How Google's AI Actually Decides What to Cite — A Technical Breakdown

Most advice about ranking in AI search reads like guesswork dressed up as strategy. "Optimize for AI." "Make your content AI-friendly." "The algorithm wants X."

The problem is that there isn't one algorithm. There's an ensemble of specialized systems, each doing something different, each triggering under different conditions. Some handle retrieval. Some handle ranking. Some handle language understanding. And they all run simultaneously.

If you want to understand how Google's AI search features — including AI Overviews — decide what to cite, you need to understand how these systems actually work together. Not the marketing version. The technical reality, drawn from Google's own documentation and public statements from the engineers who built them.

That's what this article covers.

---

The Two Core Technologies Behind AI Search

Before getting into ranking systems, it helps to understand the two mechanisms that power AI-generated search responses, as described in Google's own AI optimization guide for developers.

Retrieval-Augmented Generation (RAG)

AI Overviews don't generate answers from the model's training data alone. They use a technique called Retrieval-Augmented Generation — RAG.

Here's the sequence:

1. Google's traditional ranking systems retrieve a set of relevant pages for a query
2. Those pages are fed to a large language model as context
3. The LLM generates a synthesized response grounded in that retrieved content
4. The response includes clickable citation links back to the source pages

This is a critical distinction. The LLM isn't deciding on its own what's true or relevant. It's working with pages that Google's ranking systems already selected. The ranking systems act as a filter — and the quality of that filter determines what the AI can cite.

This means everything that matters for traditional search ranking also matters for AI citations. The ranking systems are the gatekeepers.

Query Fan-Out

The second mechanism is query fan-out. When a user asks a complex question, Google's AI doesn't just run one search. It generates related sub-queries to fetch additional results, building a more comprehensive answer.

For example, a query like "best CRM for small agencies that integrates with Slack" might trigger sub-queries about CRM comparisons, Slack integrations, and small business software recommendations — pulling results from each to construct a complete response.

This has a practical implication: your content doesn't need to match the exact query a user types. It needs to match the sub-queries that the AI system generates from that query. And those sub-queries are determined by the same ranking systems we're about to examine.

---

The Ranking Systems Ensemble

Pandu Nayak, Google Fellow and VP of Search, put it plainly in a 2022 blog post: "Search runs on hundreds of algorithms and machine learning models... Each algorithm and model has a specialized role."

Not one model. Not one algorithm. Hundreds — each specialized. Here are the major ones that matter for understanding AI citation selection.

RankBrain (2015) — Concept-Based Ranking

RankBrain was Google's first deep learning system applied to search. Its primary job: understanding the relationship between words and concepts.

What it does: Maps queries to conceptual meanings rather than matching keywords literally.

The canonical example: A user searches "what's the title of the consumer at the highest level of a food chain." RankBrain understands this maps to the concept of "apex predator" — even though those words never appear in the query.

Its role in the ensemble: Ranking. Once relevant documents have been retrieved, RankBrain helps determine the ordering of results by understanding conceptual relevance. It doesn't find documents — it helps decide which of the found documents best matches the user's actual intent.

Why it matters for AI citations: When AI Overviews synthesize responses, the underlying retrieval set has already been influenced by RankBrain's conceptual understanding. Content that thoroughly covers a concept — not just a keyword — gets ranked higher in that retrieval set, making it more likely to be cited.

Neural Matching (2018) — Concept-Based Retrieval

Neural Matching arrived three years after RankBrain and handles a different part of the pipeline entirely.

What it does: Understands "super fuzzy" concept representations. Instead of looking at individual keywords, it looks at entire queries and entire pages to determine conceptual relevance.

The canonical example: A user searches "insights how to manage a green." Without Neural Matching, this query is nearly incomprehensible. But the system maps it to management tips based on the concept of color-based personality profiles — connecting the query to pages about managing people with "green" personality types.

Its role in the ensemble: Retrieval. While RankBrain orders results, Neural Matching helps find relevant documents from Google's massive index in the first place. It's the system that says "this page is conceptually related to this query" even when the surface-level terms don't overlap.

