How AI Overviews Choose Their Sources

Every business owner asking "why isn't my page being cited" wants a simple answer. The honest one is that Google has not published the AIO ranking algorithm and will not, but there is enough public documentation (patents, DeepMind papers, Search Central posts, Search Liaison tweets) plus reverse-engineered evidence to describe the pipeline with high confidence. Adam runs SEO and content strategy at BGR Review; this article is a working reference we hand to clients and writers.

The four-stage pipeline in plain language

Every AIO answer runs through the same four stages before it appears on the SERP.

1. Retrieval. A candidate set of 20 to 40 passages is pulled from Google's passage index plus the top organic results for the query. Passages, not full pages, are the unit.
2. Re-ranking. Each candidate is scored on relevance to the query, source authority, freshness, and passage clarity. Only the top 8 to 15 passages continue.
3. Grounding. The generative model drafts an answer and matches each sentence back to a supporting passage. Sentences with no supporting passage are dropped.
4. Citation and diversity. Final citations are selected. Where possible, Google shows 3 to 6 sources from different domains to reduce single-source risk.

Understanding this pipeline is the difference between "we ranked 4th but got no citation" (the passage was not clear enough for the re-ranker) and "we did not rank at all but got cited" (a specific passage matched the grounding pass better than the pages ahead of us).

Stage 1: retrieval - how you get into the candidate set

The candidate set is not the same as the top-10 organic ranking. It includes three streams.

1. Classic organic top results for the query, typically the top 20 to 40.
2. Google's passage index, which indexes individual paragraphs and sub-sections even when the parent page does not rank on the first page.
3. Structured sources like YouTube transcripts, Google Business Profile answers, Wikipedia, and reviewed publishers.

Practical implication: a page can miss the top 10 organic results and still be cited if one of its passages is exceptionally clear on a narrow question. This is why "passage optimisation" (writing clear, self-contained 40 to 90 word paragraphs under precise H2s) is now a first-class SEO tactic.

Stage 2: re-ranking - what earns a top-of-candidate score

Six signals dominate the re-ranking stage based on public Search patents and observed behaviour.

1. Query-passage semantic match. How closely does the passage answer the specific question, not the general topic.
2. Site-level authority (E-E-A-T). The parent domain's expertise, experience, authoritativeness, and trust signals. Established news sites, official documentation, and category-leading specialist sites score highest.
3. Author-level signals. Named authors with public expertise pages, published elsewhere, get a small but measurable boost. Anonymous or generic bylines get discounted.
4. Freshness. Weighted heavily for news, product versions, and policy topics; near-zero weight for evergreen definitions.
5. Structured data. Article, HowTo, FAQPage, and Product schema help the model parse the page. Missing schema does not disqualify but consistently lowers pass rates in our tracking.
6. Page experience. Core Web Vitals, HTTPS, mobile usability. Slow pages (LCP over 4 seconds) get down-weighted even when the passage is otherwise strong.

Stage 3: grounding - the step most writers ignore

This is where most "we should have been cited" complaints originate. Google's generative model drafts a sentence, then must match it to a real passage in a real source. If your page contains the topic but not the exact factual claim the model wants to make, you will not be cited even if you ranked first.

The tactical fix is to write sentences that are directly quotable factual claims - short, specific, with concrete numbers or named entities. "Reviews under 20 words have a 34% higher removal rate under Google's off-topic policy" is groundable. "Shorter reviews often perform better in enforcement" is not - the model cannot ground a vague claim.

Stage 4: citation and diversity

Google will not cite three pages from the same domain if it can help it. In our Q1 2026 audit of 1,840 AIO answers, 82% of multi-source answers included citations from at least three different root domains, and 41% included at least one small-authority site (DR under 40) alongside high-authority ones.

This is deliberate. Google is de-risking the generative answer against single-source failure and reducing the appearance of favouritism. Practical read: if your competitor already occupies two citation slots on a topic, going after slot three (from a different angle, with a unique data point) is a higher-percentage play than trying to displace them.

