An AI doesn't choose its sources at random, and it doesn't choose many: two mechanisms decide what makes it into a generated answer. The first is the retrieval of documents relevant to the question asked — what the field calls retrieval. The second, documented by Google since 15 May 2026, is query fan-out: a complex question is broken down into several sub-questions, each processed separately, before being synthesized into the final answer. Becoming a cited source isn't a matter of some technical trick; it's about being a strong candidate at every one of these steps.
What happens between the question asked and the answer given?
A generative model doesn't answer solely from what it memorized during training. For a question that calls for recent, local, or specific information, it first queries a document index, keeps the passages most relevant to the question asked, then drafts an answer that draws on those passages — often with a citation of the source. This is the principle of retrieval-augmented generation, or RAG, which I implement myself in custom projects that need to stay factual and verifiable. The same principle, at the scale of a general-purpose search engine, explains why ChatGPT, Perplexity, or Google's generative features can cite a page published the previous week: the model didn't learn it during training, it went looking for it at the moment of answering.
What is query fan-out, and why does it change the game?
On 15 May 2026, Google Search Central publicly documented the mechanism structuring its generative search features: the core ranking and quality systems of classic search, to which generation techniques are added — retrieval-augmented generation and query fan-out, which breaks a query down into sub-queries synthesized into the final answer[1].
In practice: an executive asking who to recommend for managing their company's tax affairs in Valais isn't asking one question — without realizing it, they're asking several. The system can break this down into a sub-question about firms active in the region, another about their reputation, another about their scope of expertise — and find different sources for each, before recombining them. Being cited once isn't enough, then: you have to survive several rounds of retrieval, on angles the original question doesn't reveal.
Google I/O, four days later, confirmed the scale of this shift with the expansion of AI Mode and the first research agents capable of following a topic and comparing options across multiple criteria[2] — a mechanic built on the same decomposition principle. One honest caveat is needed here: to date, Google is the only one to have documented this mechanism in detail. Other generative environments — ChatGPT, Claude, Perplexity, Mistral — likely rely on broadly similar retrieval logic, though their publishers haven't published the exact architecture.
Why does a generated answer cite only three or four sources?
A classic results list tolerates mediocrity in seventh position: the link still exists, a trickle of residual traffic still reaches it. A generated answer doesn't tolerate this dilution. It selects a small number of sources, recombines them, and anything that fails to meet a sufficient bar of clarity and reliability simply doesn't appear. This isn't an arbitrary tightening of the rules; it's the mechanical consequence of a system that has to choose, rather than rank. The severity has increased — not the nature of the game.
What makes one source get picked over another?
Three qualities keep recurring in what the GEO Score™ grid observes, theme after theme — and they're the same ones that shape how I design a site built to be cited.
Sharp answers. A model more readily retains a passage that directly answers a precise question than a text that circles the subject without ever settling it. A page that opens with the answer, then develops it, lends itself better to extraction than a page that builds its argument over a thousand words before concluding.
Authority built over time. Dated content, signed by an identifiable author, consistent with what the same company was claiming a year earlier, carries more weight than anonymous content or content that contradicts itself from one page to the next. This is what the GEO Score™ grid observes under the theme of temporal stability: a citation that keeps shifting from one measurement to the next signals an authority that hasn't yet settled.
Entity coherence. Before deciding to cite you, a system first has to resolve who you are — a specific company, at a specific address, with a specific scope of activity — rather than one variant among several contradictory versions of your identity found across the web. This is precisely the subject of What Do AIs Already Know About Your Company?: this coherence is built before content even enters the picture.
The link to the GEO Score™'s seven themes
I built the GEO Score™ to observe, not to guess. The grid covers seven themes — citation presence, editorial prominence, citation context, factual accuracy, source grounding, cross-model consistency, and temporal stability — and each corresponds to a stage in the mechanics described above. Citation presence simply checks whether retrieval found you. Source grounding looks at what the model relied on to cite you — the outcome of fan-out, in short. Cross-model consistency tests whether your identity holds up moving from one retrieval system to another, each with its own index and its own data freshness. This grid is detailed in full in MCVA Cahier No. 1, and it's what structures the GEO Score™ pre-audit.
What you don't control
Part of this mechanism lies entirely beyond the reach of the company being measured, and that needs to be said plainly. You don't decide how a system breaks your question into sub-queries. You have no visibility into the ranking algorithm that selects the retained passages — every publisher keeps it proprietary, and rightly so. No optimization, however careful, guarantees a citation: I write this consistently, because anyone promising otherwise is overselling. And index freshness varies from one model to another — one system may cite up-to-date information while another still relies on an older version of your company, which explains part of the discrepancies observed between ChatGPT, Claude, Perplexity, Gemini, and Mistral. What you control is the quality of the candidate you present at each stage. Not the draw that results from it on any given query.
Becoming citable, then, isn't about cracking an algorithm — it's about removing the reasons a system would hesitate to select you: a site machines can read without ambiguity is precisely the purpose of sites designed to be cited, backed by honest structured data. This mechanism is just one piece of the broader question I address in What Should a Swiss SME Do About AI in 2026?
Key takeaways
— A generated answer relies on two mechanisms: the retrieval of relevant sources (RAG) and, at Google since May 2026, query fan-out, which breaks a question down into several sub-queries. — A synthesis cites few sources because it selects rather than ranks: sharp answers, authority built over time, and entity coherence make the difference. — No citation is guaranteed: you control the quality of the candidate presented, not the draw that results from it on any given query.
FAQ
Is query fan-out used by all generative models? Only Google has publicly documented it, as part of its generative search features. Other environments likely rely on similar retrieval logic, without a publicly detailed architecture — caution is warranted here.
Does paid advertising improve my chances of being cited? No, not according to the documented mechanics: the selection of sources cited in a generated answer draws on organic ranking and quality systems, not a separate advertising space.
Can content be cited without ranking on page one of Google? Yes. The selection criterion for citation isn't identical to the organic ranking criterion; the two correlate, imperfectly. Content with rigorous sourcing and a clear authorial voice can be selected without dominating the ranking.
How can I know whether my company survives these filters? By measuring it directly against the models, rather than assuming it from your Google position. That's the purpose of the GEO Score™ and its pre-audit.
Where do you stand? The GEO Score™ pre-audit measures, free of charge, where your company appears — and doesn't appear — in the answers from ChatGPT, Claude, Perplexity, Gemini, and Mistral. A quantified answer within a few days. Request a free pre-audit
Sources
[1] Google Search Central, Optimizing your website for generative AI features on Google Search, 15 May 2026. developers.google.com/search/docs/fundamentals/ai-optimization-guide [↩]
[2] Google, Google Search's I/O 2026 updates: AI agents and more, 19 May 2026. blog.google/products-and-platforms/products/search/search-io-2026/ [↩]
Jérôme Deshaie is CEO and founder of MCVA Consulting SA, an augmented agency based in Valais. Fifteen years serving major international brands, now working directly with Swiss SMEs. Background.