AI Overviews now sit above the blue links for a huge share of queries. Here is exactly how sources get selected — and the playbook to become one.
Google’s AI Overview is the synthesised answer that now sits above the blue links for a large and growing share of searches. It is not a snippet you win with a meta tag — it is a generated answer assembled from several sources, and the only thing that matters is whether your page is one of the ones it pulls from and names. This is the complete playbook: how Overviews are built, exactly how they choose sources, the five levers that get you cited, the mistakes that keep good pages out, how to measure your presence, and a 30-day plan to start showing up.
An AI Overview (AIO) is a generative summary Google places at the top of the results for queries where it believes a synthesised answer serves the searcher better than a list of links. It reads the query, gathers information from multiple pages in its index, writes a multi-sentence answer, and attaches a small set of citations you can expand. It is powered by a version of Gemini and grounded in Google’s live index — which is precisely why classic SEO signals still matter, yet are no longer sufficient on their own.
The mental shift you have to make is this: an Overview is not a ranking position you can occupy. It is a composition. Google is not asking “which single page is best?” and promoting it. It is asking “what is the best answer to this question, and which passages, from which pages, should I stitch together to write it?” Your entire job changes from producing the best page to authoring the best passages — the specific, liftable chunks worth stitching into the answer. Everything else in this guide follows from that one idea.
For two decades, SEO had one north star: rank number one, earn the click. AI Overviews quietly tear up that contract. The answer is now delivered on the results page itself, and the searcher frequently never leaves Google at all. That reframes the entire objective of organic marketing — away from earning a visit and toward earning a mention inside the answer the searcher actually reads. The destination is no longer your website; it is the sentence Google writes about your category.
This is not a passing experiment you can wait out. Overviews have rolled out across major markets and keep expanding into more query types — crucially including commercial, comparison, and product-research queries, which is exactly where buying decisions and revenue live. Google’s AI Mode, a fully conversational search surface, extends the same logic into multi-turn dialogue. Brands that dismiss this as “a feature” are making the single most expensive strategic error available to them right now, because the window to build citation authority is open widest before everyone else notices.
To optimise for Overviews, you must understand the pipeline that builds them. It runs in three stages, and each stage is a distinct place you can win or lose. Miss any one of them and you are invisible regardless of how good the other two are.
The system rarely answers your literal query. It decomposes it into a fan of related sub-questions and retrieves for each independently. A search like “best running shoes for flat feet” silently becomes a cluster: what causes flat-foot overpronation, which shoe features counteract it, which specific models experts recommend, what reviewers say about durability, and what the typical price range is. This is the mechanical reason depth and topical coverage win. A page that answers only the head term competes for one narrow slice of the fan; a comprehensive resource that addresses every sub-question competes across the whole thing, multiplying its chances of being pulled in somewhere.
For each sub-question, Google retrieves candidate passages from its index — not whole pages, but the specific chunks most relevant to that sub-question. Relevance and authority get you into the candidate pool; that is the price of entry. But the decisive factor at this stage is extractability: does a clean, self-contained passage exist that answers the sub-question directly, without the model having to rephrase it, infer around it, or risk being wrong? A page can be authoritative and still lose here if its answer is buried, hedged, or spread across three paragraphs the model has to reassemble.
Finally, the model composes a coherent answer from the strongest candidate passages and attaches citations. It favours passages that are unambiguous, corroborated by other sources, and safe to quote. When two pages assert the same fact but one states it as a crisp, sourced claim and the other buries it in cautious throat-clearing, the crisp one gets cited. The model is optimising for a confident, correct, defensible answer — and it rewards the sources that make that easiest.
The uncomfortable numbers make the stakes concrete. A Pew Research Center analysis of nearly 69,000 real searches found that when an AI summary appeared, users clicked a traditional result just 8% of the time — versus 15% when no summary was present. Only about 1% clicked a link inside the Overview, and roughly a quarter of sessions ended on the results page without any click at all.
