How to Get Cited in ChatGPT and Google AI Overviews

A generative engine optimization (GEO) playbook built on primary data — not another generic checklist.

Most of what you’re doing for SEO right now is not what gets you cited by AI. That’s not a guess — it’s what the data says.

A 100-page study by CXL found that only 37.1% of URLs cited in Google AI Overviews also rank in the organic top 10. More than a third of cited pages (36.7%) don’t appear in the top 100 organic results at all. Separately, Ahrefs analyzed 1.4 million ChatGPT prompts and found that ChatGPT cites only about half of the URLs it actually retrieves — ranking well and getting retrieved are necessary, but they’re nowhere near sufficient.

If you’ve been treating “get cited by AI” as a downstream reward for good SEO, this is the piece that explains why that assumption breaks down, what actually predicts citation, and what to do about it — with the studies to back each claim, not just another numbered checklist.

How AI citation actually works

Every AI engine that cites sources runs some version of the same four-stage pipeline: retrieval, ranking, selection, and citation. Confusing these stages is why most GEO advice falls apart in practice.

Retrieval. The engine breaks your question into several related sub-queries — a process often called query fan-out — and searches for each one, pulling back a set of candidate pages across all of them. Ahrefs found ChatGPT retrieves an average of about 33 URLs per prompt (roughly 16.6 that eventually get cited and 16.6 that don’t).

Ranking. Retrieved pages get scored for relevance to the query and its sub-queries. This is closest to traditional SEO — but it’s not the last step.

Selection. The model decides which ranked pages are actually worth quoting or attributing. This is where most of the attrition happens: Ahrefs found search-result URLs get cited 88.46% of the time when retrieved, but Reddit URLs — despite being retrieved 16.2 million times in the study — get cited only 1.93% of the time, and YouTube and academic sources fare even worse (0.51% and 0.40%). The model clearly uses Reddit heavily for context and consensus-building; it just rarely credits it publicly.

Citation. The final answer gets assembled with inline attributions or a source list. This is the only stage a reader (or a rank tracker) ever sees, which is exactly why it’s easy to mistake citation for ranking.

The practical implication: optimizing only for retrieval and ranking (classic SEO) leaves the entire selection stage unaddressed. That’s the stage the rest of this guide is really about.

It’s not one pipeline — it’s several

Google AI Overviews, ChatGPT, Perplexity, and Claude each run their own version of this pipeline against different indexes, and they don’t overlap as much as you’d expect. Citation overlap between Claude and Google runs around 64%, but between Claude and ChatGPT it’s only about 8%. ChatGPT triggers a live web search on roughly 90% of prompts; Claude does so on only about 37%, leaning more on its training data and cached context. Optimizing for one engine’s citation behavior does not transfer cleanly to the others — see the engine-by-engine notes further down.

What actually predicts citation

Four factors show up consistently across the primary research, and none of them is “domain authority.” An analysis GBG Marketing cited, covering 5 million URLs from Surfer, found almost no correlation between domain authority and AI citation probability — which means a smaller site has a real shot here, but only if it gets these four things right.

Semantic match to the query. Ahrefs measured the cosine similarity between page titles and the original prompt: cited pages averaged 0.602 similarity versus 0.484 for non-cited pages. Titles that matched fan-out sub-queries (see above) scored even higher, at 0.656. Vague, brand-first titles lose to titles that mirror how people actually ask the question.

Where the answer sits on the page. CXL’s citation-position study found 55% of AI Overview citations pull from the first 30% of a page — a pattern they call the “ski ramp” effect: steep concentration near the top, then a sharp drop-off. Independent analysis from Kevin Indig found burying a key definition deep in an article reduces its retrieval probability by roughly 2.5x compared to placing it in the introduction. The one consistent exception: self-contained FAQ question-and-answer pairs keep citing well regardless of where they sit on the page, because they’re already structured as an extractable unit.

Content age. The median cited page in Ahrefs’ search-result dataset was about 500 days old — relevance and established topical depth appear to outweigh freshness for general informational content. News-type sources flip this: younger pages (median ~200 days) win when relevance is otherwise similar, which matters if part of your GEO content is timely commentary rather than an evergreen guide.

URL and source-type signals. Natural-language URL slugs got cited 89.78% of the time versus 81.11% for opaque, parameter-heavy URLs. And as covered above, which type of source a URL is matters enormously — a forum post and a published article with the same information do not get treated the same way at the selection stage.

The practitioner checklist

Everything above points to two things AI engines are actually evaluating: can they extract a clean answer from your page, and can they trust it enough to attribute it. Organize your work under those two headings, not a flat numbered list.

