Mostly based on results of experienced examiners, endoscopic ultrasound (EUS) has been reported to be highly accurate for locoregional staging of upper gastrointestinal cancers. However, data on interobserver variability among EUS examiners, depending on their experience levels, is sparse. A study was therefore conducted to analyse well-documented videotapes of EUS examinations of 108 patients with resected cancers of the esophagus (n = 55) or stomach (n = 53) in a strictly blinded fashion by 5 examiners, all of whom were experienced in EUS (more than 300 examinations: n = 3, more than 100 examinations: n = 2). Besides the individual accuracy rates in cancer staging, a kappa-statistic was calculated to check for interobserver variability. Under the conditions described, the staging accuracy of all investigators was lower than that usually achieved under clinical routine conditions. The mean T staging accuracy was 41.1 % +/- 9.4 and 46.9 % +/- 5.4 in gastric and esophageal cancers, respectively. For N-staging the respective values were 47.9 % +/- 5.1 (stomach) and 67.7 % +/- 5.4 (oesophagus). Kappa-values were above 0.4 only in the staging of non-invasive esophagogastric tumours of the N0 and T1-category, corresponding to a fairly good agreement among the five investigators. Differences depending on experience levels could not be consistently found. Hence, it can be concluded that endosonographic cancer staging performed in a blinded manner results in a low accuracy and high interobserver variability even among experienced examiners.