How to Rank in AI Overviews: I Matched 1,764 Citations Against the Organic Results to Find Out What Google Actually Uses

Every week another list of “AI Overview ranking factors” lands in my inbox. Most of them are a poll of what people in my industry believe. That is a useful thing to read, but a belief is not a measurement, and the gap between the two is where clients lose money.

So I measured it. In September 2026 I pulled the search results for 400 US queries that show an AI Overview, spread across ten industries from health and finance to forklifts and email marketing, and I matched every link inside the AI Overview against the organic results on the same page. That is 1,764 links, 1,561 of them pointing at real outside pages on 719 different websites. I also pulled a wider pool of 3,804 keywords to see where AI Overviews show up at all. Then I read what Google itself has published about how these answers get built, including a court record most SEOs have never opened.

This post is the result. It covers how an AI Overview is actually assembled, what my data says predicts a citation, what Google says on the record, and which of the popular tactics do nothing. If you want the short version: the machine that writes the overview is fed by the same ranking system you have been optimizing for since 2004, it reaches deeper into the results than it used to, and the page-level signals matter more than the domain-level ones.

How an AI Overview gets built

You cannot optimize for a system you do not understand, so start with the mechanism. Google has described it in more detail than most people realize, and it comes down to four steps.

Step one: a normal search runs first. In May 2026 Google published a guide called Optimizing your website for generative AI features on Google Search. It says the AI features “are rooted in our core Search ranking and quality systems,” and it defines the technique behind them, retrieval-augmented generation, as “relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.” In plain English: before any AI writes a word, Google runs a search, ranks the pages the way it always has, and hands the winners to a model.

The court record backs this up. In the September 2025 remedies opinion in United States v. Google, Judge Mehta’s findings of fact describe AI Overviews as a feature that “will take the search results that come back and use a[n] [LLM] to produce a summary of those results.” Finding 44 goes further: to ground its Gemini models, Google uses a system called FastSearch, which “is based on RankEmbed signals, a set of search ranking signals, and generates abbreviated, ranked web results that a model can use to produce a grounded response.” The same finding notes it “retrieves fewer documents” than a full search. So the model sees a shortlist, and the shortlist comes from ranking signals.

Google’s own patent for this feature, Generative summaries for search results (US 11,769,017, granted September 2023), spells out how that shortlist is chosen. Candidate documents are picked using “a positional ranking,” “a selection rate” (how often people click the page for that query), “a trustworthiness measure” that can be “generated based on an author thereof, a domain thereof, and/or inbound link(s) thereto,” an “overall popularity measure,” and “a freshness measure.” Read that list twice. It is rank, clicks, links, author, domain and freshness. It is an SEO checklist from 2015.

Step two: the question gets split up. This is the part that is new. Google calls it query fan-out and defines it as “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results.” Their own example: for “how to fix a lawn that’s full of weeds,” the fan-out queries might be “best herbicides for lawns,” “remove weeds without chemicals,” and “how to prevent weeds in lawn.” Each of those runs as its own search. The patent describes the same idea as pulling in documents that are “responsive to one or more related queries.” This is why pages that do not rank for the query you typed can still show up in the overview: they rank for a sub-question the model asked on your behalf.

Step three: the model writes, then links to what it can verify. The patent is explicit that the summary is rendered “with link(s) to SRD(s) that verify portion(s) of the NL based summary,” and that a link can point to “a specific portion of the SRD (e.g., a portion of the SRD that verifies a corresponding portion of the NL based summary).” A citation is not a reward for being a good website. It is a footnote attached to a specific claim, and your page earns it by containing a passage that supports that claim. The patent also says the confidence in a portion of the summary can depend on “a quantity of the SRD(s) that verify that portion.” Facts that several sources agree on get used; a lone claim is a risk the model tends to avoid.

Step four: the gate. Google’s documentation on AI features says a page “must be indexed and eligible to be shown in Google Search with a snippet” to appear as a supporting link, and “there are no additional technical requirements.” The controls that limit AI features are the same snippet controls that have existed for years: “nosnippet, data-nosnippet, max-snippet, or noindex.” The Google-Extended token that publishers block in robots.txt “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal in Google Search,” per Google’s crawler documentation. It governs Gemini apps and Vertex AI, not AI Overviews. Since August 31, 2026 there is also a new switch: a Search generative AI control in Search Console that lets a site opt out of AI Overviews and AI Mode entirely. It is on by default. Check it anyway, because the optimization guide now says a site “must be included in Search generative AI features in Search Console to be eligible.”