Read that carefully, because the wrong conclusion is “abandon Google” and the right one is “the value moved.” The click is shrinking; the citation is now the prize. Being named in the Overview is the first thing the searcher reads — it frames the entire decision even when no click follows, the same way a recommendation from a trusted friend shapes a purchase before you ever visit a store. The brands that win the next five years treat citation share as the primary metric and blue-link position as a secondary, supporting one.
The single highest-leverage change you can make is to lead with the answer. Military and intelligence writers call it BLUF — Bottom Line Up Front. For every question your page targets, the first 40–60 words of that section should be a complete, standalone answer, delivered before any context, history, or brand preamble. If a machine can lift your opening sentence and be correct, you are a candidate for citation. If your opening is “In today’s fast-paced digital landscape, businesses must…”, you have disqualified yourself in the first line.
Answer-first does not mean shallow or short. Give the direct answer, then go as deep as the topic deserves underneath it — for the readers and the sub-questions that need the full treatment. You are deliberately writing for two audiences in the same passage: the machine that needs one liftable, correct claim near the top, and the human who wants the complete, nuanced picture below it. Done well, the two reinforce each other rather than competing.
Query: “how to descale a coffee machine.”
Loses: a page that opens with the brand’s founding story and a paragraph on the history of espresso before eventually, three scrolls down, listing the actual steps.
Wins: a page that opens with “To descale a coffee machine, run a 1:1 solution of white vinegar and water through a full brew cycle, then run two cycles of clean water to rinse. Repeat monthly in hard-water areas.” — 44 words, complete, liftable. Google pulls it straight into the Overview and names the source. Identical facts, opposite outcome, decided entirely by where the answer sits on the page.
Overviews prefer content they can lift whole, which means writing in the formats machines parse most reliably. This is not about dumbing content down; it is about removing the friction between your knowledge and the model’s ability to quote it.
The principle beneath all five is the same: reduce the work the model must do to extract a correct answer from your page. Every ambiguity you remove is one more reason to cite you rather than the competitor whose passage needs interpreting.
The academic research on generative engines is unusually actionable here. The GEO study (Aggarwal et al., presented at KDD ’24) tested empirically what actually increases how often a passage gets cited and found that adding relevant statistics, quotations from credible sources, and clear citations measurably raised citation rates — in some tested categories by 30–40%. This is one of the few places in this discipline where you can point to controlled evidence rather than folklore.
The practical instruction is blunt: stop making unsupported claims. Replace “this approach is highly effective” with “a 2025 study found this reduced X by 27%.” Attribute data to named sources with a year attached. Add a specific number, a date, a named authority to passages you want quoted. Evidence does two jobs at once — it makes the passage more trustworthy to a risk-averse model, and more useful to a human who wants proof, not adjectives. Both are direct reasons to cite. The most citable asset of all is your own original data: a survey, a benchmark, or a documented result no competitor can reproduce.
Models are risk-averse by design. They prefer to cite sources they can stand behind, because a wrong or dangerous answer is a serious, reputation-damaging failure for the engine. That is exactly why Google’s E-E-A-T framework — Experience, Expertise, Authoritativeness, Trustworthiness — is effectively a citation framework in the AI era, not just a ranking one.
Signal it concretely rather than claiming it. Use real, named authors with genuine credentials and substantive bios. Make first-hand experience visible in the writing — specifics only a practitioner would know. Cite primary sources. And, most importantly, earn corroboration from the wider web: a claim that appears only on your own site is a liability the model can’t verify, while the same claim echoed across trusted third-party sources becomes a fact it can safely repeat with your name attached. For Your-Money-or-Your-Life topics like health, finance, and safety, this bar is highest and genuinely non-negotiable — the engines will simply refuse to cite sources they can’t trust on those subjects.
None of the above matters if Google cannot see your content cleanly. Technical eligibility is the unglamorous entry ticket, and it is where a surprising number of otherwise-excellent pages quietly fail.
Beneath retrieval sits the knowledge graph: Google’s structured model of the world as entities — brands, people, products — and the relationships between them. To be cited with confidence, your brand needs to be a clearly-defined entity the model recognises, not an ambiguous string of characters it has to guess about. This is the layer most teams never think about, and it silently determines whether your citations are attributed to you or to someone with a similar name.
