Perplexity, Gemini, Claude: how citation patterns differ by engine

The four engines differ in where they show sources and in which pages they favor. Filling all four spots at once with a single optimization is hard. Perplexity openly lays out numbered source cards above its answer, while ChatGPT, when search is turned on, quietly slips links into the sentences of the body text. Gemini inherits Google's search index and AI Overviews almost as-is, and Claude, which by default builds answers from its training data, cites external sources most sparingly.

Even for the same question, the surface where citations appear differs by engine. Search Engine Land's AI answer engine citation analysis concludes that because of this difference, optimizing for just one engine cannot cover every surface. You first have to know which engine's spot is empty before you can set the order in which to work.

At a glance: a comparison of citation patterns across the four engines

Throw a single question like "where's a good place for whitening treatments in Gangnam" at all four engines at once and the differences show up immediately. The source placement, number of sources, and answer length all differ.

The table makes the four engines look very different. A closer look shows which signals they share and which differ by engine. Below, we walk through, engine by engine, what actually appears on screen and what you need to change to target each spot.

ChatGPT: citations turn on only when search is on

ChatGPT's citations depend on the mode. The default response builds its answer from training data, so no sources appear. Only when you turn on search mode do real-time results attach to the body as links.

The way they attach is also distinctive. Rather than being set apart in separate cards, the links are integrated into the sentences. You encounter the supporting link as you read the answer. OpenAI Help Center's ChatGPT search feature documentation also explains that it uses both in-sentence links in the body and a separate source panel. Usually 3 to 5 attach per answer, and for Korean-language questions, authoritative outlets like wikis, the press, and official documents get called up first.

To be cited here, your content needs two things.

• The first sentence or two of the body should already be the answer: it needs to be in a form that ChatGPT can slot straight into its answer flow. Content with a drawn-out introduction is less likely to be selected for citation.
• It should contain structured data: schema.org Organization, FAQPage, and Article markup need to be reliably in place for the engine to trust and pick up the page.

A common mistake is treating "ChatGPT with search on" and "ChatGPT with it off" as the same thing. The two behave completely differently in citation terms. That's why measurement is only meaningful when search mode is turned on.

Perplexity: sources show up before you even read

Of the four engines, Perplexity presents sources most overtly. It lays out numbered cards at the top of the answer and, via footnotes, connects each sentence of the body 1:1 to the source number it came from. Users face the source cards before they read the answer.

The pages it prefers are also clear. Semrush's Perplexity citation research points out that it most strongly favors freshness, structure, and direct-answer paragraphs. The number of cards is commonly 5–10 depending on question difficulty, rising for research-type questions. Because every question triggers a real-time search, a freshly published page has the highest probability among the four engines of making it onto a card within 24–72 hours.

For content aiming at this spot, compression is key. The structure should be one where the subheading meshes with the user's question in natural language, and the very next one or two sentences immediately become the answer. A single, tightly cut direct-answer block is more advantageous for getting onto a card than one long explanatory paragraph. AdExchanger's Perplexity referral traffic analysis shows that the referral click-through rate of pages featured in a card comes out higher than regular search results for the same question. A card brings both exposure and actual visits.

Gemini: do your SEO well and it follows naturally

Gemini is effectively an extension of Google Search. Its answers appear as cards beside the body and in the AI Overviews source panel, and the criteria for choosing sources overlap almost entirely with Google's index signals of authority, relevance, and freshness. The E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) and structured data that Google has long pushed apply directly to Gemini's citation spots.

The data shows the scale of this. Search Engine Land's AI Overviews citation research reports that AI Overviews appeared in 5.5M of the 21.9M Google searches analyzed (25.11%), and that most of those sources were domains already ranking high in SEO. A page that isn't even indexed won't appear in Gemini either. So Gemini's spot isn't a separate project you launch. Widen the SEO you're already doing by one layer and it follows naturally.

So the order is simple. First get the SEO fundamentals in shape (indexing, internal links, page speed, E-E-A-T), then layer FAQ, Article, and HowTo markup and direct-answer paragraphs on top. One thing to add: the share of Gemini answers shown on mobile is growing fast, so mobile loading speed and Core Web Vitals carry additional weight. It's an easy point to miss if you're only looking at desktop.

