How Google AI Overviews Pick Sources

Why This Question Changed

For two decades, the answer to "which page does Google show first" was governed by ranking: a list of ten blue links ordered by relevance and authority. The AI Overview, the summarized answer that now sits above those links for a large and growing share of queries, does not work that way. It is generated by a Gemini model that reads across multiple sources and writes a synthesis, attaching citations to the specific pages that informed specific claims. Being cited in that summary, and being cited in the prior list of links, are no longer the same achievement.

The practical consequence is measurable. Independent analyses of AI Overview citations report that the overlap between cited pages and the traditional organic top ten has fallen sharply, from roughly three quarters of citations in mid 2025 to closer to four in ten in later measurements. The exact figure varies by study and by query category, but the direction is consistent: ranking in the classic results no longer guarantees a place in the answer, and pages that rank lower, or not at all in the visible list, increasingly appear as cited sources.

This is the boundary where AIO, AI Optimization, separates from the SEO era. SEO optimized a page for a position in a list. AIO optimizes a business and its content so that an AI system can understand it, trust it, and reproduce its claims with confidence. Understanding how AI Overviews pick sources is the clearest available window into that shift, because Google has documented much of the mechanism.

Query Fan-Out: One Question Becomes Many

The first mechanism to understand is query fan-out. When a search triggers an AI Overview, the system does not run a single lookup against your typed words. It decomposes the question into a set of related sub-queries that together cover what a complete answer would require. Google has confirmed this behavior, and the same capability is visible in the public Gemini API, where a grounded model automatically generates one or more search queries and executes them before answering. A single user prompt can become several distinct searches.

Each sub-query produces its own set of results. The model then has a much wider pool of candidate pages than any one search would surface, and the sources that recur across several sub-queries, or that answer a sub-query unusually well, are the ones most likely to be drawn into the summary. This is why citation patterns have decoupled from the original results page. The Overview is assembled from the union of many searches, not from the single ranking a user would see.

Fan-out reframes the optimization target. The relevant question is no longer "do I rank for this keyword," but "do I clearly and completely answer the cluster of sub-questions this topic implies." A page that covers one narrow phrasing may win a position yet never enter an Overview, while a page that resolves the surrounding sub-questions can be retrieved repeatedly across the fan-out and cited as a result.

Grounding: Retrieve, Rerank, Blend

Once the sub-queries run, the system follows a grounding pipeline. Google's own developer documentation for Grounding with Google Search describes the sequence plainly: the service rewrites the prompt into search queries, retrieves results, reranks them, and blends the selected material into the model's context so it can write a grounded answer with inline supporting links. The AI Overview is the consumer-facing expression of this same retrieval-and-grounding pattern.

Grounding is also gated. In the documented API, a dynamic retrieval step assigns the prompt a prediction score between zero and one that estimates how much the answer would benefit from live search, with grounding engaged above a configurable threshold. Conceptually, the system decides both whether to ground an answer in external sources and which retrieved passages are worth placing in front of the model. Sources that are never retrieved, or that are retrieved but rank poorly in the rerank step, never reach the context window and therefore cannot be cited.

The lesson is that citation is downstream of retrievability. A page can be excellent and still be invisible to an Overview if a crawler cannot parse it, if its relevant claim is buried, or if its content does not map cleanly to the vectors the retrieval step matches against. Accessibility, one of the seven pillars of AIO, is not a courtesy here. It is the precondition for everything that follows.

Passage-Level Selection, Not Page-Level

AI Overviews operate at the level of the passage, not the whole document. The model needs a self-contained unit of text that answers a specific sub-question completely enough to be quoted or paraphrased with a citation attached. A page that contains the answer but spreads it across scattered sentences, conditional clauses, and cross-references is harder to extract than a page that states the answer directly and then supports it.

This favors content that is structured for extraction: a clear claim stated near the heading that introduces it, defined terms, direct answers to direct questions, and supporting evidence positioned close to the assertion it backs. Analyses of where citations originate within pages consistently find that the upper portion of a page contributes the largest share, with the middle next and the bottom least. That distribution reflects how the model reads, looking for the most complete and confident statement of an answer rather than mining a conclusion buried beneath preamble.

Clarity, in the AIO sense, is therefore a retrieval property, not a stylistic preference. The pillar of clarity means that a machine reading your page can locate one unambiguous answer to a given question and reproduce it without having to resolve hedges or reconcile contradictions. Pages that meet that bar give the model something safe to cite.

