Simulated DAG models may exhibit properties that, perhaps inadvertently,\nrender their structure identifiable and unexpectedly affect structure learning\nalgorithms. Here, we show that marginal variance tends to increase along the\ncausal order for generically sampled additive noise models. We introduce\nvarsortability as a measure of the agreement between the order of increasing\nmarginal variance and the causal order. For commonly sampled graphs and model\nparameters, we show that the remarkable performance of some continuous\nstructure learning algorithms can be explained by high varsortability and\nmatched by a simple baseline method. Yet, this performance may not transfer to\nreal-world data where varsortability may be moderate or dependent on the choice\nof measurement scales. On standardized data, the same algorithms fail to\nidentify the ground-truth DAG or its Markov equivalence class. While\nstandardization removes the pattern in marginal variance, we show that data\ngenerating processes that incur high varsortability also leave a distinct\ncovariance pattern that may be exploited even after standardization. Our\nfindings challenge the significance of generic benchmarks with independently\ndrawn parameters. The code is available at\nhttps://github.com/Scriddie/Varsortability.\n