How to Get Cited by AI Search: What the Data Says, and What Google Ignores

The Short Answer

If you want one instruction, take it from the source that decides half the question. In its guide to optimizing for generative AI features, last updated on 10 July 2026, Google states that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

The same page says you do not need special structured data, you do not need to chunk your content, and you do not need to write differently for machines.

That is an unsatisfying answer for anyone hoping to buy a shortcut, and it is the accurate one. What has genuinely changed is not the technique but the payoff: fewer people click, so the citation itself is now the visibility.

What the Click Data Shows

The most methodologically transparent study is from the Pew Research Center, published 22 July 2025. It tracked the real browsing of 900 US adults across 68,879 searches in March 2025, of which 12,593 returned an AI summary. The results:

Ahrefs reached a similar direction from a different angle. Comparing 150,000 keywords that trigger AI Overviews with 150,000 that do not, it reported a 34.5% lower click-through rate at position one in April 2025, and 58% in its February 2026 re-run. Read these as modelled counterfactuals from a company that sells SEO tooling, not as measured causation. Ahrefs says so itself.

How Much Traffic AI Sends

Less than the discourse implies, and it lands in unexpected places. According to Similarweb data reported by Search Engine Journal on 30 July 2026:

• Search still had about 3.3 billion monthly unique visitors against roughly 655 million for chatbots.
• 461 million of ChatGPT's 494 million users, about 95%, also used Google between March and May 2026. These are overlapping habits, not a migration.
• Only 6.8% of US ChatGPT answers included an external link in May 2026, up from roughly 1%.
• 65% of cited URLs sat two or three folders deep, while 58.8% of AI referral traffic landed on a homepage.

The crawl-to-referral imbalance is starker. Cloudflare measured HTML requests per HTML referral across its network between 19 and 26 June 2025 and reported roughly 70,900 requests per referral for Anthropic and about 0.1 for Mistral.

Cloudflare notes the caveat that Claude's native app sends no referrer header, so the ratio overstates the gap. The honest summary: AI assistants read far more than they send back, and that asymmetry is structural.

The Crawlers That Gate Citations

This is the part most teams get wrong, because blocking the training crawler and blocking the citation crawler are different decisions with the same-looking syntax.

Three practical consequences. First, if you block everything with a blanket rule, you have removed yourself from AI answers while gaining nothing on the training question you actually cared about — check the OpenAI, Anthropic and Perplexity documentation for the exact tokens.

Second, user-initiated fetchers behave differently: OpenAI says robots.txt rules may not apply to ChatGPT-User, and Perplexity says Perplexity-User generally ignores robots.txt. Third, Google-Extended is not an AI Overviews opt-out; for that, Google's AI features documentation points to nosnippet, data-nosnippet, max-snippet and noindex.

What Does Not Work

Google's optimization guide is unusually blunt about the current crop of tactics, and it is worth quoting rather than paraphrasing.

• llms.txt. On these files, Google writes that adding one "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them." The llms.txt specification site itself claims only that thousands of sites publish one and that AI labs publish them for their own developer documentation — not that any assistant reads yours when answering a question. No AI vendor's own documentation that we could open says its systems consume your llms.txt.
• Special AI markup. Google states there is no special schema.org structured data you need to add, and no machine-readable AI text file requirement.
• Content at scale. Generating many near-identical pages violates Google's scaled content abuse policy.
• Buying mentions. Google's wording: "Seeking inauthentic 'mentions' across the web isn't as helpful as it might seem."

How to Get Cited by AI Search

1. Serve your text in the HTML. If the content only appears after JavaScript runs, you are relying on every retrieval system rendering it. Server-rendered content removes that dependency, which is one of the reasons we build the way we describe in why frontend matters.
2. Answer one question per page, in the first two paragraphs. Retrieval favours passages that stand alone. Pew found that 60% of the searches producing AI summaries were questions, and 53% of ten-word-plus queries triggered one.
3. Make your facts quotable. Put figures in tables with units and dates, and attribute them to a named source. An assistant cannot cite a claim it cannot attribute.
4. Write FAQ answers that survive being extracted. Two or three sentences that name the subject, with no dependency on the paragraph above.
5. Keep deep pages strong. Since 65% of cited URLs sit two or three folders deep, your detailed guides matter more than your homepage for citation, even though the homepage receives most of the referral traffic.
6. Fix the basics. Google lists crawlable pages, unblocked JavaScript, semantic HTML, and good page experience among the things that do help — the ground covered in Core Web Vitals.
7. Measure it. Segment AI referrers in your analytics so the discussion runs on your own numbers rather than industry averages; our Google Analytics guide covers the setup.

Conclusion

AI search has not replaced search. Similarweb's panel data shows chatbot visits growing about 70% year on year, while the overwhelming majority of those users still search too.

What it has done is raise the cost of being unremarkable. Pages that merely restate what everyone else says now get summarised away; pages with original figures, clear structure, and verifiable sources get named. That is the same discipline good publishing always required, with a sharper penalty for skipping it.