Extractability

1. Answer first, explanation second. Put the direct answer to the implied question in the first 1–2 sentences of a section, then expand. Don’t make the model — or the reader — dig for it.
2. Write headings as questions. “How does AEO work?” extracts more cleanly than “AEO methodology.” Match the phrasing real people search with, not internal jargon.
3. Add FAQPage and HowTo schema. Structured data doesn’t guarantee citation, but it makes extraction mechanically easier and costs little to implement.
4. Unblock the AI search crawlers. AI companies now run separate bots for search and for model training, and blocking one does not block the other. For citations, the ones that matter are OAI-SearchBot (ChatGPT search), Claude-SearchBot (Claude) and PerplexityBot (Perplexity). GPTBot and ClaudeBot collect training data, so you can block them and still appear in AI answers. Google AI Overviews use the normal Googlebot; the control for those is the generative AI setting in Search Console. You can skip llms.txt: Google says you can ignore it.
5. Use natural-language URLs. /blog/generative-engine-optimization-guide over /blog/p?id=4471.

Trustworthiness

6. Name a real author with real credentials. Anonymous or generic “Team” bylines are a weak trust signal for a model deciding whether to attribute a claim to you.
7. Cite your own sources. Pages that name where their numbers come from are easier for a model to treat as reliable secondary sources themselves.
8. Earn third-party mentions. Being referenced by other sites and being discussed on platforms the model retrieves from (review sites, industry press, Reddit threads even if they don’t cite you back) builds the consensus signal that feeds selection.
9. Refresh on a real cadence, not a one-time push. The HOTH’s engine-by-engine research found pages refreshed quarterly saw 502% more ChatGPT referral traffic than stale ones — an update date and a genuinely revised page beat a “last modified” stamp with no real changes.
10. Test with real prompts, monthly. Run the actual questions your buyers would ask through ChatGPT, Perplexity, and Google AI Mode and note whether you show up — don’t rely on rankings as a proxy.

Worked example: a B2B SaaS company

Say a mid-market project-management SaaS wants to get cited when someone asks ChatGPT “what’s the best project management tool for remote teams.” A generic company page (“Features” → “Pricing” → “About”) gives the model nothing extractable. A page structured around the actual question — a direct one-paragraph answer up top, a comparison table against 4–5 named competitors, an FAQ section covering “is it good for remote teams,” “how much does it cost,” and “does it integrate with Slack,” published under a named product marketer with a linked bio, and refreshed each quarter with current pricing — hits all ten items above without reading like an ad. That’s the difference between a page that ranks and a page that also gets quoted.

Engine-specific notes

Each engine has its own quirks worth knowing:

• Google AI Overviews and AI Mode. Appears on 40%+ of searches. ~75% of citations come from pages already in the top 12, but only 37.1% from top-10, so ranking well helps but isn’t enough alone. AI Mode queries run 2-3x longer, a clear sign of query fan-out: optimize for the whole sub-question cluster, not one keyword.
• ChatGPT. Runs a live web search on ~90% of prompts, the most search-dependent of the major engines. Semantic title match and source type matter most; it favors a direct, well-organized answer over marketing copy.
• Perplexity. Shows sources inline on nearly every answer, making it the fastest engine to test tactics on.
• Claude. Searches live on only ~37% of prompts, leaning more on cached knowledge and Brave’s index. Overlaps ~64% with Google’s citations but only ~8% with ChatGPT’s, so ranking well in Google still helps Claude visibility.
• Gemini. Grounds live sources on only ~41% of answers, versus 91%-98% for the other three (per Cloro’s grounding study), and cites just 3.2 sources on average when it does, the fewest of the four; the rest come from training data with no citation at all. The AI Citation Index found only 12% source overlap with ChatGPT, and Gemini’s citation links are unverifiable Google redirects rather than real URLs. Don’t assume a ChatGPT citation carries over to Gemini.
• India. The pipeline doesn’t change by country, but what a model can find and trust in your market does. Google added AI Mode support for 7 more Indian languages in October 2025, on a model trained for them directly, not machine translation. An English-only page is invisible to Hindi or Tamil queries even if it ranks well in English, and the same trust signals (reviews, credentials, clear pricing, see our guide to Indian service-business pain points) matter regardless of language. We found no reliable India-specific breakdown of which domains get cited most, so treat such figures elsewhere with scepticism and test your own prompts monthly instead.

Don’t treat “AI search” as one channel. If AI-referred traffic matters to your business, track these four separately from day one.

How to measure it

Start with what you already have, then add dedicated tracking once GEO is a real budget line.

In GA4, segment referral traffic from chatgpt.com (and the older chat.openai.com), perplexity.ai, and other AI domains into their own channel group rather than letting it fall into generic “referral.” It’s a rough proxy — not every AI answer includes a clickable link — but it’s free and immediate.

Dedicated AI-citation tracking tools (Peec, Profound, and similar platforms) run your target prompts on a schedule across ChatGPT, Perplexity, Gemini, and Claude, and report whether and how often you’re cited. Worth adopting once you’re targeting more than a handful of prompts, since manual testing doesn’t scale.

Manual prompt testing is the cheapest starting point: pick 10–15 real buyer questions, run them through each engine monthly, and log whether you appear, in what position, and alongside which competitors.

