AI citations are not just a source-order game

TL;DR: If you want AI systems to cite you, do not obsess over being first in a source list. Make your page easier to parse, quote, and trust.

Does source order decide who gets cited by AI?

Search Engine Journal’s “AI Citation Test Finds Source Order Matters Less Than It Looks” by Matt G. Southern reports a useful correction to a tempting belief: AI citation behavior is not only a ranking-order problem.

The controlled test found that source order looked important in raw data, but less so after closer analysis. That matters because a lot of AI search advice has drifted toward a simple idea: get mentioned early, get cited more. Nice if true. Too clean to trust.

Southern’s report points to a messier pattern. Order can appear to matter, but the underlying format and rewrite structure can shift how citation credit gets handed out. In plain English: models are not just reading a list from top to bottom and assigning citations like a search engine results page with footnotes.

That should change how operators think about “AI visibility.” Traditional SEO still cares about crawlability, authority, links, and query fit. AI answers add another layer: whether the model can lift a clean claim, associate it with the right entity, and decide that this source deserves attribution in the generated answer.

Ken Ashe runs Lucky Domains, which works on SEO and search visibility for websites, so this is not an abstract debate for me. The practical question is whether your content survives being chopped, summarized, rewritten, and compared against other sources.

What does “structured rewrite” change?

The interesting part of Southern’s report is not that source order is irrelevant. It is that structured rewrites changed how citation credit was distributed.

That phrase should get the attention of anyone publishing reference content, product pages, research explainers, comparison pages, or documentation. If a rewrite changes citation distribution, then the model is sensitive to presentation. Not just facts. Presentation.

That does not mean stuffing schema everywhere or turning every post into a FAQ farm. It means the page should make claims in a way that can be separated from fluff. Clear definitions. Named entities. Direct answers. Primary evidence. Dates where dates matter. No cute preamble before the actual answer. No burying the most useful sentence under five paragraphs of brand voice.

The catch is that “structured” can get overdone. Pages written only for extraction often read like commodity content. AI systems may quote them, but humans may bounce. The better target is dual-use writing: clear enough for a model to parse, specific enough for a human to trust, and original enough that another site cannot replace you with a bland paraphrase.

What should publishers test now?

I would not rebuild a content program around one citation test. Search Engine Journal reported a controlled result, not a universal law of AI search. We also do not get enough from the summary alone to make hard claims about every model, query class, or citation surface.

But the direction is useful. Stop treating AI citations as a black box where the only move is “rank higher.” Run your own tests. Ask common buyer, research, and support questions in the AI systems your audience uses. Track which pages get cited, which competitors appear, and which claims get repeated without attribution.

Then rewrite a small set of pages. Keep the same factual substance, but change the structure. Move the direct answer up. Add concise definitions. Put evidence near the claim. Clarify who did what, when, and why it matters. Remove filler that forces the model to infer your point.

Measure again. Not once. Across query variants.

The operator move is simple: pick five pages that should be cited but are not, rewrite them for extractable clarity, and test them against the same prompts every week. The catch most readers miss is that AI citation optimization is not only about pleasing the model. It is about making your best evidence hard to misread, hard to detach from your brand, and easy to use in an answer.