Asking a large language model the same question twice gives different answers, so a result based on one response mixes the model's underlying forecast with random variation between calls. We query the model 500 times for each quarterly inflation forecast in three economies, using target quarters that fall after the model's knowledge cutoff, and average the responses. Averaging lowers forecast error by 7 to 35 percent, and most of the gain arrives within ten repetitions. An exact decomposition separates the error of the average forecast from the removable variation between calls; that share is typically 16 percent of single-response error but reaches 76 percent, and it governs how much averaging helps. A more capable model varies less between calls and gains correspondingly less. Repeated querying should form part of any evaluation in which language model outputs are treated as data, and the number of repetitions should be reported.