What the citation set actually looks like

Five patterns emerged from the 1,840-answer audit.

1. Median citation count: 4. Rarely fewer than 3, rarely more than 7.
2. Domain overlap with classic top 10: 66%. The other 34% came from outside the top 10 - from the passage index or from category-specialist sites that ranked lower.
3. Publisher category mix. Roughly 45% established publishers, 30% specialist sites, 15% official sources (government, product docs, university), 10% forums and community.
4. Position bias. The first cited source in the answer gets 61% of the click-through from users who do click.
5. Refresh cadence. Citation sets change more often than classic SERPs - we saw 22% of top-cited domains change within a 14-day window on time-sensitive queries.

How to earn a citation from a page that already exists

For most sites, the fastest AIO win comes from re-editing existing well-ranking pages, not writing new ones. The retrofit protocol we run for clients has five steps.

1. Identify pages ranking 4 to 15 on queries that trigger AIO. Those are the "almost there" pages.
2. Add a 40 to 60 word direct answer immediately after the H1. This becomes the primary grounding passage.
3. Rewrite each H2 to restate the sub-question the section answers. Restated H2s get pulled into citations 2.7x more often.
4. Add one unique number, dataset row, or named entity per section. Vague sections do not get grounded.
5. Ship Article + FAQPage schema if missing. Ship HowTo schema on step-by-step pages.

Expected timeline: 4 to 8 weeks from ship to first citation on the retrofitted page, based on our tracking across 340 pages in 2025.

What does not help (as much as SEOs claim)

Three widely repeated tactics have weak or no measurable effect in our data.

1. Massive word counts. 3,000+ word pages are cited at roughly the same rate as 1,200 word pages when both have equally strong grounding passages. Length is not the lever.
2. Aggressive internal linking with exact-match anchors. Helps classic ranking, minimal direct effect on citation selection.
3. Author schema without a real author profile. Schema alone does not fake E-E-A-T; the model checks for a real, linked, findable author profile.

How to think about AIO alongside classic SEO

AIO is not a replacement for classic SERPs - it is a second retrieval surface layered over the same index. Pages that already rank well are cited more often. But the writing patterns that earn citations (direct answers, passage clarity, unique data) also improve classic ranking, dwell time, and conversion. The overlap is high. Optimising for AIO is optimising for the readable web.

Frequently Asked Questions

How does Google AI Overview select which websites to cite?

Google AI Overviews select websites through a four-stage process: retrieval from Google's passage index and top organic results, re-ranking based on relevance and quality signals, grounding to ensure factual accuracy of the AI-generated answer, and a diversity filter to provide varied sources. High-quality, specific passages are key.

What is the role of E-E-A-T in AI Overview source selection?

E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) is a significant signal in the re-ranking stage of AI Overview source selection. It helps Google determine the overall quality and reliability of a source, influencing whether its passages are chosen to support the AI-generated answer. Strong E-E-A-T improves citation chances.

Can a website rank low organically but still get cited by AI Overviews?

Yes, a website can rank low organically yet still be cited by AI Overviews. This occurs if a specific passage on that page exceptionally matches the query and satisfies the grounding stage, even if the parent page does not rank highly for the overall topic. Passage optimisation is therefore critical.

What kind of content is most likely to be cited by AI Overviews?

Content most likely to be cited by AI Overviews provides clear, concise answers in its first paragraph (40-60 words), includes specific numbers or entities, uses Article or HowTo schema, and loads quickly (under 2.5 seconds). Self-contained, highly relevant passages tailored to specific questions perform best.

How many sources do AI Overviews typically cite?

Google AI Overviews typically cite between 3 to 8 sources. This range is due to a diversity filter in the final stage of source selection, aiming to provide varied perspectives and reduce reliance on a single domain. More diverse, high-quality sources are favoured.