diff --git a/narwhals/_compliant/group_by.py b/narwhals/_compliant/group_by.py
index d9271170be..ead3ee9674 100644
--- a/narwhals/_compliant/group_by.py
+++ b/narwhals/_compliant/group_by.py
@@ -108,20 +108,19 @@ def _temporary_name(key: str) -> str:
             key_str = str(key)  # pandas allows non-string column names :sob:
             return f"_{key_str}_tmp{'_' * (tmp_name_length - len(key_str) - 5)}"
 
-        output_names = _evaluate_aliases(compliant_frame, keys)
-
+        keys_aliases = [expr._evaluate_aliases(compliant_frame) for expr in keys]
         safe_keys = [
             # multi-output expression cannot have duplicate names, hence it's safe to suffix
             key.name.map(_temporary_name)
             if (metadata := key._metadata) and metadata.expansion_kind.is_multi_output()
             # otherwise it's single named and we can use Expr.alias
-            else key.alias(_temporary_name(new_name))
-            for key, new_name in zip(keys, output_names)
+            else key.alias(_temporary_name(new_names[0]))
+            for key, new_names in zip(keys, keys_aliases)
         ]
         return (
             compliant_frame.with_columns(*safe_keys),
             _evaluate_aliases(compliant_frame, safe_keys),
-            output_names,
+            list(chain.from_iterable(keys_aliases)),
         )
 
 
diff --git a/tests/frame/group_by_test.py b/tests/frame/group_by_test.py
index ea30ff90be..d8e22e471c 100644
--- a/tests/frame/group_by_test.py
+++ b/tests/frame/group_by_test.py
@@ -558,14 +558,19 @@ def test_group_by_raise_drop_null_keys_with_exprs(
 
 
 def test_group_by_selector(constructor: Constructor) -> None:
-    data = {"a": [1, 1, 1], "b": [4, 4, 6], "c": [7.5, 8.5, 9.0]}
+    data = {
+        "a": [1, 1, 1],
+        "b": [4, 4, 6],
+        "c": ["foo", "foo", "bar"],
+        "x": [7.5, 8.5, 9.0],
+    }
     result = (
         nw.from_native(constructor(data))
-        .group_by(nw.selectors.by_dtype(nw.Int64))
-        .agg(nw.col("c").mean())
+        .group_by(nw.selectors.by_dtype(nw.Int64), "c")
+        .agg(nw.col("x").mean())
         .sort("a", "b")
     )
-    expected = {"a": [1, 1], "b": [4, 6], "c": [8.0, 9.0]}
+    expected = {"a": [1, 1], "b": [4, 6], "c": ["foo", "bar"], "x": [8.0, 9.0]}
     assert_equal_data(result, expected)