Build entity clarity with a consistent name and description across every property, Organization and Person schema, unambiguous author identities, and presence in the authoritative references models lean on — Wikipedia and Wikidata where genuinely warranted. When the model is certain who you are, it can attribute facts and citations to you correctly and repeatedly. When it isn’t, it either conflates you with another entity or leaves you out of the answer entirely to stay safe — and you never even know it happened.
Most pages that genuinely deserve to be cited but aren’t are making one or more of these specific, fixable errors:
The encouraging part: fixing these is often faster than writing new content. It is usually a restructuring, evidencing, and signalling job performed on pages that already carry authority — which means the payoff can arrive within a crawl cycle or two rather than a quarter.
You cannot manage what you don’t measure, and traditional rank tracking will actively mislead you here — you can sit at position three and be completely invisible if the Overview cites four other sources above you. The metric that matters is citation share: across the queries in your category that trigger an Overview, how often are you one of the named sources? Track it over time, per query cluster, and against your direct competitors.
Practically: build the list of your priority buyer questions, check which of them currently trigger Overviews, record who gets cited on each and why, and monitor it continuously — because Overview sources change silently and often, with no notification when you drop out. This exact volatility is what DUNkē was built to watch, so a lost citation surfaces the same week it happens rather than in a quarterly review long after the damage is done and the competitor has entrenched.
A focused month beats a vague year of good intentions. Here is a concrete starting sequence you can run against any set of priority pages:
Run this on five pages, learn precisely what moves your citation share, then scale the winning pattern across the site. Overview visibility compounds — the topical authority you build for one cluster of questions makes the next cluster meaningfully easier to win, which is why starting now, on a focused set, beats waiting for a perfect site-wide plan.
It is tempting to treat Overviews as “featured snippets 2.0,” but the difference is fundamental and worth internalising. A featured snippet lifts one passage from one page and displays it verbatim, with a link. It is extraction. An AI Overview synthesises a new answer from many passages across many pages, in the model’s own words, citing several sources. It is generation. The old game was “own the snippet.” The new game is “be one of several sources the model chooses to build its answer from.”
This matters for strategy because winning is no longer winner-take-all. In the snippet era, one page got the box and everyone else got nothing. In the Overview era, three, four, or five sources are cited for a single answer — which means there is room for more players, but also that being “the best page” is no longer enough. You have to be one of the most quotable and corroborated, because the model is assembling a committee, not crowning a winner.
Teams are often baffled when a page gets cited for one query and ignored for a near-identical one. The explanation is the fan-out. Two similar queries can decompose into different sets of sub-questions, and the Overview assembles each answer from whichever passages best serve that query’s specific fan. Your page might contain the perfect passage for one sub-question but nothing for the slightly different sub-question the sibling query generated.
The practical response is coverage. Rather than writing one page per keyword, build comprehensive resources that answer the full spread of sub-questions a topic generates — definitions, comparisons, steps, costs, pros and cons, edge cases. The more of the fan you cover with clean, liftable passages, the more queries in that neighbourhood you become eligible to be cited for. Depth is not a vanity exercise here; it is direct citation surface area.
Sometimes you audit a priority query and a competitor is already cited while you are absent. This is a solvable problem, not a dead end, because Overview sources are far less entrenched than classic rankings and change constantly. Start by reading the cited passages closely and asking what they do that you don’t — usually it is answer-first structure, a specific statistic, or corroboration you lack.
Then out-execute on exactly those dimensions: publish a cleaner, more directly-quotable answer; add stronger, more specific evidence than the incumbent; shape it into the format the query rewards; and earn a couple of third-party mentions to match or beat their corroboration. Because the model re-evaluates sources continuously and rewards the most citable passage rather than the oldest URL, a focused improvement can displace an incumbent in weeks — a pace that is simply impossible in traditional link-driven SEO.
“Ranking #1 used to mean you were the first thing read. Now the Overview is. If you’re not in the answer, you’re not in the conversation — and being cited is the only ranking that compounds.” Rahul Shrivastava · Founder, The Age’X & DUNkē
Get a free GEO audit — the same analysis behind every article here.