Trust, Entities, and the Confidence to Cite

Retrievability and clarity get a passage into consideration. Whether the model is willing to attach its name to that passage is a question of trust. AI Overviews draw on the same authority and quality signals that Google has long associated with experience, expertise, authoritativeness, and trustworthiness, and they appear to apply them as something closer to a gate than a gradient: a source that fails the trust assessment is unlikely to be cited regardless of how well-written its answer is.

Entities are central to this assessment. Google maintains a Knowledge Graph of people, organizations, products, and concepts, and content that is densely and accurately tied to recognized entities is easier for the system to place, disambiguate, and verify. A page that names its claims, its sources, and the organization behind it in ways that connect to known entities gives the model corroboration. A page that floats free of any verifiable entity context gives the model nothing to check against, and an answer engine biased toward caution will tend to cite the corroborated source instead.

This is where four AIO pillars converge: evidence, validation, expertise, and entity strength. Evidence is the verifiable support attached to a claim. Validation is the presence of independent confirmation, including third-party signals the model can cross-reference. Expertise is the demonstrable credibility of the source. Entity strength is how legibly the business and its claims are wired into the structures, like the Knowledge Graph, that AI systems use to reason about the world. Together they produce the recommendation confidence a model requires before it will cite you.

Why Aggregators and Forums Appear So Often

A recurring observation in citation studies helps explain the model's preferences. Across AI systems, a small set of domains appears disproportionately: community platforms such as Reddit, video on YouTube, reference sources such as Wikipedia and, in some categories, fan wikis, and professional networks such as LinkedIn. In Google's AI surfaces specifically, encyclopedic and community sources rank prominently among cited domains. These sites are not winning on conventional authority alone. They win because they offer clearly structured, consensus-bearing, frequently corroborated text that a model can ground in with low risk.

It is important not to overread the concentration. The same studies find that even the single most-cited domain on a given platform rarely accounts for more than a small percentage of total citations, and that the large majority of citations are spread across a long tail of thousands of individual domains. The visible giants set the tone, but most citations go elsewhere, to specific pages that answered a specific sub-query well. That long tail is where a focused business actually competes.

The takeaway is not to imitate a forum. It is to notice what those sources have in common that an ordinary business page often lacks: a direct answer, visible corroboration, a clear entity behind the content, and a structure a machine can read at a glance. Those are pillar properties, available to any site willing to build them.

What This Means for Practice

If the goal is to be picked as a source, the work is to make a business legible and trustworthy to a retrieval-and-grounding system, not to chase a ranking position. That means writing pages that answer the sub-questions around a topic, not only its single keyword, because fan-out rewards coverage. It means stating answers directly and early, because passage extraction rewards self-contained clarity. And it means surrounding every claim with verifiable evidence and recognized entities, because the model cites what it can corroborate.

Consistency does quiet but decisive work across this pipeline. When a business describes itself, its products, and its facts the same way across its own pages, its profiles, and the third-party sources that mention it, the model encounters agreement rather than contradiction, and agreement raises confidence. When the same facts appear in different forms in different places, the model has reason to hesitate, and hesitation suppresses citation. Consistency is therefore not housekeeping. It is a direct input to whether you are chosen.

None of this requires manipulation, and none of it rewards it. The mechanisms reward content that is genuinely clear, genuinely supported, and genuinely tied to a real and recognizable organization. That alignment between honest substance and machine legibility is the core premise of AIO, and it is why AIOFacts treats AIO as distinct from SEO rather than merely extending it.

From Ranking to Recommendation

The shift captured here is larger than one feature. Search is moving from returning documents to recommending answers, and the unit of victory is moving from position to citation. An AI Overview is an early, well-documented instance of a pattern now common to ChatGPT, Gemini, Perplexity, and other answer engines: decompose the question, retrieve broadly, ground the answer in sources the system trusts, and cite the few it can stand behind.

Because the pattern is shared, the optimization response generalizes. AIOFacts uses AIO as the umbrella term for being understood and recommended by AI systems. GEO, Generative Engine Optimization, and AEO, Answer Engine Optimization, are narrower terms focused on particular surfaces, but the underlying requirement is the same across all of them: build recommendation confidence. A business that earns it in Google's Overviews tends to earn it elsewhere, because every answer engine is asking a version of the same question about whether you are safe to cite.

That is the durable point. The systems will keep changing, models will be replaced, and thresholds will be retuned. What persists is the demand for sources that are clear, consistent, evidenced, validated, expert, accessible, and entity-strong. Pages that meet that standard are the ones AI Overviews, and the answer engines that follow them, choose.