What Gets Cited in AI Answers? Query Type Beats Topic

“How do I get cited by AI?” usually gets answered with topic advice: write about your niche, build authority in your vertical. New data says that’s the wrong first question. The bigger lever is the shape of the question your content answers, not the subject you write about.

cloro’s AI Search Index ran the same frozen basket of 735 queries through ChatGPT, Perplexity, Copilot, Gemini, and Google AI Mode, with an equal number of prompts in every combination of eight industries and six query types. Because the basket is balanced, it can separate two things a normal sample tangles together: does citation behavior depend on the topic, or on the kind of question? The answer is lopsided.

Query type moves citations ~2× more than topic

Here’s how often an AI answer carried at least one citation, broken out by query archetype, pooled across all five engines at 120 prompts each, running from the most commercial intent (“best CRM for startups”) down to the most definitional (“what is a REST API?”):

That’s a 10.5-point spread from top to bottom, and it isn’t random: the ranking runs cleanly from commercial intent down to definitional intent. “Best X” and buying-guidance queries, where the answer hinges on current, specific facts like prices, rankings, and product specs, pull a citation nine times out of ten. Definitions, which an LLM can answer confidently from its own training, reach for an external source least often.

Now compare that to the same metric cut by topic. Every vertical gets the same 90 prompts, so this is a like-for-like read. On the same scale, the bars barely move:

The entire eight-vertical range fits inside 5.4 points. Setting aside education (82.4%), the other seven sit within about two points of each other (85.6–87.8%). Whether you’re in travel or SaaS or health, an AI answer is about equally likely to cite someone.

Put the two charts side by side and query type accounts for roughly twice the swing that topic does. The variation lives in the query, not the subject.

Why query shape beats your niche for GEO

Most generative-engine-optimization advice is organized around topics and authority. Become the trusted voice in your niche and the citations follow. That’s not wrong, but it’s the second question. The first is whether the queries you’re targeting are the kind that get cited at all.

A page built around a definitional head term (“what is X”) is swimming against a 10-point current before it starts: those answers cite a source only ~81% of the time, and when the model does cite, it’s choosier. A page built around a top commercial intent (“best X”, “which X should I buy”) is aimed at answers that cite ~90% of the time. Comparison and how-to shapes land a notch lower (~83%), still comfortably above definitions. Same effort, meaningfully different odds of appearing in the citation list.

That changes the order you plan content in. Ranked lists and buying guides sit at the top of the citation-presence chart (91% and 90%), so if getting cited is the goal, those commercial formats are where the spend goes furthest.

Definitions and how-tos aren’t a lost cause, they just need arming. They get cited less because the model can usually answer them itself, so give it something it can’t generate: named data and dated figures, the kind of claim an engine can point back at you. A definition carrying a specific, verifiable statistic is a better citation candidate than one that restates common knowledge.

The thing you can mostly stop worrying about is your vertical. Your industry barely changes your citation odds, so the effort going into “is my niche too obscure to get cited?” is better spent on the query shapes above.

For the fuller framework of schema, structure, and formatting that makes a page quotable, see our generative engine optimization guide and the GEO checklist.

The engine caveat: presence is only half the picture

Getting cited at all is query-driven. How many sources an engine attaches, and which ones, varies enormously by engine, and that is a separate lever. In the same index, four of the five engines cite on 98–100% of answers while Gemini cites on just 33%; Google AI Mode backs an answer with ~20 sources where Copilot uses ~5. We broke the anatomy of that down across ~13,000 answers in how each AI engine cites sources, and mapped which engines ground their answers in live sources in AI grounding by engine.

The takeaway: a page that earns citations reliably overall can still go uncredited on Gemini specifically, so track citation presence per engine, not as one blended number, if any single engine matters to your traffic mix.

There’s also a durable pattern in what gets cited: pool the sources across engines and the top of the list is user-generated platforms, led by YouTube and Reddit, not publisher homepages. That’s why Reddit shows up in AI citations far more than its traffic would suggest, and it’s another reason the query-shape question matters: the formats that win citations (lists, comparisons, first-hand experience) are exactly what those platforms are full of.

An independent analysis of roughly 8,000 AI citations by Search Engine Land shows the same divide from the source side: ChatGPT leans on authoritative references (Wikipedia was ~27% of its citations) and almost never cites forums, while Google’s AI surfaces pull heavily on Reddit and community content. It also found that query intent, B2B versus B2C, reshapes which sources an engine trusts, echoing what the balanced basket shows about query shape.

What do AI citations look like?

An AI citation is a link attached to an answer that points at the page a claim came from. Engines show it as a numbered marker, a source card, or a list under the answer, and they differ most in how many they attach. In the same index, Google AI Mode backs an answer with ~20 sources, Copilot with ~5, and Gemini cites on just 33% of answers (how each AI engine cites sources). The cited page also does not need to rank: 80% of URLs cited by ChatGPT, Perplexity, and Copilot do not appear in Google’s top 100 for the same query.

Three moves to earn more AI citations, in order

1. Audit your target queries by intent, not just volume. Sort your priority prompts into the six archetypes above. The ones sitting in “best-in-class”, “buying guidance”, and “comparison” are your best citation bets; treat definitional and how-to targets as needing extra attributable substance.

2. Earn the citation once you’ve picked the intent. Presence is a population average: sitting in a high-citation query is the ante, not the win. Four levers move you from “an answer here gets cited” to “this answer cites you”.

