Google AI Overviews: how they work and who gets cited

An AI Overview is the block of AI-generated text Google places above the traditional results for queries it judges answerable by synthesis. Instead of ten links, the searcher gets a few paragraphs assembled by Google's models from multiple web pages, with source links attached along the side or woven into the text. It grew out of Google's Search Generative Experience experiments and has been expanding across query types and countries ever since, moving from a labs curiosity to the first thing a large share of searchers read.

The strategic weight of that placement is hard to overstate. Search results have always had a hierarchy of attention, and the Overview now occupies the top of it: it is the default first read for the queries it covers, and for many searchers it is the only read. McKinsey's October 2025 research found 50 percent of consumers already use AI-powered search intentionally as a primary way to find information and make buying decisions, and the Aeolyft 2026 U.S. Search Trends Report found 58 percent of Americans use AI weekly. The Overview is where those habits and Google's distribution meet.

For brands, the right mental model is that a new answer layer has been inserted between your pages and your buyers, on the world's largest search engine. Everything below is about how that layer selects its material and what the selection means for visibility and clicks.

Ask how often AI Overviews trigger and you get answers that disagree by a factor of three, so let us present the spread rather than a false precision. Some trackers report AI Overviews on roughly 48 percent of tracked queries by early 2026, up about 58 percent year over year. Other panels report substantially lower figures, in the 13 to 25 percent range. Both camps are measuring honestly; they are measuring different things.

The gap comes from query composition. Overviews fire disproportionately on informational, question-shaped and research queries, and far less on navigational or purely transactional ones. A tracker whose panel skews toward long-tail question queries will report high prevalence; a panel weighted toward brand names and short commercial heads will report low. Neither number is "the" truth, and any vendor quoting one without naming the panel is selling confidence rather than measurement.

What matters for planning is your own query mix. The buying questions that precede considered purchases, comparisons, "best X for Y", how-to and evaluation queries, sit squarely in the Overview-heavy zone. If your funnel begins with those searches, your effective Overview exposure is closer to the high estimates regardless of what the blended average says, and the trend in every panel points the same direction: up.

The single most important mechanical fact about AI Overviews is where their material comes from: Google's AI features cite from Google's organic index. There is no separate AI index, no side door, no parallel submission process. When an Overview is generated, Google's models draw on pages the regular crawler has fetched, the regular index has stored, and the regular ranking systems consider retrievable for the query and its variants. Being indexed and rankable is the precondition for everything else.

That fact should genuinely comfort SEO teams, because it means fifteen years of accumulated discipline still applies. Crawlability, internal linking, indexing hygiene, content that earns rankings: all of it remains the entry ticket. The synthesis layer then adds its own selection on top, choosing which retrievable pages actually feed the generated text, and that second selection favors a recognizable profile: pages that state answers directly, carry evidence, and offer passages a model can lift cleanly into a summary.

On that second layer, the experimental evidence from generative engines generally applies. The Princeton and Georgia Tech GEO study, the first academic treatment of generative engine optimization, found that adding quotations, statistics and cited sources boosted visibility in generative responses by up to roughly 40 percent, while keyword stuffing landed near the bottom. A page built as evidence rather than assertion is more liftable, and liftable is what the Overview generator is shopping for.

Because Overviews draw on the organic index, cited pages tend to be pages that already rank somewhere for the query's territory, though the citation list is not a copy of the top ten: Google's models routinely cite pages from deeper positions when those pages answer the specific sub-question the Overview is synthesizing. That is the opening for focused content. A page that answers one precise question completely can be cited above bigger domains that answer it vaguely, which is the same long-tail dynamic the citation studies found across AI engines.

Now the finding that upsets checklists. Ahrefs ran the clean experiment in May 2026, adding JSON-LD schema to 1,885 pages, and found no measurable citation lift for ChatGPT or AI Mode, and a statistically significant decline in AI Overviews citations. The honest caveats: the studied pages were already heavily cited, and schema may still matter for initial parsing and discovery. Bing's Fabrice Canel has said schema helps LLMs understand content, and SE Ranking found high structured-data incidence on cited pages, about 65 percent for Google AI Mode, though that is correlation. The industry consensus is genuinely conflicted, and we lay out both sides in schema markup for AI search.

Our reading: schema is cheap and fine to ship, and it is not a lever for Overview citations. The lever is the content itself, answer-first structure, evidence density and indexing, plus presence on the third-party pages Google already ranks for your buying questions, since an Overview citing a comparison article that features you is a win you did not have to host.

The click data explains why AI Overviews dominate every SEO conversation. 2026 tracking studies measured organic click-through dropping about 61 percent when an AI Overview is present: the summary answers the question, and the majority of clicks that would have flowed to the results below simply do not happen. For publishers whose economics depend on informational traffic, that is a structural break, and no amount of optimization brings those sessions back.

