Which Sources Do AI Engines Cite? Citation Tendencies by Engine

Same question, different sources

Ask generative engines the same question and the sources behind the answers from ChatGPT, Perplexity, and Gemini are often different from one another. This is not a coincidence — it happens because each engine connects a different search infrastructure and a different evidence-selection policy behind its answers. Generative engines generally gather documents that are semantically close to the question (search), pick the evidence to use directly in the answer (evidence selection), and then synthesize it into sentences (see What content does AI cite). Even when they go through the same three steps, which index they use, how much they value freshness, and whether they label sources all differ by engine, so the results diverge.

This piece qualitatively summarizes the general tendencies observed in public documentation and the GEO paper ↗.^[1] Concrete figures such as per-engine citation rates or rankings are not officially published and they fluctuate with the query and the moment, so here we deal only with direction rather than numbers.

Citation tendencies known per engine

Each engine's citation style shows up fairly consistently in its product characteristics.

Perplexity is well known for exposing numbered sources next to each answer sentence. Because it searches in real time for pages related to the question to gather evidence, recent, self-contained paragraphs that bear directly on the topic are an advantage for being cited.

Gemini is connected to the Google Search and Knowledge Graph infrastructure, so it is understood to lean on domains where entities such as companies, products, and people are organized consistently and structured data ↗ (schema.org/Organization ↗) is applied as evidence.^[2][3]

ChatGPT behaves differently depending on whether web search is turned on. When search is on, external source links are attached to the answer, but when it answers from learned knowledge alone, sources may not be labeled. If you are aiming for citations, you have to factor in the very conditions under which web sources get attached to the answer. Here, whether crawlers such as GPTBot ↗ can read the page is also a prerequisite.^[4]

Why the differences arise — causes and effects

The differences in citation tendency stem largely from three axes. The first is the index. Engines that rely on their own index and engines that pull in external search in real time have a different candidate pool to begin with. The second is the freshness weight. For topics that change, such as prices, versions, or policies, the more strongly an engine prefers recent documents, the less it cites older pages. The third is trust signals. How much an engine favors signals like domain reputation, entity consistency, and explicit source labeling differs by engine.

The result these differences create is clear. A page that is frequently cited by one engine may barely appear in another. So judging your overall visibility from the single observation that "it shows up well in one engine" creates blind spots. You have to start from the premise that citation is a phenomenon that happens separately per engine.

Action — measure from a multi-engine view and solidify a common foundation

The fact that engines have different criteria means that measurement and optimization must also be separated by engine. In practice, the following order is realistic.

• Measure several engines together. Throw the same query at multiple engines and look separately at where you are cited and where you are dropped (see multi-engine measurement).
• Solidify the structures that work across the board first. Self-contained paragraphs, clear evidence (definitions, figures, sources), and organized entities raise the probability of citation regardless of engine (see citable content structure).
• Close the per-engine gaps. If you are omitted only in a specific engine, reinforce the signals that engine values (freshness, structuring, explicit sourcing).

In Korea, you have to look not only at the global engines but also at domestic real-usage environments such as Naver-family engines and the characteristics of Korean-language queries. The sources cited when you ask in Korean can differ from English-language tendencies. From this multi-engine, multi-dimensional view, BOIDA sets out to track several engines such as ChatGPT, Claude, Gemini, and Perplexity together with domestic engines, and to provide a visibility diagnosis that combines multiple angles such as exposure, citation, and trust, stably through repeated measurement. The focus is on broadly comparing per-engine gaps rather than on a one-off observation of a single engine.

Summary

• Generative engines go through the same three steps (search → evidence selection → synthesis), but because of differences in index, freshness, and trust signals, the sources they cite diverge by engine.
• Perplexity is known for labeling sources per sentence and favoring fresh web documents, Gemini for organized entities and structured domains, and ChatGPT for treating sources differently depending on whether web search is used.
• Concrete rates and rankings are not published, so it is safer to understand them as direction rather than numbers.
• Optimizing for one engine leaves blind spots. The efficient order is to measure several engines together, first solidify the common foundation of self-contained paragraphs, clear evidence, and organized entities, and then close the per-engine gaps.