How Google AI Overviews and AI Mode pick sources (2026)

Google's two AI search features run on the Google Search index, but they don't simply quote the top organic results. Google documents the basics: who is eligible, that the features fan out into several searches, and that no special markup is needed. The rest comes from patents and vendor studies. Each figure below names its sample, its date and how strong the evidence is. Both features change often, so treat every number as a dated snapshot.

Which pages are eligible for AI Overviews and AI Mode

A page is eligible for AI Overviews and AI Mode if Google has indexed it and it is allowed to show a snippet in Search. Google's documentation on AI features says there are no extra requirements: no special schema, no special files and no separate AI markup. That is confirmed by Google.

Pages lose eligibility through the same controls that limit ordinary snippets:

• A noindex directive, or a page blocked from Googlebot.
• A nosnippet directive, or max-snippet set to 0.
• data-nosnippet on the passage that would otherwise be quoted.

Google also says AI Overviews and AI Mode “may use different models and techniques,” and its I/O 2025 post says a custom Gemini model runs both. Extra markup does not buy a place. Ahrefs' controlled study of JSON-LD (May 2026, 1,885 pages vs about 4,000 matched controls) found a tiny 4.6% drop in AI Overview citations and a non-significant 2.4% rise in AI Mode.

How fan-out works in AI Overviews and AI Mode

Google builds an AI Overview or AI Mode answer by running several related searches, called fan-out queries, and combining what they return. Google's documentation describes “issuing multiple related searches across subtopics and data sources.” For AI Mode, Google's I/O 2025 post describes “breaking down your question into subtopics and issuing a multitude of queries simultaneously,” and says Deep Search issues “hundreds of searches.” Those statements are confirmed. The counts below are not:

• AI Mode: about 8–12 searches per turn in third-party measurements (plausible).
• Gemini 3 API grounding: an average of 10.7 searches, from 3 to 28, per Seer (plausible).
• AI Overviews: little fan-out is visible in the page source, and Google says only that the feature “may” fan out (plausible).

Google's patents describe the idea in more detail. The “Thematic search” patent (US12158907B1) groups a query into themes and extracts passages for each. “Search with stateful chat” (US20240289407A1) builds query variants from the conversation. A patent shows what Google designed, not what runs in production.

Ranking for those sub-queries predicts citations better than ranking for the head term. Surfer's study of 10,000 keywords and 173,902 URLs found a Spearman correlation of 0.77 between the number of fan-out queries a page ranked for and its AI Overview citations. Ranking for the main query plus a fan-out query made a page 161% more likely to be cited. That is correlational, and Surfer generated its fan-outs with Gemini as a proxy. optmz's fan-out queries view lists the searches engines actually ran for a tracked question.

Does organic rank still matter for AI Overviews? 38% vs 17% in 2026

Organic rank still matters for AI Overviews, but much less than it did, and most citations now come from outside the top 10. Studies disagree on the exact share because they count differently: top 10 vs top 100, URLs vs domains, and parsers that changed between runs.

Being in the index and ranking somewhere still counts. Ahrefs' 20-day study found that pages ranking anywhere in Google were about 7 times more likely to stay cited for a week or more in AI Overviews and AI Mode (strong). The working rule: rank somewhere for the sub-queries. A top-10 position for the head term is neither necessary nor sufficient.

AI Overviews vs AI Mode: why they cite different pages

AI Overviews and AI Mode usually cite different pages for the same question. Ahrefs found only 13.7% URL overlap between them, even though their answers were 86% semantically similar. Put another way, they cite different sources about 87% of the time. That is a strong finding.

AI Mode also cites more sources per answer and leans on different kinds of sites:

• Citations per answer: AI Mode 15.2, AI Overviews 11.1, Gemini app 6.6 (Profound, plausible).
• Domains per answer: AI Mode 16.4, AI Overviews 10.2, Gemini app 3.4 (GetIntel, plausible).
• AI Mode leans on Google-owned or partner properties such as YouTube, Reddit and LinkedIn (Semrush, plausible).
• The Gemini app shares only 27–37% of its citations with AI Mode (BuzzStream and BrightEdge, directionally replicated).

AI Mode draws on sources AI Overviews mostly don't: Knowledge Graph entities, Shopping and Maps data, and opt-in Personal Intelligence, which uses Search history and, if the user agrees, Gmail and Calendar data (confirmed by Google). Treat AI Overviews and AI Mode as separate surfaces in client reports.

How Google lifts passages, and how to find them with #:~:text= fragments

Google quotes specific passages, not whole pages, and the cited link often shows which passage it used. Citation links in AI Overviews and AI Mode carry #:~:text= text fragments that point at the lifted text, as RESONEO's inspector documents. Click a citation and the browser scrolls to that passage and highlights it.