Here is the metaphor I use with clients. An AI Overview is a research assistant with a library card, not a librarian with opinions. It can only cite books the library already put on the front shelf for your question or for the questions it asked next. Getting on that shelf is the job, and that job has a name.

One boundary before the data: this is how Google’s features work, and it does not transfer. ChatGPT, Perplexity and the other assistants retrieve from their own systems, and Ahrefs found that only 12% of the pages they cite rank in Google’s top ten for the same prompt. I covered those separately in How to Rank in ChatGPT. This post is about the one that sits on top of Google’s own results.

What 400 AI Overviews actually cited

Now the data. The method is at the bottom of this post; here is the shape of it. I sampled 3,804 US keywords with three or more words and 100 to 20,000 monthly searches, across ten verticals, and recorded whether each had an AI Overview. From the ones that did, I drew 40 per vertical, pulled the Ahrefs SERP snapshot for each (all captured between September 1 and September 24, 2026), and matched every link in the AI Overview block against the organic results in the top ten positions of the same page. I then pulled page-level metrics for every cited page that did not appear in those results.

Where AI Overviews show up

Across the pool, 68.7% of keywords had an AI Overview. That number is high because the pool is skewed toward longer, question-shaped queries, which is what most of the web is trying to rank for. The spread between industries is the useful part.

68.7% of 3,804 keywords showed an AI Overview. Health 97.4%, home services 42.4%.

Two patterns inside those numbers. Queries phrased as questions (how, what, why, can, does, is) triggered an overview 86.4% of the time against 64.7% for everything else. Queries with local intent triggered one only 23.4% of the time. If your business lives on “plumber near me,” the AI Overview is mostly not your problem yet. If it lives on “why is my water heater leaking,” it is.

Finding 1: organic rank is still the main road in

Of the 1,561 outside links in the 400 overviews, 796, or 51.0%, pointed to a URL that was already ranking in the organic top ten for that exact query. Count by domain instead of exact page and it rises to 63.7%, meaning another 12.7% of citations went to a different page on a site that was already on page one. In 347 of the 400 overviews, at least one cited page was an organic top-ten result.

Where 1,561 AI Overview citations pointed

Matched against the organic top ten of the same search. 400 US SERPs, September 2026.

The cleaner way to see it is to flip the question: if you hold an organic position, how often does the overview cite you?

How often the AI Overview cites each organic position

Share of organic results cited by the AI Overview on the same page. 400 US SERPs, September 2026.

Position 1 is cited 56% of the time. Position 8, under 10%.

That curve is the single most important chart in this post. The first organic result is cited more often than not. By position four the odds have fallen by more than half. By position eight you are a coin flip away from irrelevant. Nothing about this is exotic: it is the same click curve we have watched for twenty years, applied by a machine instead of a thumb.

For context, Ahrefs found that 76.1% of citations came from the top ten in July 2025 and only 37.9% in March 2026 across four million cited URLs. seoClarity measured 32% in October 2025 and found the position-one page cited 43% of the time. My 51% sits between those studies, partly because I counted only the links Google shows first and partly because my sample leans informational. The direction everyone agrees on is that the overlap is falling as fan-out reaches deeper, which brings us to the second finding.

Finding 2: fan-out is real, and it favors pages that already rank for something

The other 49% of citations, 765 links, went to pages that were not in the organic top ten for the query. A third of those, 252, were YouTube videos, which Google pulls into overviews through its video systems. That leaves 513 citations to ordinary web pages that were not on page one for the query.

Those pages are not obscure. Their median Domain Rating was 79. Only 12.1% came from domains with a DR under 30. All but four of them (99.2%) rank in Google’s top 100 for at least one keyword, with a median of 45 keywords per page, and 77% rank for ten or more. In other words, they are indexed, established pages that rank for something adjacent. They got in through the side door the fan-out opened: the model asked a sub-question, Google ran that search, and these pages were on that shortlist.