But the same studies found the redistribution inside the shrunken pool: brands cited inside the AI Overview see roughly 35 percent higher click-through. The searchers who do click, click on what the Overview endorsed. So the Overview performs a double sorting: it removes casual clicks entirely, and it concentrates the remaining, more deliberate clicks on its cited sources. Being cited is now worth more than ranking positions the Overview has buried. We analyze the full zero-click picture in AI Overviews and your clicks.

The deeper shift is that a growing share of brand influence now happens inside the answer, with no click at all. A buyer whose first three research queries all returned Overviews naming the same two vendors arrives at vendor websites with the shortlist pre-formed. Winning that pre-click layer is a different discipline from winning the click, and it is the discipline this whole blog exists to teach; the practical playbook for this surface specifically lives in how to show up in Google AI Overviews.

Google now operates three distinct AI answer surfaces, and conflating them leads to muddled strategy. AI Overviews augment the classic results page. AI Mode is a separate, fully conversational search experience where the entire session is chat-shaped. And the Gemini app is a general assistant, which surpassed 750 million monthly users, that answers with or without live search. They share models and index plumbing, and they behave differently as citation surfaces.

All three draw on Google's index for live material; they differ in interface, session shape and how their citations can be measured.

The good news is that the underlying work overlaps heavily: indexed, rankable, answer-first, evidence-dense pages feed all three. The measurement stories differ sharply, though, which is the subject of the next section. For the deeper dives, see Google AI Mode explained and Gemini and your brand.

It is tempting to treat AI Overviews as the whole AI visibility problem because they sit on the biggest search engine, and the numbers argue for a wider frame. ChatGPT reached roughly 900 million weekly active users in early 2026, about double a year earlier, and G2 found 51 percent of B2B software buyers now start research with an AI chatbot more often than with Google, up from 36 percent seven months prior. One 2026 estimate puts AI assistant sessions at around 45 billion worldwide, equal to roughly 56 percent of global search engine volume. The Overview is one front. The war is multi-engine.

The encouraging part is how much the fronts share. The page profile that earns Overview citations, indexed, answer-first, dense with sourced evidence, is the same profile the GEO research validated for generative engines broadly, and the third-party mentions that feed Google's index feed the other engines' retrieval too. A brand that builds for the answer layer as a whole collects wins across surfaces, rather than optimizing one interface at a time.

The sobering part is the baseline: G2 found 51 percent of B2B tech brands have zero citations across ChatGPT, Perplexity and Gemini. Half the market has not entered the answer layer anywhere, which also means the brands that move early are competing against absence. If you are choosing where to start, start where measurement is honest and effort is verifiable, then let the same content carry you into the Overview.

Here is the uncomfortable truth about every AI Overviews tracking dashboard on the market: Google provides no official API for AI Overviews. Every tool reporting Overview presence, citations or share of voice is scraping search results, against terms of service, through proxies, for an experience Google personalizes by location, history and ongoing experiments, and rotates continuously. Scraped coverage breaks silently: when the scraper fails or the layout shifts, the dashboard keeps rendering numbers that no longer mean what they meant.

This is why Reachroller made the unfashionable choice to leave Google AI Overviews out of its engine lineup until an official API exists. Reachroller tracks ChatGPT, Claude, Gemini, Perplexity and Grok through official APIs only, ChatGPT live today and the others built and rolling out, because a score you cannot audit back to a reliably collected raw answer is decoration. Gemini matters here: it shares Google's models and index plumbing and is trackable through an official API, making it the honest proxy for how Google's AI describes your brand.

In the meantime, measure Overviews the honest manual way: pick your core buying queries, search them repeatedly across days in a clean browser profile, record whether an Overview appears, whether you are cited, and who is. Treat any single observation as one sample from a distribution, and look only at trends. It is imprecise, and imprecise-but-honest beats precise-looking-and-broken.

The response to AI Overviews compresses into three moves. First, protect the precondition: keep every page you care about crawlable, indexed and genuinely rankable, because the Overview only cites what the organic index holds. Second, restructure your money pages around liftable answers: the question in the heading, the complete answer in the first paragraph, evidence throughout. Third, extend the same thinking to the third-party pages Google ranks for your buying questions, because Overviews cite comparison articles, reviews and community threads, and a citation naming you on someone else's page is still your win.

Then close the loop on the surfaces where honest measurement exists. Run your buying questions through the engines with official APIs, find the ones where rivals get named and you do not, publish the fix, and verify the flip. Reachroller automates exactly that loop, from question runs to publish-ready fix pages with indexing steps to the recheck, for $29 per month on Starter with a three-day, 50-credit, no-card trial. The same evidence-dense pages that win those engines are the pages Google's Overview generator likes to lift, so the work compounds across surfaces, including the one nobody can officially measure yet.