On timelines: be honest that this is slower and noisier than traditional SEO. The quarterly-refresh stat cited above (502% more ChatGPT referral traffic) implies compounding gains over a quarter-plus horizon, not weeks. Given AI citation databases update on their own schedules (independent of any single engine’s crawl cadence), expect to wait at least one full refresh cycle before drawing conclusions from a change you made.

Common mistakes

• Chasing rankings instead of extractability. Good organic position raises the odds of retrieval; it says nothing about selection. Treat them as two different jobs.
• Treating “AI search” as one engine. Optimizing only for what works in ChatGPT and assuming it transfers to Google AI Overviews or Claude ignores the ~8% citation overlap between ChatGPT and Claude noted above.
• Skipping schema and crawler access. These are close to zero-cost and often overlooked — confirm OAI-SearchBot, Claude-SearchBot and PerplexityBot are allowed in robots.txt, and that your CDN or firewall isn’t blocking them, before doing anything else.
• Generic, undifferentiated messaging. Pages that open with broad, interchangeable claims (“the leading solution for X”) give a model nothing distinctive to extract or attribute. Specific, named claims outperform.
• No measurement plan. Shipping GEO changes without a baseline (manual prompt tests or a tracking tool) makes it impossible to tell whether anything worked.

Quick-start checklist

The same checklist from above, condensed into something you can run against a page right before you hit publish:

• Direct answer appears in the first 1–2 sentences of its section
• Headings are phrased as the questions people actually ask
• FAQPage and/or HowTo schema is implemented
• robots.txt allows the AI search crawlers: OAI-SearchBot, Claude-SearchBot and PerplexityBot
• Search Console’s generative AI setting matches your decision on appearing in AI Overviews and AI Mode
• URL is natural-language, not parameter-based
• A named author with a real bio is credited
• Claims cite their own sources
• Update date reflects a genuine quarterly (or better) refresh
• The page has been tested with 3–5 real prompts across ChatGPT, Perplexity, and Google this month

Running your own pages against this list — or having someone audit them for you — is a faster starting point than rebuilding a content strategy from scratch.

FAQ

How do I get cited in AI Overviews? Structure the specific answer to a specific question in a self-contained block near the top of a page, back it with FAQ or HowTo schema, and make sure Googlebot can crawl and index it, since AI Overviews draw on Google’s normal index. For ChatGPT, Claude and Perplexity, also allow OAI-SearchBot, Claude-SearchBot and PerplexityBot. Ranking well helps but isn’t enough on its own — only 37.1% of AI Overview citations come from top-10 organic results.

How do I appear in AI Overviews if I’m not already ranking well? Focus on a narrow, well-answered question rather than a broad head term. CXL’s data shows self-contained FAQ-style answers cite well regardless of their position on the page, which is the one part of the “ski ramp” pattern that favors smaller or newer sites.

What’s the difference between SEO, AEO, and GEO? SEO optimizes for ranking in traditional search results. AEO (answer engine optimization) optimizes for being the direct answer in featured snippets, voice search, and AI answer boxes. GEO (generative engine optimization) optimizes specifically for being retrieved, selected, and cited by generative AI tools like ChatGPT, Perplexity, and Google’s AI Overviews. In practice the three overlap heavily and most of the same on-page work supports all three.

Does domain authority affect AI citations? Barely, according to the data. Surfer’s analysis of 5 million URLs found almost no correlation between domain authority and citation probability — semantic relevance, page structure, and source type matter far more than backlink profile strength.

How long does it take to get cited by ChatGPT? Expect a quarter or more before you can draw real conclusions. Pages refreshed on a genuine quarterly cadence saw 502% more ChatGPT referral traffic in The HOTH’s data — which points to compounding gains over months, not a one-time fix.

How do I show up in Google AI Overview specifically? Google AI Overviews lean more heavily on existing organic rankings than the other engines (about 75% of citations come from pages already in the top 12), so traditional on-page and technical SEO still carries real weight here even though it isn’t sufficient by itself.
Does Gemini cite sources the same way ChatGPT and Perplexity do? No. Gemini only grounds on live web sources for a minority of answers, roughly 41% in one engine-by-engine study, versus 91%-98% for ChatGPT, Perplexity and Google AI Overviews, and cites far fewer sources when it does. On most answers it responds from its trained knowledge with no citation at all, so a page being cited by ChatGPT is no guarantee it will ever be surfaced by Gemini.
Is GEO different for Indian-language searches? The underlying mechanism is the same, but Google has been actively expanding AI Mode into seven more Indian languages beyond English and Hindi, using a Gemini model built for those languages rather than machine translation. If part of your audience searches in Hindi, Tamil or another regional language, an English-only page cannot be retrieved for those queries, regardless of how well it ranks in English.

Ready to get cited?

This is the exact framework we run for clients — from technical AI-crawler audits to full GEO and AEO implementation. See how it’s worked for other companies, or book a free AI citation audit to see where your site stands today.