Surfer measured the same effect from the other direction in a 10,000-keyword study: pages that ranked for both the main query and its fan-out queries were cited 51.2% of the time, against 19.6% for pages ranking for the main query alone, a 161% lift, with a 0.77 correlation between ranking for fan-out queries and being cited. The practical translation is that a page which answers the neighboring questions on the same URL has more doors into the overview than a page that answers one.

One caution, straight from Google’s guide: creating “separate content for every possible variation of how people might search (for example, by focusing on other queries that people have asked, or fan-out queries)” primarily to manipulate rankings or AI responses “violates Google’s scaled content abuse spam policy.” Cover the sub-questions inside a strong page. Do not spin them into forty thin ones.

Finding 3: page strength beats domain strength

This is the finding that surprised me. Everyone assumes AI Overviews are a big-brand game, and at the domain level the data looks that way at first: 55.9% of the organic results in my sample came from DR 80 or higher domains, and those results were cited 31.6% of the time against 21.6% for DR under 30. But that gap is mostly rank, not authority. When I compare cited and uncited pages at the same organic position, the median DR is almost identical: 90 versus 85 at position one, 87.5 versus 85 at position two, and 84 versus 85 at position three. Once you are on page one, the model does not seem to care much how strong your domain is.

What it does seem to care about is the page. Among all 2,769 organic results, pages with zero referring domains were cited 16.6% of the time. Pages with one to nine referring domains, 26.9%. Ten to 99, 33.1%. One hundred or more, 39.2%. Restrict it to the top three positions and the gradient gets steeper: 22.3% for pages with no links pointing at them, rising to 57.3% for pages with a hundred or more linking domains.

Same rank, same DR: the page with more links to it gets cited.

The keyword count is the fan-out story again. A page that ranks for a hundred queries is, by definition, a page that ranks for the sub-questions. The referring-domain count is the patent’s “trustworthiness measure” based on “inbound link(s)” showing up in the wild. Median numbers tell the same story: cited pages had a median URL Rating of 6 against 4, 26 referring domains against 10, and 2,951 monthly organic visits against 1,086 for the pages beside them that were not cited.

The lesson for a small business is not “give up.” It is that links to the specific page you want cited matter more than your overall domain score, and that a page which thoroughly covers a topic outperforms a strong site with a thin page. I have seen this on my own AI-built site, which has a Domain Rating of zero and still ranks for more than a thousand queries on content depth alone.

Finding 4: video and forums take a fixed slice, and it varies wildly by industry

YouTube was the single most cited domain in my sample: 265 citations, 17.0% of all outside links, appearing in 145 of the 400 overviews. Reddit was second at 5.1%. Add Quora, Facebook, TikTok, LinkedIn and Instagram and forums plus social take 23.6% of all citations. Government sites took 6.2%, driven by the IRS, state courts, the National Park Service and the NIH.

Finance cites what ranks. Software cites what plays.

Look at the two ends. In finance, the overview almost always cites what already ranks, YouTube barely appears, and the sources are Fidelity, Schwab, Experian and the IRS. In health, Reddit was cited exactly zero times across 151 citations, and the top-ranked page was cited 78% of the time. Those are the YMYL categories, and Google is visibly conservative in them. At the other end, software and industrial queries pulled a quarter to a third of their citations from YouTube, because “how does a refrigerated air dryer work” and “how to create a pivot table” are questions people want shown, not told. If you sell equipment or software and you have no how-it-works video on YouTube, you are conceding a third of the citation slots before the game starts.

Ahrefs’ running tally of the most-cited sites in AI Overviews across three million queries puts YouTube at 22.9% and Reddit at 18.5% of the top-50 pool as of September 2026, so my sample, which is lighter on entertainment and shopping queries, is probably a floor for how much video and forum content you are competing with.

Finding 5: the concentration is real but smaller than the scary headlines

You have probably seen the line that one percent of domains take half of all citations. In my sample the 1,561 outside citations spread across 719 domains. The ten most-cited domains took 30.5% of the links. Forty-four domains were cited five or more times and together took 45.2%. But 547 domains, three quarters of the list, were cited exactly once, and those one-off citations were 35% of the total. There is a head, and it is mostly YouTube, Reddit, Home Depot, the Cleveland Clinic and Fidelity. There is also a long tail, and it is where most businesses live.