   Format comes first, because commercial intents get cited as ranked lists, comparison tables, and buying guides. Publish that shape rather than a wall of prose; it’s what the engine is already reaching for. Then write for extraction. Engines lift self-contained “chunks,” so make claims that stand on their own, named and dated and specific enough to quote out of context. Semrush’s guide to AI citations lands on the same rule.

   Authority is what tips a close call, and it comes from original data and third-party mentions. That same Search Engine Land analysis found foundational search authority precedes AI citations rather than following them, so AI visibility is an outcome of broad web credibility rather than a separate trick. And none of it counts if a crawler can’t parse the page, so prefer server-rendered HTML, add schema markup (Google recommends JSON-LD), and publish an llms.txt so engines can reach and read your content.

3. Measure whether it’s working. Run your priority prompts across every engine on a schedule and track whether your domain lands in the citation list, per engine, over time. That’s exactly the pipeline cloro’s AI visibility tracking is built for, so you see whether a reformat actually moved your citation odds instead of guessing, and it’s the same data that produced the AI Search Index behind this post.

So before you ask “what should I write about,” ask “what kind of question does this answer.” On the current data that’s the bigger lever on whether AI cites you at all.

What content changes actually increase AI Overview citations?

Far fewer than the GEO checklists claim, and we measured it on our own site. We scored 249 of our pages on 27 on-page features (takeaway blocks, FAQ sections, quotable sentences, numeric density, extractable claims and the rest) and tested each against whether engines actually retrieved and cited the pages. No single on-page feature predicted citation on its own. Two features did separate retrieved pages from ignored ones: stating concrete prices and naming competitors. Pages that named no competitor at all were retrieved for 19% of eligible prompts against 50% for pages whose prose named the field, across that corpus.

The honest implication: treat checklist features (TL;DR blocks, FAQs, schema) as hygiene that makes a page liftable once retrieved, and treat retrieval as the fight. What gets a page into more answers, in our data, is being the kind of page that commits to specifics: real dollar figures, named vendors, claims an engine can attribute. A takeaways block on a page nobody retrieves changes nothing, and most GEO audits cannot tell you which problem you have.

Two outside studies measure things our feature set did not, and both point at specifics and placement. CXL’s 100-page study found 55% of AI Overview citations came from the first 30% of a page and 21% from the last 40%, so the answer belongs near the top. Surfer’s analysis of 57,000+ URLs found cited pages covered 31% of their topic’s key facts against 24% for pages not cited. Neither result makes a checklist feature predictive; both reward a page that states the facts early.

Does llms.txt or schema markup affect whether AI engines cite you?

There is no measured citation lift from either in our data, and we track both. Schema markup earns its place for a different reason: it labels facts for extraction and feeds Google’s rich results, which is classic SEO value with or without AI answers. An llms.txt file is cheap and harmless, and cloro publishes one, but across our corpus and our tracked prompts we have not observed pages gaining citations because the file exists. The llms.txt explainer covers what it is; treat claims that it drives citations as unproven until someone publishes the measurement.

What that means in order of operations: fix crawlability first (a blocked bot reads nothing), commit to specifics second (the retrieval levers above), and file schema and llms.txt under low-cost hygiene rather than strategy.

How do I improve my brand’s chance of being cited by ChatGPT?

Give ChatGPT the source types it actually quotes, and know that vendor blogs are mostly not among them. We read the search queries ChatGPT issued while answering our tracked prompts, and on the buying-intent ones it appended the word “official” to its own searches, then cited an official API reference, a news article and an academic paper while scoring zero for every vendor-published roundup in the retrieved set, ours included. Domain-wide across our prompt corpus, ChatGPT’s citations concentrate on independent editorial, documentation and academic sources; the one category it consistently declines to quote is a vendor’s own editorial about the vendor’s category.

Three concrete moves follow, in order of leverage:

1. Make your documentation excellent and public. Docs are a source type ChatGPT demonstrably quotes; a vendor’s reference pages can win retrievals its blog cannot.
2. Publish original data under your name. Studies and measured results get cited where marketing does not, and first-party numbers are claims only you can be the source for.
3. Earn third-party coverage. An independent page saying it carries more weight with this engine than your page saying it; the outreach method in the next section tells you which pages to pitch.

The Search Engine Land analysis above points the same direction from a different dataset.

How do I find the sites I should get mentioned on to appear in AI answers?

Read them out of the answers themselves: the sources engines retrieve for your prompts, ranked by how often they are pulled, are the outreach queue. With per-answer source data the process is mechanical:

1. Run your priority prompts daily across the engines you care about, capturing each answer’s full source list.
2. Aggregate the retrieved URLs over a couple of weeks, counting distinct answers per URL.
3. Drop your own domain and your competitors’ domains from the list.
4. Rank what remains by retrieval count. The result is a ranked list of third-party pages engines already trust for your category, most-retrieved first, and the top of it is your pitch list. Getting named on a page engines already pull is the shortest path into answers, because the retrieval step is already solved.

In our own tracking this correlates strongly: across 20 brands we measured, presence on more of the most-retrieved third-party pages tracked with higher brand visibility at r=0.88. That is correlation on one corpus, so treat the direction as the finding: the pages to pitch are the ones the engines already read, and per-answer source data (what the API returns) tells you exactly which those are.

All figures in this post are from cloro’s AI Search Index (Edition 01, July 2026), a recurring measurement of AI citation behavior on a fixed, balanced 735-prompt basket, and from cloro’s internal retrieval calibration (249 pages, 27 features) and tracked prompt corpus. Correlational throughout; see each study for methodology.