In 2025, Ahrefs found that 76% of the pages cited in Google's AI Overviews also ranked in the top ten organic results for the same search. When it repeated the study in 2026 across about 863,000 searches, the overlap had fallen to 38%. The rest of the citations split almost evenly between pages ranking 11 to 100 and pages outside the top 100 entirely.
That is a large change in a single year, and it cuts both ways. Ranking first no longer guarantees you are the source. Ranking twentieth no longer rules you out.
AI answers cite the page that answers a precise sub-question most clearly and most checkably, which is often a different page from the one that ranks for the broad search. Being cited is a page-level writing problem as much as a domain-level ranking problem.
Why rank and citation drifted apart
The likely mechanism is what Google calls query fan-out. When an AI feature answers a question, it breaks the question into several narrower searches, gathers results for each, and assembles an answer. A broad query such as "best way to ship groceries to Norway" might fan out into customs rules, delivery times, cold-chain options and specific retailers. The page cited for "delivery times to Norway" does not need to rank for the original question. It needs to answer that one narrower question better than anything else.
This is consistent with what we see in client analytics. On an online store we run, only 16% of visitors arriving from ChatGPT this year landed on the homepage. The rest landed on individual product pages, a page explaining delivery across Europe, and a seasonal guide. None of those pages ranks for broad category searches. Each one answers one narrow question completely, in plain text.
What gets cited
The sources AI answers draw on are not a random sample of the web. In Pew Research Center's study of real Google searches, Wikipedia, YouTube and Reddit together made up 15% of the sources cited in AI summaries. Government sites made up 6% of summary sources, three times their 2% share of standard results. Assistants favor sources that look authoritative, specific and checkable.
Here is the concession: a small business will rarely outrank Wikipedia or a government site as the source for a general fact, and should not try. The opening is in questions only you can answer authoritatively: your prices, your delivery areas, your process, your results, your product specifications. Nobody else is the primary source for those.
Five things that make a page citable
1. One question, answered in the first two sentences. If the page is about delivery to Norway, the first sentence should say whether you deliver to Norway, how long it takes and what it costs. Answers buried under an introduction are skipped by people and by machines alike.
2. Facts stated as sentences, not implied by design. A price inside an image, a delivery area shown only on a map, or a specification inside a downloadable PDF is invisible to most AI fetchers. If it matters, write it as text on the page.
3. Content in the HTML the server sends. Several AI crawlers do not run JavaScript reliably, so content that only appears after scripts load may never be seen. We build client sites to render their main content on the server for this reason; the cost is a little more engineering on interactive pages.
4. A clear, consistent entity. Assistants need to know who is making a claim. Use one organization name, one description and the same facts everywhere: your site, your Google Business Profile, directory listings and social profiles. On our own site we declare a single Organization in structured data and removed a personal profile that was confusing the company record with a person.
5. Mentions elsewhere. Research from Ahrefs and others has found that brands mentioned widely across the web are cited in AI answers more often than brands with similar link profiles and fewer mentions. These are correlations, not proof of cause, but they point in the same direction as everything above: assistants trust what other sources confirm.
What does not matter as much as people think
An llms.txt file is optional; Google has said it does not use it, and we publish one only because it costs nothing. Schema markup helps machines read a page but does not make a vague page citable. And writing "for AI" in a stiff question-and-answer format tends to produce pages people dislike, which defeats the purpose, because the AI-referred visitor still has to buy from you when they arrive.
Where this stops applying
Citation studies measure Google's AI Overviews more than ChatGPT, Perplexity or Claude, and each assistant chooses sources differently. Being cited also does not guarantee a click: in Pew's data, people clicked a link inside an AI summary in only 1% of visits. Citation builds the brand and wins the visitors who are ready to act. It does not restore the traffic informational pages used to get, a shift we cover in who is eating your clicks.
Questions to ask of your most important pages
- For each key page, what single question does it answer, and is the answer in the first two sentences?
- Are your prices, delivery areas and lead times written as text, or only shown in images and PDFs?
- Does your main content appear in the HTML before any JavaScript runs?
- Would someone comparing your site, your Google profile and your directory listings find the same facts?
We run this page-level audit as part of our SEO work, and we build new sites to pass it from the start through our website development process.