One more detail worth knowing: the organic results Ahrefs classified as articles were cited 31.3% of the time, landing pages 37.1%, listing pages 33.3%, and video pages hosted on ordinary sites just 9.8%. The overview is not allergic to commercial pages. It is allergic to pages with nothing quotable on them.

The factor list, sorted by evidence

Here is how I would rank the factors people argue about, using three sources: my data, Google’s published statements, and the large controlled studies. The gates come first because if you fail one, nothing below it matters.

Gates (pass or fail). Googlebot can crawl the page. The page is indexed. It is eligible for a snippet, meaning no nosnippet or max-snippet:0 tag. The Search generative AI control in Search Console is still set to include. The site is not tripping a spam policy, and note that since May 15, 2026 Google’s spam policies explicitly cover “attempting to manipulate generative AI responses in Google Search.”

Tier one: measured, strong, and confirmed by Google. Ranking in the organic top ten for the query (56% cited at position one, under 10% at position eight). Ranking for the fan-out sub-queries (Surfer’s 161% lift; my 513 side-door citations that all rank for something). Links pointing at the specific page (22% to 57% in the top three). Breadth of what the page covers (pages ranking for under ten keywords were cited 8.6% of the time; pages ranking for a hundred or more, 34.5%). Specific, verifiable facts on the page, which is what the patent’s “verify portion(s)” language means in practice; Surfer’s separate 57,000-URL study found cited pages covered 31% of a topic’s key facts against 24% for uncited pages.

Tier two: supported, but indirect. Brand mentions across the web. Ahrefs’ 75,000-brand study found web mentions had the strongest correlation with AI Overview visibility (0.664), with brands in the top quartile for mentions earning over ten times the AI Overview mentions of the next quartile. That is a correlation between being talked about and being cited, and it lines up with the patent’s “popularity measure,” but it is not something you can buy: Google’s guide lists “seeking inauthentic ‘mentions’ across the web” among the things you do not need to do, and the May 2026 spam clarification puts attempts to manipulate AI responses under the same policy as classic ranking manipulation. Video on YouTube, for the industries above where a fifth to a third of citations are video. Corroboration, meaning your facts agree with what other trusted sources say, because the model weighs how many documents verify a claim. Freshness, which the patent lists but which the data does not reward much: Ahrefs’ 17-million-citation study found AI Overview citations averaged 1,432 days old, the same as organic results.

Tier three: weak or none. Structured data. Google’s guide says “structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add,” and when Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 controls, AI Overview citations moved by minus 4.6%, which they called noise. Keep schema for the rich results it earns; do not expect it to buy a citation. An llms.txt file. Google’s guide: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.” Ahrefs checked 137,210 domains and found 28% publish one and 97% of those files got zero requests in a month. SE Ranking found no correlation with citations across 300,000 domains. When John Mueller was asked in January 2026 whether Google endorsed the file, his answer was “to be direct, no.” Chunking content into bite-sized pieces for the model. Danny Sullivan on Google’s own podcast in January 2026: “We don’t want you to do that. I was talking to some engineers about that. We don’t want you to do that. We really don’t.” Word count. Ahrefs found a 0.04 correlation between length and citation position, with 53.4% of citations going to pages under 1,000 words. Rewriting your content “for AI.” Google’s guide lists that as a myth too.

What this means for your site

Most of what I would tell a client after this study is the same thing I told them in 2019, with two additions.

Rank first, because that is where the shortlist comes from. Google put it plainly in the May 2026 guide: “optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” Every fashionable acronym is a rebrand of that sentence.

Answer the neighboring questions on the same page. Write out the sub-questions a searcher would ask next, the ones Google would generate as fan-out queries, and answer each one with a heading and a direct paragraph. That is how a page ends up ranking for a hundred queries instead of one, and it is what the citation data rewards. It is also the opposite of chunking: one thorough page, not many thin ones.

Give the model a sentence it can quote. Every important claim on the page should be specific, checkable and attributed. “Refrigerated dryers deliver a pressure dew point of roughly 38°F” is citable. “Our dryers deliver superior performance” is not. If you have first-party numbers, publish them, because a fact nobody else has is a fact the model can only get from you.

Earn links to the page, not just the homepage. The strongest separator between cited and uncited pages at the same rank was referring domains to the URL itself. A single well-placed link from a trade association or a supplier to your guide page does more for citations than ten links to your homepage.