Claude: one page won't do it; the whole domain matters

Claude produces citations most sparingly. Its default answer is built from training data, and external sources enter the body only when the web search tool is on. Its answers are long and its reasoning dense, but at 0–3 citations per question, it has the fewest.

The criteria for choosing what to cite are also different. Anthropic's Claude Citations feature documentation defines this as a "tool-augmented response." It calls up external sources only when the question has strong search intent or the search tool is turned on. And it weighs the authority consistency of the domain as a whole more heavily than a freshly published one-off page. Wired's Claude citation behavior analysis also reports that Claude weighs domain-level authority signals more heavily than ChatGPT or Perplexity.

So Claude's spot can't be won by editing a single piece of content; you earn it by making your whole domain coherent. In practice, there are two things to take care of.

1. Use consistent brand naming: if basic information like company name, service name, and founding year varies across your own site, press coverage, and social media, the engine can't tie it into a single entity. Start by making those names and details consistent.
2. Steadily accumulate original assets worth citing: material that others can pull in as evidence, such as your own measurement data or research, needs to build up on the domain.

Can you satisfy all four engines at once: the order PION follows

Filling all four engines with a single effort is hard. Still, there are shared signals. Direct-answer paragraphs, clear subheadings, structured data, and consistent brand naming all help on every engine. Establishing these first benefits every engine.

What differs by engine is freshness and source preferences. Perplexity is sensitive to fresh pages 24–72 hours after publication, Gemini weighs Google index authority, and Claude weighs long-term domain authority. ChatGPT wavers somewhere in between depending on the mode and the question.

As an agency, the order we follow in client work is this. First, we throw the category's core cluster of questions directly at all four engines to measure the current state of each engine's citation spot. Then we start with the engine that's emptiest. Which engine to tackle first is decided by the measured numbers, not by a textbook priority order.

In Korean-language categories, the Perplexity and ChatGPT spots are usually empty first. Gemini recovers relatively quickly if there's some SEO foundation. Claude only shows up reliably after domain authority has built up, so we don't push it for this month's results; we run it separately as 6–12 month cumulative asset work. Because even the same effort yields different results depending on where you put it.

Frequently asked questions

Comparing ChatGPT, Perplexity, Gemini, and Claude citation patterns

The four engines differ in where they show sources, their diversity, and their update cycles. Perplexity lays sources out most overtly as numbered cards above the answer, while ChatGPT incorporates them as in-body links in search mode. Gemini appears as a card-style source panel like Google AI Overviews, and Claude shows restrained citations only when external search is turned on.

How does AEO optimization differ for each AI engine?

There are four shared signals: direct-answer paragraphs, clear subheadings, structured data, and consistent brand naming. All four help on every engine. What differs by engine is freshness (Perplexity, 24–72 hours), Google index authority (Gemini), long-term domain authority (Claude), and search mode (ChatGPT); it's most efficient to measure each engine's current state and then fill the emptiest spot first.

Is optimizing for Perplexity different from ChatGPT?

Yes. Perplexity triggers a real-time search on every question and connects sources 1:1 via numbered cards above the answer, so freshness and short, compressed direct-answer paragraphs are key. ChatGPT turns citations on according to the search toggle, and links appear within the body, so structured markup and paragraph design that's easy to slot into an answer come first.

Is Gemini's citation spot separate work from SEO?

No, it's not separate work. Gemini inherits Google's search index and AI Overviews almost as-is, so the SEO fundamentals (indexing, internal links, E-E-A-T, structured data) apply directly to the citation spot. For the same reason, a page that isn't even indexed won't appear in Gemini either.

Can you quickly increase Claude's citation spots?

In the short term, it's hard. Claude's default is a training-data-based response, and even its external-search citations weigh long-term domain authority more heavily, so rather than fixing a single page, unifying brand naming across your site, the press, and social media and steadily publishing your own measurement data and research is more effective. It's realistic to run it as asset work on a 6–12 month cycle.