Studies Reveal AI Citation Clues

There are no guidelines directly from ChatGPT, Gemini, and other generative AI platforms on how to appear in their answers.

Microsoft’s recent “AEO and GEO” guide offered only commonsense tips.

We’re left with independent research to inform citation optimization tactics. Two recent studies offer helpful takeaways.

• Kevin Indig is an organic search consultant and the former head of SEO for G2, the software review platform. He analyzed 1.2 million ChatGPT results, which contained 18,012 citations.

• Daniel Shashko is a senior search engine optimization specialist with Bright Data, a research firm. He studied 42,971 citations across 520 queries on six platforms: Grok, AI Mode, Perplexity, Gemini, Copilot, and ChatGPT.

He found that Grok delivered 33 citations per query, while ChatGPT averaged just 1.5. Roughly 70% of Google’s AI Mode and Gemini used citations that included embedded #:~:text= fragment, which linked to the exact cited sentence in the answer.

Here are key findings from the studies.

Optimizing Citations

The closer to the top, the better

Both studies found that the platforms tend to cite answers from the top third of pages.

Kevin’s study found that 44.3% of ChatGPT’s citations originated from the first 30% of the page’s text.

Daniel’s study revealed that 74.8% of citations in AI Mode and Gemini appeared in the first half of the page, with 46.1% being in the first 30%. (The other four platforms do not link directly to the cited sentence and do not appear prominently in Daniel’s study.)

The takeaway from both of these studies is clear: make sure to answer the most important question or problem in the first third of your page.

Emphasize brevity

Daniel’s study introduced “atomic facts,” which he defines as “… a self-contained, single-claim sentence that makes sense on its own.”

For AI Mode and Gemini, Daniel found:

• Sentences of 6 to 20 words accounted for 92.4% of citations.

• All citations (100%) were full sentences. No single citation started or ended in the middle of a sentence.

In other words, avoid long introductions or unclear or irrelevant dialogue. Get to the point.

A new free tool tracks the number of “atomic facts” on a page.

No Google overlap

Daniel found just 4.5% of AI Mode’s cited domains appear in Gemini, and just 13.2% of Gemini’s domains are in AI Mode.

The two LLMs appear to follow similar patterns in selecting sources, yet the citations are largely unique.

Citations vs. visibility

The two studies focus solely on citations, not on general visibility, i.e., unlinked references. To optimize the latter, ensure your brand is well-positioned in the training data.