Put the how-it-works content on YouTube as well as on your site, if you sell equipment, software, or anything a person wants shown. In those verticals a quarter to a third of the citation slots go to video, and a video from a manufacturer beats a video from a hobbyist for trust.

Be present where the corroboration happens. Reddit, Quora, trade forums and review platforms take a fixed slice of citations because they are where “what people are saying” lives. Show up in those conversations as yourself. Do not seed them, and do not pay for mentions; Google’s guide calls inauthentic mentions unhelpful, and its spam policy now covers attempts to manipulate AI responses.

Measure the right thing. As of August 31, 2026, Search Console has a Generative AI performance report that shows impressions from AI Overviews and AI Mode, by page, country and device. It does not show clicks or queries, and it does not separate the two features. Clicks from both are still blended into the Web search type in the regular Performance report, and an AI Overview click inherits the block’s single position. Google Analytics cannot tell an AI Mode click from a blue-link click. So your reporting should pair the new impressions report with a third-party tracker that records which prompts cite you, and it should stop treating a position-one ranking as a traffic guarantee, because Ahrefs measured the position-one click-through rate on keywords with an AI Overview falling by 58% between December 2023 and December 2025.

Do not spend money on llms.txt, AI-specific schema, content chunking, or anyone selling “AI citation placements.” The first three do nothing that Google or the data can find, and the fourth is the kind of manipulation Google’s spam policy now names.

What I tell clients about AI Overviews

I have done SEO since 2004, and I will say it again: nobody is an expert in AI and SEO right now, including the people selling you certainty. What I can offer is a measurement you can check, from data anyone with an Ahrefs login can reproduce, and a reading of Google’s own documents that most of the lists you have seen never bothered to open.

The honest summary is that AI Overviews did not invent a new ranking system. They bolted a reader onto the old one, taught it to ask follow-up questions, and told it to footnote its claims. The pages that win are the pages that rank, cover the topic completely, say something specific enough to quote, and have other sites vouching for that page in particular. That was already the assignment. The overview just grades it faster.

People lie. Results don’t. If you want this study run against your own keywords and competitors, with the fan-out queries mapped to pages you already own, that is exactly what my AI SEO services engagement does. Or book a free consultation and I will tell you whether it is worth doing, including the cases where it is not.

Method. Keyword pool: 50 seed terms (five per vertical, listed in the first table) expanded with Ahrefs Keywords Explorer, US, terms match, three or more words, 100 to 20,000 monthly searches, 80 keywords per seed ordered by volume, de-duplicated to 4,046 keywords; presence rates use the 3,804 keywords whose SERP was captured on or after August 1, 2026 and count the ai_overview SERP feature. Study sample: from keywords with an AI Overview and a snapshot dated September 2026, 48 were drawn at random per vertical, off-topic matches were removed by hand (celebrity news, slang, unrelated uses of a seed word) and 40 per vertical were kept, 400 in total. For each, the Ahrefs SERP overview was pulled with organic, AI Overview and AI Overview link rows within the top ten positions; snapshot dates run September 1 to 24, 2026. Because the AI Overview block occupies one position, the organic rows cover organic positions one through eight in most SERPs (2,769 organic rows). Citation matching used normalized URLs (scheme, www and trailing slashes ignored, YouTube matched on video ID); domain matching used the registrable domain. Of 1,764 links in the AI Overview blocks, 115 were Google Shopping product cards and 88 were opaque Google redirects that could not be resolved, both excluded, leaving 1,561 outside citations. Domain Rating, URL Rating, referring domains, traffic and keyword counts for organic results come from the same SERP snapshot; the same metrics for the 747 unique cited pages outside the organic results come from Ahrefs Batch Analysis (exact URL mode) on September 24, 2026. Ahrefs captures the links Google displays with the overview (median three per overview here, up to 14), not every source in an expanded view, so overlap figures describe the most visible citations. This is a single snapshot of US desktop results, verticals are defined by seed topics, and I did not crawl the cited pages myself, so the structured data and llms.txt conclusions rest on the Google statements and the Ahrefs and SE Ranking studies linked above rather than on my sample. Correlation is the claim throughout; no single table proves causation.