Google's patent on generative summaries (US11769017B1) describes writing a summary and then linking portions of it to documents that verify them. The verifying source can come from a separate search on the summary text, so a cited page may not be the page the sentence was generated from. The patent is confirmed as a disclosure; that it runs this way in production is plausible.

Gemini grounding may or may not work the same way as AI Overviews, but it is the closest documented case. DEJAN's tests suggest a budget of about 2,000 words per query, spread across sources by rank, with roughly a third of any one page getting through. Navigation, boilerplate and time-sensitive promotions are dropped, and the model sees the query, title, snippet and publication time (plausible). CXL found 55% of AI Overview citations came from the first 30% of a page, but on only 100 examples.

The practical result: put a clear, quotable answer near the top of each section, and keep promotional and navigation text away from the facts you want lifted.

Freshness: AI Overviews are the exception

AI Overviews do not favor newer pages the way other AI assistants do. Ahrefs' freshness study (17 million URLs) found that AI assistants overall cite pages 25.7% newer than Google's organic results. AI Overviews were the exception: their cited pages were on average 16 days older than the organic results. That is a strong finding for this specific feature.

For AI Overviews, a recent date is not a lever by itself. Update a page when its facts change, keep the visible date honest, and don't refresh dates to look new. Google warns against that, and it is close to spam. Freshness matters more for ChatGPT and Perplexity, which do skew toward newer pages.

How volatile AI Overviews are, and how to measure them

AI Overviews change every few days. Ahrefs (Nov 2025, 43,000 keywords, at least 16 observations each) measured a 70% chance that an AI Overview changed between observations, an average persistence of 2.15 days, and 45.5% citation turnover per change. The wording changed more than the meaning: semantic similarity stayed at 0.95. Popular queries were no more stable than rare ones. This is confirmed by repeated observation.

Over longer windows the churn adds up. Profound (Jul 2025, about 80,000 prompts per platform) found that 59.3% of the domains cited in AI Overviews were different after one month (strong, vendor data). SISTRIX (Apr 2026, 82,619 prompts over 17 weeks) found 56% weekly domain churn in AI Mode, though 86% of prompts kept a stable core of 1–5 domains (plausible).

One check tells you little. Sample each question repeatedly and report rates with ranges, as the guide to measuring AI visibility with sampling and confidence ranges describes. optmz's AI visibility tracking can add a weekly check of AI Overviews and AI Mode alongside its engine samples.

Google-Extended: what it controls and what it doesn't

Google-Extended controls whether a site is used for the Gemini apps and Vertex AI grounding, but it does not control Google Search, which includes AI Overviews. Google's crawler documentation confirms the split. Blocking Google-Extended can remove a site from Gemini app grounding while leaving it in AI Overviews.

To limit what AI Overviews and AI Mode can show, use the Search snippet controls: nosnippet, max-snippet and data-nosnippet. Search Console also added a setting in 2026 that removes a site from all generative AI features (strong, from coverage of the launch). Every one of these costs visibility, so decide deliberately. The AI crawler user-agent list has the exact Google and other AI crawler tokens.

How to see AI Overviews and AI Mode in Search Console

Search Console's Generative AI report shows impressions from AI Overviews and AI Mode for each page, but no queries and no clicks. Google launched the report on June 3, 2026, rolled it out to all sites by August 31, 2026, and backfilled data to May 18, 2026 (strong, from coverage of the launch).

• Impressions only, broken down by page, country, device and date.
• AI Overviews and AI Mode are combined in one figure.
• No queries, no clicks and no API. The Search Console API has no AI option in its type field.
• These impressions are already counted inside the Web totals.
• Brand mentions without a link don't count.

The Generative AI report is the best first-party check on whether a page's exposure in Google's AI features changed after an edit. Export it from the interface, compare edited pages with pages you didn't change, and expect noise given how often the features change.

Checklist for AI Overviews and AI Mode in 2026

To be cited in AI Overviews and AI Mode, a page must be indexed and snippet-eligible, rank somewhere for the sub-queries, and hold a passage Google can lift. Check these in order.

1. Confirm the page is indexed and has no noindex, nosnippet, max-snippet:0 or stray data-nosnippet.
2. Let Googlebot fetch the page. Treat Google-Extended as a separate decision about Gemini.
3. List the sub-queries a question fans out into and check where the page ranks for each, not only the head term.
4. Open each section with a direct, quotable answer, away from navigation and promotions.
5. Click existing citations and read the #:~:text= fragment to see which passage Google lifted.
6. Track AI Overviews and AI Mode separately, since they share about 13.7% of URLs.
7. Don't count on fresh dates or extra schema to win AI Overview citations.
8. Sample repeatedly, and use the Search Console Generative AI report as a first-party check.