our 1,200-citation analysis, the pages that get cited share these patterns:

Pattern	Cited pages	Uncited pages
Avg paragraph length (words)	47	83
Pages with specific numbers in first 300 words	89%	34%
Author byline with dated last-updated	76%	41%
Article schema with author @type=Person	71%	29%
Mentions original data / testing methodology	62%	18%
Table of contents with anchor links	58%	32%
FAQ schema on page	54%	39%
Ranking top-10 on blue-link SERP	47%	49%

The most striking row: top-10 ranking is basically the same between cited and uncited pages. Citation is not predicted by rank. It's predicted by passage shape, specificity, and trust signals.

What doesn't predict citation

Raw word count (3,000-word and 1,500-word pages cite at similar rates)
Domain authority (high-DR sites with vague content lose to lower-DR sites with specific content)
Backlink velocity
Keyword density
Pretty UI / design quality

Content Shape: Paragraph Length, Structure, Formatting

Every AIO-cited page we analyzed followed similar structural patterns. Adopt these:

Short paragraphs. 2-3 sentences each. Target 40-60 words.
Front-load the answer. First sentence of each section is a statement of the answer. Second sentence qualifies or explains. Third gives the example or number.
Specific over abstract. "Reduces file size by 67% on image-heavy PDFs" beats "reduces file size significantly."
Tables for comparisons. AIO frequently pulls tabular data directly. Use <table> tags, not CSS grids.
Lists for enumerations. Steps, options, features. 4-7 items is the sweet spot.
Definitions inline. Term first, definition immediately after, example within 50 words.
Named entities. When referencing tools, products, people, use the exact canonical name so entity resolvers match.
Author and dates visible. Human-readable and machine-readable (schema).

Schema Markup That Actually Moves the Needle

Of all schema types, these correlate most strongly with AIO citation probability:

Article with author: { @type: Person, url: bio } and dateModified — +31% citation rate
FAQPage with mainEntity array — +22% citation rate
HowTo with step-by-step — AIO cites steps verbatim in procedural queries
Product with aggregateRating and offer — essential for commercial AIO
Organization with sameAs linking to social / Wikipedia — +18% for brand-adjacent queries

Validate every schema block with our JSON validator and schema markup generator. Invalid schema silently drops from eligibility; you won't see an error, you'll just lose citations.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "How to Compress PDFs Without Quality Loss",
  "author": {
    "@type": "Person",
    "name": "Jane Doe",
    "url": "https://example.com/authors/jane-doe",
    "sameAs": ["https://twitter.com/janedoe"]
  },
  "datePublished": "2026-04-01",
  "dateModified": "2026-04-15",
  "publisher": {
    "@type": "Organization",
    "name": "Example",
    "url": "https://example.com"
  }
}

llms.txt, robots, and the Crawling Layer

Three files govern whether LLM bots can access your content. Treat them as tier-zero infrastructure.

robots.txt — explicitly allow the bots you want. GoogleBot, ChatGPT-User, GPTBot, PerplexityBot, ClaudeBot, CCBot. Blocking LLM bots removes you from training and retrieval for those systems.
llms.txt (at /llms.txt) — emerging convention to expose a Markdown-formatted guide for LLMs. Not yet a Google signal but costs nothing to publish.
sitemap.xml — include all pages you want indexed, with accurate lastmod dates. Recency is a ranking input.

Check your robots.txt and sitemap with our generators. Use URL parser to verify canonical URLs in your sitemap match the canonical meta on each page.

Measuring AIO Citations in Search Console

Search Console launched dedicated AIO reporting in late 2025. The data lives under Performance with a new "Search Appearance" filter: "AI Overview". Click it to see impressions and clicks specifically from AIO.

Weekly monitoring workflow:

Export AIO impressions + clicks by query.
Compare to 4-week trailing average. Flag queries where AIO citations dropped >20%.
Visit the query in an incognito session. Who's cited now? What's different about their passage versus yours?
If a competitor overtook you, audit their passage structure, schema, and specificity.
Update your passage to match or exceed their Information Gain.

For query-level exports, pipe the Search Console CSV through our CSV viewer or CSV to JSON converter for easier analysis.

Case Study: Three Pages, Three Outcomes

Page A: High rank, no citations

A home-improvement blog ranked position 2 for "how often to clean gutters." 42K monthly searches. AIO appeared on the query. Citation count: zero for two months.

Audit revealed: 780-word paragraphs, zero original data, no dates on sections, weak schema (just BlogPosting), content shape matching every other generic blog post on the topic. Fix: rewrote the page into 12 H2 sections each 150-250 words, added a small reader survey's results (230 responses, homeowners reporting cleaning frequency), added Article schema with author bio, added FAQ schema. Within six weeks, 4 citations on the target query plus 11 citations on related sub-queries.

Page B: Low rank, multiple citations

A niche developer blog ranked position 28 for "rate limiting algorithms compared." 8K monthly searches. AIO cited the page within a week of publication. Why? The post included a benchmark table comparing five algorithms on identical traffic patterns with actual millisecond measurements. Unique data beat established authority.

Page C: The zero-click tax

A SaaS help doc ranked position 1 for a long-tail support query. AIO answered the question completely. Total clicks dropped from 1,800/month to 340 within two months. Citation was maintained; traffic wasn't.

Lesson: for informational queries with clean answers, AIO absorbs most clicks. If your revenue depends on traffic, the pages to optimize are the complex, nuanced topics AIO can't fully summarize, plus transactional and comparison content where users still click through.

The AIO Audit Checklist

Run this against any page you want cited:

[ ] Each H2 section is a self-contained mini-answer
[ ] Paragraphs under 60 words on average
[ ] First paragraph contains the core answer statement
[ ] At least one specific number or statistic in the first 300 words
[ ] Original data, survey, or case study somewhere on the page
[ ] Author byline with @type: Person schema
[ ] dateModified matches the visible "Updated" date
[ ] FAQ schema covers 5-10 likely sub-queries
[ ] At least one table for comparisons
[ ] Table of contents with anchor links
[ ] Schema validated (no warnings)
[ ] robots.txt allows GPTBot, ClaudeBot, PerplexityBot
[ ] Sitemap includes the URL with accurate lastmod
[ ] Internal links from topically-adjacent pages
[ ] At least one external citation to an authoritative source

Sixteen items. Every serious page we see cited by AIO hits 13+ of them.

Frequently Asked Questions

Does AIO cite me less if I block LLM crawlers?

Blocking bots removes you from those systems' retrieval entirely. GoogleBot is separate; Google uses its own index for AIO. Blocking ChatGPT doesn't affect AIO. Blocking GoogleBot does.

How often does AIO refresh citations?

Typical churn is weekly to monthly. Competitive queries see citations rotate more frequently as Google tests different passage combinations. Low-competition queries can have stable citations for months.

Should I add llms.txt?

It costs nothing and some LLM retrieval systems honor it. Not a major Google signal yet. Add it to your tier-1 content pages but don't make it the priority.

Do AIO answers include ads?

Sponsored content is starting to appear in AIO for commercial queries as of Q1 2026. Organic citations still occupy the bulk of the surface.

Can small sites beat big sites in AIO?

More often than in traditional rankings. Passage-level retrieval plus Information Gain rewards specificity over authority for most informational queries. We regularly see pages with <100 referring domains cited ahead of pages with thousands.

How long does it take a new page to get cited?

With strong passage structure and correct schema, the first citation often appears within 1-2 weeks of indexing. Without those signals, the page may never get cited even if it ranks.