Does AI-written content get cited?

The cheapest way to produce an article is to have a language model write it. So the pitch writes itself: if AI decides the answers now, feed it content that AI wrote. A whole cottage industry sells exactly this. The measured result runs the other way.

The appeal is easy to understand. A human writer costs a few hundred euros a page and takes days. A model does it in seconds for cents. And the volume is already enormous: 52% of new articles published online are now AI-generated, a share that first passed human writing in November 2024 (Graphite). The internet is, by article count, already more machine-written than human-written.

So you would expect AI answer engines — trained on and searching that same web — to quote mostly machine text. They do not.

The number that settles it

Hold the two figures next to each other. AI content is the majority of what gets published and the minority of what gets cited. The pool the models draw from is more than half machine-written, yet more than four out of five pages they actually quote came from a person. Producing at volume does not buy you a place in the answer. If anything it buries you in the half the models skip over.

Why the machines avoid machine text

The answer sits one layer down, in ordinary search ranking. In the same study, 86% of the articles ranking in Google Search are human-written, and only 14% are AI-generated. Near the top the gap widens: just 7% of the pages ranking first are machine-made (Graphite). When an AI-written and a human-written page compete for the same keyword, the human one ranks higher, and the difference is statistically firm.

Answer engines lean on the pages that already rank and already carry trust signals. The ranking disadvantage of thin AI content therefore carries straight through into citations. The model is not running an AI-detector and rejecting robots; it is reaching for the sources that look authoritative, and mass-produced content rarely does.

What this means if you are a team of five

The volume game was never one you could win anyway. You will not out-publish a content farm running a model around the clock, and now the data says the farm is playing a losing hand regardless. That is good news, because it points you at the one thing you have that the farm does not: the work itself.

A model can assemble a generic page on "how to choose a heating contractor" in seconds, and so can a thousand other sites, which is precisely why none of them get cited. What it cannot write is the real job you finished last month — the actual quote, the constraint you hit, the number it came in at, the mistake you would not make again. That page exists only if you write it, and it is the kind of specific, first-hand answer the models keep reaching for.

So the useful order of operations is unglamorous. Use AI where it earns its keep: a first draft, a cleanup pass, a title you were stuck on. Then put your own facts, numbers and judgement into it, because that is the part that gets quoted. One page that only you could have written beats fifty the model could have written for anyone.

The shortcut everyone is selling — generate it all, publish it all — optimises for the one metric that does not convert into a citation. The slower path is also the cheaper one in the end, because a page that gets named keeps working, and a page in the ignored 18% never did anything at all.