When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models

Language models now draft literature reviews and choose which papers to cite, and the checks applied to their output mostly ask whether the references exist. We show that a second failure remains even when every reference is real. Eleven language models from three vendors concentrate citations onto the same papers, and eight domain experts shown the same candidates show no concentration beyond their null. Our benchmark replaces retrieval with uniform random exposure, hides prestige metadata, and excludes references that match no shown record. Each of 120 prompts shows a uniformly random panel of 30 real knowledge-distillation papers with real titles and abstracts but fabricated authors, randomly reassigned years, and no venue or citation count. The model may cite at most 10, and every run is compared with uniform selection on the same panel and number of citations. Every model concentrates its citations. The top decile of papers takes 23.3--30.2\% of citations against 15.6\% under the null. The eleven preferences are also nearly one preference. A single principal component carries 68--73\% of the variance across the eleven maps, so the best cross-fitted mixture of the eleven models still retains 54.6\% of the average single model's excess. The experts, selecting from 52 of the same panels under the same cap, stay at their null. The shared preference follows the text. Paraphrase, content-slot crossover, and design resampling attribute 89\% of one model's preference variance to what the titles and abstracts say. When the benchmark runs for 12 rounds with model-written papers diluting the catalogue, the shared preference does not transfer to the generated papers. On average a real paper is nonetheless cited on 63\% of its appearances by round 11 against 33\% at round 0, as a fixed preference predicts when real papers grow rare. Within the benchmark, the narrowing is neither a retrieval problem nor a prestige problem, and no check for nonexistent references will find it.