Why it matters for AI citations: This is the system that determines whether your page even makes it into the candidate set that gets fed to the LLM. If Neural Matching doesn't recognize your content as conceptually relevant, the AI never sees it — and can never cite it.

BERT (2019) — The Language Understanding Layer

BERT — Bidirectional Encoder Representations from Transformers — represented the single biggest leap in Google's ability to understand natural language. And unlike RankBrain and Neural Matching, which have focused roles, BERT does both retrieval and ranking.

What it does: Understands word combinations and sequences in context. Specifically, it understands how small words change meaning — words that previous systems often dropped as unimportant.

The canonical example: A user searches "can you get medicine for someone pharmacy." The phrase "for someone" is the entire point of this query — the user wants to know about picking up a prescription on behalf of another person. Previous systems might have ignored "for someone" as filler words and returned generic pharmacy results. BERT understands that those two small words completely change the query's meaning.

Its role in the ensemble: Both retrieval and ranking, for almost every English-language query. Google confirmed that BERT is used in nearly every English search. It doesn't replace RankBrain or Neural Matching — it adds a layer of language comprehension on top of them.

Why it matters for AI citations: BERT's influence on AI citations is pervasive. Because it processes almost every query, it affects which documents get retrieved and how they get ranked for virtually all searches. Content that uses natural, precise language — where word relationships and context are clear — aligns with what BERT is optimized to understand.

MUM (2021) — The Most Powerful System Google Barely Uses

MUM — Multitask Unified Model — is, by Google's own description, 1,000 times more powerful than BERT. It understands and generates language. It's trained across 75 languages simultaneously. It's multimodal, processing both text and images.

And Google uses it for almost nothing in general search.

Where MUM is actually deployed: Specific, narrow applications. Vaccine information quality. Some featured snippet improvements. Visual search through Google Lens. Crisis information responses.

Where MUM is NOT deployed: General search ranking. General retrieval. The day-to-day queries that make up the vast majority of search traffic.

Why this matters: The marketing narrative around MUM has far outpaced its actual deployment. Plenty of "optimize for MUM" articles exist. But optimizing for a system that doesn't affect your queries is wasted effort. MUM matters for the future of search — Google will almost certainly expand its use — but for today's AI citations, BERT, Neural Matching, and RankBrain are doing the heavy lifting.

---

The Supporting Cast — Systems That Shape the Candidate Set

The four systems above get the most attention, but several other ranking systems play direct roles in determining which pages make it into the retrieval set that feeds AI Overviews.

PageRank — Still Alive, Still Counting

PageRank has evolved dramatically from its original 1998 formulation, but the core principle — that links from other sites serve as signals of authority and trust — remains active in Google's ranking systems.

The practical impact hasn't changed: pages with genuine backlinks from authoritative sources rank higher. For AI citations, this means the LLM is more likely to encounter and cite pages that have earned real editorial links. No amount of content quality overcomes a complete absence of authority signals.

Original Content Systems

These systems specifically identify and elevate original reporting and original research over derivative or rewritten content. When five sites cover the same story, the site that broke it gets priority.

For AI citations, this creates a clear advantage for brands that publish original data, proprietary research, first-hand case studies, and novel analysis. The AI is more likely to cite the original source than the tenth blog post summarizing someone else's findings.

Passage Ranking

This is one of the most underappreciated systems in the context of AI search. Passage ranking allows Google to identify and evaluate specific sections within a page independently from the page as a whole.

A 3,000-word guide about CRM implementation might have one section about data migration, another about user training, and another about integration architecture. Passage ranking means each of those sections can independently match different queries — and each can be independently cited by AI Overviews.

This has a direct structural implication: well-organized pages with clear section headings and self-contained passages are more "citable" than pages where information runs together without clear delineation. The AI can point to a specific passage and say "this answers the question."

Freshness Systems

Some queries demand recent content. "Best project management tools 2026" has different freshness requirements than "how does TCP/IP work." Google's freshness systems identify time-sensitive queries and weight recent content accordingly.

For AI citations on trending or evolving topics, recently published or recently updated content has a structural advantage. A page last updated in 2023 is less likely to be cited for a query where freshness matters than a page updated this quarter.

