How Do Ai Overviews Choose Sources

• A page that ranks first can still be left out of an AI Overview, and a page on page two can be cited.
• Traditional SEO sets the entry conditions.
• Passage clarity, entity density, and original detail are the content features that appear to influence citation, more than format.
• Authority has become a weaker predictor than many assume.
• The citation layer refers to the attribution step, where an AI Overview ties specific sentences to specific documents. For instance, according to Ziptie.ai's analysis, organic-AIO overlap collapsed from 76% to 38% in under a year, based on Ahrefs' study of 863K keywords.

A page that ranks first can still be left out of an AI Overview, and a page on page two can be cited. Google describes AI Overviews as built on its core Search ranking and quality systems, so organic ranking is an input rather than a guarantee. Google also states that AI Mode and AI Overviews may use query fan-out, issuing several related searches across subtopics and data sources. A source can therefore be cited for one subtopic even when it does not rank for the literal query. Third-party research points the same way. According to Ziptie.ai, organic-AIO overlap fell from 76% to 38% in under a year, based on Ahrefs' study of 863K keywords. The practical takeaway is that rank remains necessary for eligibility but no longer predicts selection on its own.

Google publishes no selection formula. Treat third-party figures as measurements of patterns, not rules. Google formally rolled out AI Overviews in May 2024, positioning them to answer complex, multi-step queries with a summary that prominently includes links to learn more (per Google's May 2024 blog post).

Traditional SEO sets the entry conditions. Google's documentation states there are no additional technical requirements for AI Overviews beyond being indexed and eligible to show with a snippet. A page must be accessible, crawlable, and eligible for snippets, and it needs no special AI markup or file.

Crawlable visible text matters because Google's JavaScript SEO guidance notes that rendered pages can wait in a rendering queue, and server-side or pre-rendering is still recommended. Google also recommends non-commodity content and a clear technical structure.

Eligibility gates the candidate pool, but it does not decide which candidate is quoted. That decision depends on the passage-level factors covered in the next section. According to Seer Interactive's September 2025 study (3,119 informational queries, 25.1 million impressions), organic CTR dropped 61% (from 1.76% to 0.61%) when AI Overviews were present.

Passage clarity, entity density, and original detail are the content features that appear to influence citation, more than format. Ziptie.ai reports that pages with 15 or more recognized entities had 4.8 times higher selection probability. That finding is a correlation from a vendor analysis, not a Google disclosure. Format signals are weaker than commonly assumed.

Rankera's study of 30,613 overviews, cited by Exalt Growth, found that FAQ schema and question headings barely predicted citation, with an AUC of 0.51, close to a coin flip. Schema remains useful for machine readability but should not be treated as a citation lever. Write passages that name the mechanism, the tool, and the figure, because specificity is what a model can attribute to a source.

Authority has become a weaker predictor than many assume. Ziptie.ai reports that the correlation between Domain Authority and AI Overview citation fell from r=0.43 to r=0.18. Corroboration matters more. Exalt Growth's reading of Google's US11900068 patent describes an AI Overview sentence being verified by two documents, each with its own link.

If that mechanism operates as described, a claim that other reputable sources independently support is easier to attribute. Google's own guidance frames E-E-A-T as quality evaluation rather than a citation switch, so the defensible position is that a claim repeated across independent, credible sources is structurally easier to cite than an isolated one.

The citation layer refers to the attribution step, where an AI Overview ties specific sentences to specific documents. According to Exalt Growth, Google's patent US11769017, 'Generative summaries for search results', was filed in March 2023 and granted in September 2023. The same analysis reads US11900068 as claiming that one AI Overview sentence is verified by two documents, each with its own link.

This implies attribution works per claim, not per page. A single page may be cited for one sentence while a competitor supplies another. For publishers, the implication is that each factual passage should stand alone and carry its own verifiable detail.

The traffic effect is measurable and large. According to Seer Interactive's September 2025 study, which covered 3,119 informational queries and 25.1 million impressions, organic CTR fell 61%, from 1.76% to 0.61%, when AI Overviews were present. Paid CTR fell 68%, from 19.7% to 6.34%, on the same query set.

According to The Digital Bloom's 2026 AI Citation Position and Revenue Report, brands cited as AI Overview sources gained 35% more organic clicks from branded searches. Press Gazette's 2025 Trends Report estimated global Google-referred traffic fell about 33% in 2025. Publishers can measure citations with Citation Analytics-style tracking across engines, since Search Console alone does not show which overview sentences cite a page.

What remains unknown is the source-selection formula itself, which Google does not publish, along with the weights of any signal and the thresholds that govern citation. Most figures in circulation come from vendor studies with different sample frames, so cross-study comparisons should be treated cautiously. The patent evidence describes mechanisms that may or may not match production behavior.

Google Search Console now offers a Search generative AI control, with pages set to exclude appearing in AI Overviews or AI Mode and include as the default; per the AI Syndicate article, the rollout date is August 31, 2026. Claims that a specific schema, word count, or format guarantees citation lack published evidence, so verify such advice against your own citation data before acting on it.

Related guides

• How ChatGPT Picks Sources: Retrieval, Bing, and Citations
• How AI Answer Engines Pick Sources: The Signals Behind
• How to Appear in AI Search Results: The AEO Playbook