The State of AEO

Four engines, four different answers

For each query we took the set of domains each engine cited and measured how much any two engines overlapped. Across 99 engine pairings over 30 queries, the mean overlap was 0.16. In plain terms, pick two AI engines, ask them the same question, and roughly one in six of the sources behind their answers is shared. The other five-sixths are different.

The answer you get depends nearly as much on which AI you ask as on what you ask.

For anyone trying to optimize for AI answers, this is the whole problem. There is no single ranking to climb. Being the source ChatGPT trusts buys you almost nothing with Gemini.

Every engine has a citation personality

The engines do not just disagree on which sources to cite. They disagree on how widely to look. Gemini drew on 310 distinct domains; OpenAI on 80, leaning hard on a short list of big tech-review publishers.

OpenAI behaves like a curated shortlist; Gemini like a wide net; Perplexity leans on video. If you only optimize for one engine's taste, you are invisible to the others.

YouTube is the most-cited source on the answer web

Across the whole corpus, the single most-cited domain was not a news site or Wikipedia. It was youtube.com, cited in 21 of 30 queries. Video, review media, and personal-finance explainers dominate the top of the table.

Wikipedia, the backbone of the old search-result page, was just 1.1 percent of citations. Forums and Reddit, despite the discourse about them feeding the models, were under 1 percent.

It is a long tail, not a winner-take-all

You might expect AI answers to collapse onto a handful of trusted megasites. They do not, at least not yet. The 30 queries pulled in 546 distinct domains, and the top 10 accounted for only 16.7 percent of all citations. The citation pool is wide and shallow. That is good news for anyone outside the incumbent set: the door is not shut. It is also why the engines disagree so much. With a wide pool and no shared ranking, each model picks its own slice.

What it means

• For brands and publishers. There is no "rank #1 in AI." Track citations across every engine your audience uses, because a win on one is not a win on another. Measure, do not assume.
• For the open web. The wide tail means visibility is still winnable from outside the incumbent set, but the dominance of video and a few explainer publishers shows where the models currently place their trust.
• For the agentic web. If agents shop and decide by reading these answers, the 16 percent agreement means the agent's recommendation depends on which model it runs on. The discovery layer is not neutral, and it is not consistent.

Methodology & honest caveats

• Firsthand and open. 30 representative queries across product, how-to, technical, health, finance, and comparison categories, run against Claude, Perplexity, OpenAI, and Gemini via their citation-returning APIs. Every cited URL was captured. The tool and query set are open source.
• Read-only. We send queries and read the answers. We do not fetch the cited pages or probe anything.
• A snapshot, not a verdict. Answer engines are non-deterministic; the same query can cite different sources run to run. The value is in the pattern across many queries and in re-running over time, not in any single answer.
• Citations, not rankings. Being cited is a presence signal, not a position. We count where a domain appears, not how prominently.