diff --git a/src/spatialdata_plot/pl/render.py b/src/spatialdata_plot/pl/render.py
index 5cd6688e..eec55481 100644
--- a/src/spatialdata_plot/pl/render.py
+++ b/src/spatialdata_plot/pl/render.py
@@ -58,6 +58,7 @@
 )
 from spatialdata_plot.pl.utils import (
     _ax_show_and_transform,
+    _check_obs_var_shadow,
     _convert_shapes,
     _datashader_canvas_from_dataframe,
     _decorate_axs,
@@ -370,6 +371,8 @@ def _render_shapes(
     groups = render_params.groups
     table_layer = render_params.table_layer
 
+    _check_obs_var_shadow(sdata, element, col_for_color, render_params.table_name)
+
     sdata_filt = sdata.filter_by_coordinate_system(
         coordinate_system=coordinate_system,
         filter_tables=bool(render_params.table_name),
@@ -766,6 +769,8 @@ def _render_points(
     groups = render_params.groups
     palette = render_params.palette
 
+    _check_obs_var_shadow(sdata, element, col_for_color, table_name)
+
     if isinstance(groups, str):
         groups = [groups]
 
@@ -1687,6 +1692,8 @@ def _render_labels(
     groups = render_params.groups
     scale = render_params.scale
 
+    _check_obs_var_shadow(sdata, element, col_for_color, table_name)
+
     sdata_filt = sdata.filter_by_coordinate_system(
         coordinate_system=coordinate_system,
         filter_tables=bool(table_name),
diff --git a/src/spatialdata_plot/pl/utils.py b/src/spatialdata_plot/pl/utils.py
index 35844c6a..c6ae3350 100644
--- a/src/spatialdata_plot/pl/utils.py
+++ b/src/spatialdata_plot/pl/utils.py
@@ -104,6 +104,38 @@
 }
 
 
+def _check_obs_var_shadow(
+    sdata: SpatialData | None,
+    element_name: str | None,
+    value_to_plot: str | None,
+    table_name: str | None,
+) -> None:
+    """Raise if ``value_to_plot`` exists in both ``table.obs.columns`` and ``table.var_names``.
+
+    Upstream ``_get_table_origins`` uses an ``elif`` chain, so a key that lives in
+    both locations is silently resolved to ``obs`` — masking the user's likely
+    intent of plotting gene expression. Catch this here before any value fetch.
+    Any ``None`` parameter short-circuits the check.
+    """
+    if (
+        value_to_plot is None
+        or table_name is None
+        or element_name is None
+        or sdata is None
+        or table_name not in sdata.tables
+    ):
+        return
+    if table_name not in get_element_annotators(sdata, element_name):
+        return
+    table = sdata.tables[table_name]
+    if value_to_plot in table.obs.columns and value_to_plot in table.var_names:
+        raise ValueError(
+            f"`color={value_to_plot!r}` is ambiguous: it exists in both "
+            f"`table[{table_name!r}].obs.columns` and `table[{table_name!r}].var_names`. "
+            "Rename one of them (or drop the obs column) so the intended source is unambiguous."
+        )
+
+
 def _gate_palette_and_groups(
     element_params: dict[str, Any],
     param_dict: dict[str, Any],
diff --git a/tests/pl/test_utils.py b/tests/pl/test_utils.py
index 0a277f05..ab279d45 100644
--- a/tests/pl/test_utils.py
+++ b/tests/pl/test_utils.py
@@ -422,6 +422,22 @@ def test_color_column_collision_on_annotating_table_raises():
     sdata.pl.render_shapes("s", color="#ffa500")
 
 
+def test_color_key_obs_var_shadow_raises():
+    # regression test for #621
+    pts = PointsModel.parse(pd.DataFrame({"x": [1.0, 2.0], "y": [1.0, 2.0]}))
+    obs = pd.DataFrame({"instance_id": [0, 1], "region": ["pts"] * 2, "GeneA": [0.9, 0.6]}, index=["0", "1"])
+    table = TableModel.parse(
+        AnnData(X=np.zeros((2, 1)), obs=obs, var=pd.DataFrame(index=["GeneA"])),
+        region=["pts"],
+        region_key="region",
+        instance_key="instance_id",
+    )
+    sdata = SpatialData(points={"pts": pts}, tables={"t": table})
+
+    with pytest.raises(ValueError, match=r"'GeneA'.*ambiguous.*obs\.columns.*var_names"):
+        sdata.pl.render_points("pts", color="GeneA", table_name="t").pl.show()
+
+
 def test_explicit_table_name_honored_when_element_has_same_column():
     # regression test for #620: explicit table_name= must not be silently
     # discarded when the element has a same-named column with different values.