Helpful Content Signals

Google's standalone "Helpful Content System" was absorbed into core ranking in March 2024. This wasn't a demotion — it was a promotion. The signals that identify people-first, genuinely helpful content are now baked into the core ranking systems rather than running as a separate overlay.

What this means: the distinction between "helpful" and "unhelpful" content isn't a separate filter anymore. It's woven into how every ranking system evaluates pages. Content created primarily to manipulate rankings — rather than to inform readers — is penalized by the core systems themselves.

SpamBrain

Google's AI-powered spam detection system runs continuously and gets regular updates. It identifies manipulative link building, auto-generated content created solely for rankings, cloaking, and other policy violations.

For AI citations, SpamBrain acts as a negative filter. Pages flagged by SpamBrain won't make it into the retrieval set, regardless of how well they might otherwise perform. The system is specifically designed to catch AI-generated spam content — a rapidly growing category.

---

15% of Queries Have Never Been Searched Before

Here's a statistic that reframes everything: Google has confirmed that 15% of its daily queries are completely new. The search engine has never processed them before.

Fifteen percent of billions of daily queries is hundreds of millions of never-before-seen searches. Every single day.

This is why Google invested in concept-understanding systems like RankBrain, Neural Matching, and BERT rather than maintaining lookup tables of queries to results. Pre-programmed rules can't handle queries that don't exist yet. Concept matching can.

For content strategy, this has a specific implication: you cannot anticipate every query that might surface your content. You don't need to. If your content comprehensively covers a concept — with the right depth, structure, and authority signals — Google's systems can match it to queries that neither you nor anyone else has ever imagined.

This is the technical reality behind the advice to "write for topics, not keywords." RankBrain, Neural Matching, and BERT don't match keywords. They match concepts, language patterns, and meaning. Content organized around comprehensive concept coverage gets retrieved for a long tail of queries that keyword-targeted content misses entirely.

---

The Ensemble Effect — Why No Single Trick Works

Here's where most AI search optimization advice falls apart. It assumes there's a lever to pull — one system to optimize for, one technique to apply.

But the systems operate simultaneously:

• RankBrain evaluates whether your content conceptually matches the query's intent
• Neural Matching determines whether your page belongs in the candidate set at all
• BERT assesses whether your language precisely addresses the query's actual meaning
• PageRank weighs whether other authoritative sources vouch for your content
• Original Content Systems check whether you're the source or just echoing someone else
• Passage Ranking evaluates whether specific sections of your page answer specific questions
• Freshness Systems check whether your content is current enough for the query type
• Helpful Content signals assess whether the page was written for people or for algorithms
• SpamBrain checks whether any of your tactics cross into manipulation

A page needs to perform well across all of these simultaneously. Gaming one system while neglecting others doesn't work — and gaming any single system is nearly impossible when the systems collectively evaluate conceptual depth, language precision, structural clarity, authority, originality, freshness, user intent, and manipulation signals all at once.

The only content strategy that performs well across every system in the ensemble is the one Google keeps repeating because it's literally how the technology works: create genuinely useful, original, well-structured content that demonstrates real expertise.

That's not marketing advice. It's an engineering constraint. The systems were designed to reward exactly that.

---

What This Means for AI Citations — Practical Implications

Understanding how these systems work together points to specific, actionable conclusions about how to earn AI citations.

Concept Coverage Beats Keyword Targeting

RankBrain and Neural Matching both operate on concepts, not keywords. A page that thoroughly covers "customer data platform implementation for mid-market retailers" will match queries that use none of those specific words — as long as the conceptual coverage is comprehensive.

Write naturally about your subject matter. Use the terminology your audience uses. Cover adjacent concepts that a knowledgeable person would expect to find. The systems are built to recognize topical authority.

Comprehensive Pages Beat Scattered Thin Pages

Passage ranking means Google can cite specific sections of a longer page for different queries. One thorough, well-structured page about your domain area can rank for dozens of different queries through passage-level matching.

The old SEO playbook of creating separate pages for every keyword variation works against this system. Instead: create definitive resources organized with clear headings, where each section is a self-contained, citable passage.

This is exactly the mechanism behind AI Overview citations. When an AI Overview cites a specific paragraph from your page, passage ranking identified that section as the most relevant passage for that particular query or sub-query.

Original Research and Data Create Structural Advantages

Original Content Systems specifically reward first-party research, proprietary data, novel analysis, and original reporting. If your content is a summary of what five other sources already published, those other sources get the citation priority.

For brands, this is an argument for publishing:

• Proprietary data from your own platform or customer base
• Original survey results with your own methodology and sample
• First-hand case studies with specific metrics and outcomes
• Novel frameworks you've developed from real practice
• Primary analysis of industry data that nobody else has done

Each of these creates content that the Original Content Systems can identify as source material rather than derivative commentary.

Authority Still Requires Real Links

PageRank has evolved, but link analysis still underpins authority measurement. Content without any external authority signals — no backlinks from relevant, authoritative sources — faces a structural disadvantage regardless of its quality.

This isn't about link building in the traditional SEO sense. It's about creating content valuable enough that industry publications, professional communities, and other authoritative sources reference it naturally. Original research, original data, and genuinely novel insights earn links. Repackaged conventional wisdom does not.

Traditional Search Ranking and AI Visibility Are the Same Problem

This might be the most important practical takeaway: there is no separate "AI algorithm."

AI Overviews use RAG, which means the AI generates responses from pages that Google's existing ranking systems retrieved. The ranking systems that determine traditional search results are the same systems that determine what the AI can see and cite.

Improve your traditional search ranking and your AI visibility improves with it. The two are architecturally coupled. There is no shortcut that improves AI citations without improving your underlying search performance, because the AI citations are downstream of search performance.

---

The Technical Reality vs. The Marketing Narrative

A significant gap exists between how AI search actually works and how it's described in most marketing content.

The marketing narrative: AI search is a new paradigm requiring new optimization strategies. Previous SEO work is obsolete. You need to "optimize for AI" using new techniques.

The technical reality: AI search features are built on top of the same ranking systems that power traditional search. RAG uses those ranking systems as the retrieval layer. The AI can only cite what the ranking systems surface. The ranking systems evaluate the same signals they always have — conceptual relevance, language quality, authority, originality, freshness, and user intent.

The marketing narrative: One system (often MUM or "the AI") decides what to cite.

The technical reality: Hundreds of specialized systems each evaluate different aspects of quality and relevance. They run simultaneously. No single system is in charge. The ensemble design is specifically engineered to prevent gaming.

The marketing narrative: You need to understand how the AI "thinks" to optimize for it.

The technical reality: The AI doesn't think. It generates responses from retrieved pages. The retrieval is done by ranking systems that evaluate well-understood signals. Understanding those signals is the entire game.

This distinction matters because it determines where you invest effort. Chasing "AI optimization" as a separate discipline leads to busywork. Investing in content quality, topical authority, information architecture, and genuine expertise compounds across every ranking system simultaneously — including the ones that feed AI search features.

---

A Framework for Thinking About AI Citations

Based on how these systems actually work, here's a framework for evaluating whether your content is positioned to earn AI citations:

Every "yes" makes your page more likely to appear in the retrieval set that feeds AI-generated responses. Every "no" is a reason the AI might cite a competitor instead.

---

Where This Goes Next

Google's ranking systems ensemble continues evolving. MUM's capabilities — multilingual understanding, multimodal processing, generative capacity — will almost certainly expand into broader ranking applications. New systems will be added. Existing systems will be refined.

But the architectural pattern is set: multiple specialized systems, each evaluating different quality signals, all running simultaneously. This ensemble approach makes the systems collectively robust against manipulation while rewarding genuine expertise, originality, and usefulness.

For brands monitoring their visibility across AI search, the implication is that sustainable AI citation performance comes from the same place sustainable search performance has always come from — being the most authoritative, most original, most comprehensive, most clearly structured source on topics within your domain.

The technology behind AI search is genuinely sophisticated. The strategy for earning citations from it is not complicated. Those two things can both be true at the same time.

---

The ranking system details in this article are drawn from Google's Ranking Systems Guide, Google's AI Optimization Guide, and Pandu Nayak's blog post "How AI Powers Great Search Results" (February 2022). System descriptions reflect their documented roles as of the publication date.