The Delulu Blog

How AI Overviews Select and Combine Sources

By Eric Barker · August 2026 SEO & Search

"We outrank them on that exact phrase. Why did the AI Overview cite them and not us?" It's a fair question, and it has a real, documented answer — it's just not the answer classic rank-tracking intuition expects. AI Overviews don't work like a single ranked list responding to a single typed query. Here's the actual mechanism, as Google has described it.

Step one: query fan-out

Google's own documentation describes AI Overviews and AI Mode using a technique it calls "query fan-out" — a single query gets expanded into multiple related searches across subtopics before a response is generated. Type one question, and the system is quietly running several related searches behind it, not just the exact phrase you typed.

This is the first reason rank-for-the-exact-phrase intuition breaks down: your visibility for a given AI answer doesn't map one-to-one to your position for the literal query someone typed. You're being evaluated against a wider, related set of searches happening underneath the one visible question.

Concretely, a query like "best CRM for a small landscaping business" might fan out into related searches for CRM pricing comparisons, landscaping-industry software reviews, and general small-business CRM buying guides — several distinct searches, each pulling its own candidate pages, before anything gets synthesized into one answer. A page that ranks #1 for the exact typed phrase but says nothing relevant to those adjacent searches is competing on only one of several fronts that actually determine the final answer.

Step two: candidate selection

Which pages become candidates in the first place? Crawled, indexed, ranking reasonably well in the standard organic index — AI Overviews draw from that same index, not a separate ranking system. The best available independent evidence for how selection actually skews in practice comes from Ahrefs' analysis of 1.9 million citations across 1 million AI Overviews: 76.1% of cited pages already rank in the organic top 10, and 86% rank somewhere in the top 100. The median rank of the first-listed citation is position 2.

Read plainly: organic ranking strength is a strong input into candidacy, not a guarantee and not the only input, but a heavy thumb on the scale. A page with no organic footing at all has little shot at candidacy, let alone citation.

The practical implication is blunt: there's no AI-citation shortcut that bypasses ordinary organic ranking. A page with no meaningful presence in the standard index isn't being quietly considered through some separate AI-specific channel — it's simply not a candidate. This is the same honest point the rest of this cluster makes from different angles: the foundation is still the foundation. Nothing about AI Overviews replaces the work of being crawlable, indexed, and reasonably well-ranked in the first place.

Step three: synthesis and grounding

Once candidates are selected, Google's systems generate the actual answer using a technique it describes, in its own words, as retrieval-augmented generation, "also known as grounding" — a method "used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index." From there, per the same documentation, "systems then review the specific information from those retrieved pages to generate a more reliable and helpful response, showing prominent, clickable links to relevant web pages that support the information in the response."

That's the mechanism behind why citations appear at all in an AI Overview — it isn't a courtesy link bolted onto a generated answer after the fact. The citation is the record of what the answer was actually grounded in, which is also why a page that never got retrieved in the first place can't end up cited, no matter how good its content is.

This is also the specific thing that separates a grounded answer from an ungrounded one — worth naming because it's easy to conflate the two. A system with no retrieval step is generating an answer purely from its trained-in knowledge, with no live source to check against and nothing concrete to cite. A grounded system is doing something structurally different: retrieving actual current pages first, then generating an answer constrained by what those pages say. The citations aren't decoration bolted onto a confident-sounding paragraph — they're the mechanism's own audit trail, and a business's job is to be one of the retrieved pages that trail can point to.

What determines which sources get combined versus excluded

Here's the honest limit of what's publicly documented, stated plainly rather than papered over: Google has said AI-generated answers surface more diverse links than classic search results typically do, but it has not published a full ranking formula for the combination step specifically — which candidate, among several plausible ones, actually gets synthesized into the final answer, and why one gets included while another equally-relevant one doesn't. This is the one genuine gap in an otherwise unusually forthcoming disclosure. Treat any claim describing this step as a fully solved, fully transparent algorithm with real skepticism — Google's own documentation doesn't go that far.

Picture two pages, both ranking well, both crawlable, both plausibly relevant to the same fanned-out query set — and only one ends up in the final synthesized answer. Google's public documentation doesn't give a formula that predicts which one wins that specific comparison. That's not an oversight in this article; it's an honest description of where the publicly available mechanism actually ends. Anyone claiming to have reverse-engineered the exact combination formula is claiming more certainty than Google's own documentation supports.

Why this isn't a fixed algorithm

Google states directly that crawling, indexing, and serving are never guaranteed outcomes, and that these systems are actively iterated over time — the mechanism described above reflects how AI Overviews work as of the date this article was written (August 2026), not a permanent specification. Query fan-out, the weight given to organic rank, and the grounding process itself are all subject to change as Google continues developing the underlying models and techniques. Anyone citing this mechanism as fixed, timeless fact — including this article, read a year from now without checking for updates — is making a claim Google itself doesn't make.

The practical takeaway isn't to chase the current mechanism's exact parameters — by the time any specific weighting is confirmed, it may already be out of date. It's to keep investing in the parts of this that don't depend on any one snapshot of the algorithm: genuine organic ranking strength, clean crawlability, and content substantive enough to be worth grounding an answer in. Those hold up regardless of which specific technique Google is using this quarter.

What this means practically

None of this changes what actually earns visibility — it explains the mechanism behind why that work translates into AI citation specifically. A page still needs to be crawlable, indexed, and ranking reasonably well to become a candidate at all; content still needs to be genuinely useful and clearly structured to be the piece grounding actually pulls from once it's a candidate. The mechanism adds detail to that picture — it doesn't replace it.

In practice, that means the same short list shows up again and again across this cluster, because it's genuinely the foundation underneath every mechanism described above: be crawlable and indexed, rank reasonably well organically, keep entity information consistent, and write content specific enough to be worth grounding an answer in. None of that is new advice dressed up as an AI-era discovery — it's the same foundation, with a documented mechanism now explaining exactly why it still matters.