A page can sit at the top of Google’s classic results and still be missing from the AI answer printed above them. And Google is not one AI answer anymore. It is three: the AI Overview box on the results page, the conversational AI Mode tab, and the standalone Gemini app. Each is a different system that pulls from different sources, so “rank in Google’s AI” is at least three separate jobs. This is the Google entry in our per-platform set, after the ChatGPT and Perplexity guides. We lean on the largest public studies for the hard claims, add our own measurements where they are solid, and say plainly where the data is thin.
Three Googles, not one
Before any tactic, get the map right. These are three products, and people lose months optimizing for one while measuring another.
| Surface | What it is | Where it reads from |
|---|---|---|
| AI Overviews | The summarized answer box on the results page. Appears for some queries, not all. | The live Google results page |
| AI Mode | A full conversational tab inside Search, with follow-up questions. Longer answers. | The live Google results page |
| Gemini | The standalone Gemini model and app. | The model, grounded on Google Search when it decides to |
They reach similar conclusions and cite different sources. Ahrefs compared 730,000 AI Mode and AI Overview response pairs and found only 13.7 percent of citations overlapped, even though the answers were 86 percent semantically similar. Victorious, on 1,540 queries, found roughly a third of citations carried over. So winning the overview box does not win the AI Mode tab, and the Gemini app is different again. The single most useful thing in this whole post is to stop treating “Google AI” as one target.
How Google decides what to cite
Two mechanics matter more than any individual tactic.
The first is query fan-out. Instead of answering your literal query, Google’s AI breaks it into a set of narrower sub-questions, runs each as its own search, and fuses the results; the pages that show up across several of those sub-results are the ones that get pulled into the answer. Google’s own Gemini grounding documentation describes the shape: a classifier decides whether a search would help, the model issues one or more queries, and it returns grounding chunks (the source pages) mapped to the specific sentences they support. The practical read: you do not need to dominate one broad head term. You need to be the best answer to one specific sub-question that the fan-out happens to ask.
The second is that ranking and citation are coming apart. Ahrefs found the share of AI Overview citations coming from the top 10 organic results fell from 76 percent in mid-2025 to 38 percent by early 2026, with the rest spread across positions 11 to 100 and beyond. Ranking near the top still helps, but it no longer decides the citation. A page on your exact sub-topic can get pulled in from page three while the number-one result is skipped.
What differs across the three surfaces
We track all three Google surfaces separately, which is the only way to see this. One honest note on scope first: the numbers below come from a small, category-skewed slice of the brands we monitor, weighted toward how-to-heavy categories like WordPress and ecommerce tooling, and our AI Overviews data is only a few weeks deep. So read the absolute levels as sample-specific. The cross-surface contrast is the durable part, because it compares the same brands across surfaces.
For the same brands, the two surfaces that read the live results page (AI Overviews and AI Mode) leaned on video and community sources. YouTube ran around 11 to 13 percent of their citations in our sample. The Gemini model, answering the same kinds of questions, pulled YouTube closer to 2 percent. Same brands, same categories, very different source mix, so the gap is the surface, not the brand. The takeaway is concrete: a strong YouTube presence can earn you AI Overviews and AI Mode and do almost nothing for the Gemini app. That alone kills the idea of one “Google AI” content plan.
One more pattern worth knowing, and it changes how you measure. Across our Google measurements, the answer often names a brand in the prose without attaching a clickable source, where Perplexity and Claude almost always attach one. So if you judge your Google visibility by clickable citations alone, you undercount how often Google actually mentions you. Count the named mentions too, not just the links.
How to get cited: the on-page work
This part is mostly shared across the three surfaces, and it lines up with the one controlled study the field has. The Princeton-led Generative Engine Optimization paper (KDD 2024) tested specific edits and found that adding statistics, citing authoritative sources, and adding direct quotations raised visibility in AI answers by up to roughly 40 percent, while keyword stuffing did worse than the baseline. Build from that.
- Answer first. Open the page, and each major section, with a direct 40 to 60 word answer before any backstory. AI surfaces lift these capsules almost verbatim, so a buried answer is an uncited one.
- Structure for extraction. Listicles, numbered steps, and comparison tables get pulled more than long prose, because a model can take one row or one step without grabbing the whole article.
- Be specific and sourced. Replace vague claims with concrete, cited numbers, and make definitive statements rather than hedged ones. This is the Princeton result in practice.
- Mark it up and keep it current. Schema helps Google understand and index your page (the engines themselves read the markup as plain text), and recently updated content gets preference on AI surfaces. We read the structured data the most-cited pages actually use in our schema for AI breakdown.
- Be a recognizable entity. Google leans on the brands its knowledge graph already associates with a topic, so consistent coverage under your own name compounds.
Then the surface-specific part. Because AI Overviews and AI Mode read the results page, ordinary Google SEO still feeds them: be crawlable, earn the ranking, and win the clean answer that fan-out can lift. A genuinely useful YouTube video is a real lever on those two, as our own source mix shows. The Gemini app rewards the same clear, sourced content but does not lean on video, so do not expect a YouTube push to move it.
How to track whether it worked
Two rules follow directly from the data above. Track each surface on its own, because an AI Overview win is not an AI Mode win and neither is a Gemini win. And count named mentions, not just clickable links, or you will undercount Google specifically. A per-surface, presence-aware view is the gap we built SurfacedBy to close, and the broader case for measuring each engine separately is in our study of citation overlap across five engines.
Two related posts go deeper. Getting into the index each assistant reads is its own problem, and Google’s surfaces all read Google, which the index map walks through. And the on-page work above is the same craft we broke down in how to write content AI platforms cite.
The honest limits
It is worth being clear about which claims rest on what. The three-surfaces split, the fan-out mechanic, and the ranking decline come from large public studies and Google’s own documentation, not from us. Our first-party contribution is the cross-surface source-mix contrast, and it is drawn from a small, category-skewed sample with only a few weeks of AI Overviews data, so treat the exact percentages as directional. The mechanism, though, is solid enough to act on now: Google is three AI surfaces, they cite differently, ranking number one no longer guarantees a citation, and the work that earns one is the specific, sourced, answer-first page. Write that page, then measure each surface on its own.
