The AI Citation Study: How ChatGPT, Claude, Gemini and Perplexity Pick Who to Cite

Perplexity is built on retrieval, and it shows: 60.6% of its answers cite at least one specific source. Gemini sits in the middle at 43.5%. ChatGPT and Claude are twins at the bottom, citing in 32.6% and 32.4% of answers.

That gap decides where your effort pays off. On Perplexity, there is a citation slot to win in most answers. On ChatGPT, two answers out of three contain no link at all: there is nothing to rank for, because the model is answering from its training memory.

So the same content strategy produces completely different results per engine. Perplexity and Gemini reward fresh, crawlable, source-shaped pages. ChatGPT and Claude mostly reward being known: being in the training data, the entity graph, and the places the model absorbed before the chat ever started.

We tagged 265 losing answers with a reason. The single biggest, at 34%, is brutal and mostly invisible to SEO tools: the AI never searched the web at all. It answered from memory and recommended whatever brands it absorbed during training. Your page never had a chance to compete, because there was no retrieval step to compete in.

Second, at 23%: the engine searched, but never pulled your page into the answer set. Classic retrieval failure: wrong keywords on the page, weak topical authority, or the page simply is not in the index the engine searches.

Third, at 17%: forum consensus won. The engine found Reddit threads and community posts and trusted the crowd over any brand's own site.

And fourth, at 14%, the quiet heartbreaker: the engine cited your page as a source, then recommended someone else in the same answer. Being cited is not being chosen.

When an engine did cite a competing source, we logged what kind of site won. Brand and company sites took the most citation slots, followed by Google's own official pages, with Reddit close behind.

Put those together and AI engines lean on two pillars. Official sources for facts, and community consensus for recommendations. Reddit alone out-cited YouTube, Wikipedia, and every social publishing platform in our sample, and came within two citations of Google's own official pages.

The practical move is uncomfortable for most brands: you need a genuine presence where the consensus forms, not just a better-optimized page. The engines are reading the room, and the room is a forum.

Three plays follow directly from the loss data.

One: split your strategy by engine. Perplexity and Gemini are retrieval games: win them with pages shaped like sources, clear claims, real data, clean structure. ChatGPT and Claude are reputation games: win them with entity presence, consistent naming, and being talked about in the places models train on.

Two: fix retrieval before polish. 23% of tagged losses were pages that never entered the answer set. Before rewriting copy for the hundredth time, make sure the page matches the exact question wording and is reachable, indexable, and specific.

Three: treat communities as a ranking surface. Forum consensus decided 17% of tagged losses. A helpful, real presence in the communities where your buyers ask questions is now direct answer-engine optimization.