diff --git a/klausur-service/backend/grid_editor_helpers.py b/klausur-service/backend/grid_editor_helpers.py
index 90c55b9..a1831c3 100644
--- a/klausur-service/backend/grid_editor_helpers.py
+++ b/klausur-service/backend/grid_editor_helpers.py
@@ -1424,6 +1424,8 @@ def _build_zone_grid(
     # Split word boxes that straddle column boundaries (e.g. "sichzie"
     # spanning Col 1 + Col 2).  Must happen after column detection and
     # before cell assignment.
+    # Keep original words for colspan detection (split destroys span info).
+    original_zone_words = zone_words
     if len(columns) >= 2:
         zone_words = _split_cross_column_words(zone_words, columns)
 
@@ -1431,11 +1433,11 @@ def _build_zone_grid(
     cells = _build_cells(zone_words, columns, rows, img_w, img_h)
 
     # --- Detect colspan (merged cells spanning multiple columns) ---
-    # A word-block that extends across column boundaries indicates a merged
-    # cell (like Excel cell-merge).  Detect these and replace the split
-    # cells with a single spanning cell.
+    # Uses the ORIGINAL (pre-split) words to detect word-blocks that span
+    # multiple columns.  _split_cross_column_words would have destroyed
+    # this information by cutting words at column boundaries.
     if len(columns) >= 2:
-        cells = _detect_colspan_cells(zone_words, columns, rows, cells, img_w, img_h)
+        cells = _detect_colspan_cells(original_zone_words, columns, rows, cells, img_w, img_h)
 
     # Prefix cell IDs with zone index
     for cell in cells: