diff --git a/doc/source/whatsnew/v0.21.0.txt b/doc/source/whatsnew/v0.21.0.txt
index ac0e960a348b6..2f2089b4f8ad7 100644
--- a/doc/source/whatsnew/v0.21.0.txt
+++ b/doc/source/whatsnew/v0.21.0.txt
@@ -324,6 +324,7 @@ Indexing
 - Bug in reindexing on an empty ``CategoricalIndex`` (:issue:`16770`)
 - Fixes ``DataFrame.loc`` for setting with alignment and tz-aware ``DatetimeIndex`` (:issue:`16889`)
 - Avoids ``IndexError`` when passing an Index or Series to ``.iloc`` with older numpy (:issue:`17193`)
+- Allow unicode empty strings as placeholders in multilevel columns in Python 2 (:issue:`17099`)
 
 I/O
 ^^^
diff --git a/pandas/core/frame.py b/pandas/core/frame.py
index 027a427555253..94cce1b4d05b5 100644
--- a/pandas/core/frame.py
+++ b/pandas/core/frame.py
@@ -2134,10 +2134,18 @@ def _getitem_multilevel(self, key):
                 result = self._constructor(new_values, index=self.index,
                                            columns=result_columns)
                 result = result.__finalize__(self)
+
+            # If there is only one column being returned, and its name is
+            # either an empty string, or a tuple with an empty string as its
+            # first element, then treat the empty string as a placeholder
+            # and return the column as if the user had provided that empty
+            # string in the key. If the result is a Series, exclude the
+            # implied empty string from its name.
             if len(result.columns) == 1:
                 top = result.columns[0]
-                if ((type(top) == str and top == '') or
-                        (type(top) == tuple and top[0] == '')):
+                if isinstance(top, tuple):
+                    top = top[0]
+                if top == '':
                     result = result['']
                     if isinstance(result, Series):
                         result = self._constructor_sliced(result,
diff --git a/pandas/tests/test_multilevel.py b/pandas/tests/test_multilevel.py
index 0b2dc9ba70f03..a765e2c4ca1bf 100644
--- a/pandas/tests/test_multilevel.py
+++ b/pandas/tests/test_multilevel.py
@@ -1675,24 +1675,31 @@ def test_int_series_slicing(self):
         expected = self.ymd.reindex(s.index[5:])
         tm.assert_frame_equal(result, expected)
 
-    def test_mixed_depth_get(self):
+    @pytest.mark.parametrize('unicode_strings', [True, False])
+    def test_mixed_depth_get(self, unicode_strings):
+        # If unicode_strings is True, the column labels in dataframe
+        # construction will use unicode strings in Python 2 (pull request
+        # #17099).
+
         arrays = [['a', 'top', 'top', 'routine1', 'routine1', 'routine2'],
                   ['', 'OD', 'OD', 'result1', 'result2', 'result1'],
                   ['', 'wx', 'wy', '', '', '']]
 
+        if unicode_strings:
+            arrays = [[u(s) for s in arr] for arr in arrays]
+
         tuples = sorted(zip(*arrays))
         index = MultiIndex.from_tuples(tuples)
-        df = DataFrame(randn(4, 6), columns=index)
+        df = DataFrame(np.random.randn(4, 6), columns=index)
 
         result = df['a']
-        expected = df['a', '', '']
-        tm.assert_series_equal(result, expected, check_names=False)
-        assert result.name == 'a'
+        expected = df['a', '', ''].rename('a')
+        tm.assert_series_equal(result, expected)
 
         result = df['routine1', 'result1']
         expected = df['routine1', 'result1', '']
-        tm.assert_series_equal(result, expected, check_names=False)
-        assert result.name == ('routine1', 'result1')
+        expected = expected.rename(('routine1', 'result1'))
+        tm.assert_series_equal(result, expected)
 
     def test_mixed_depth_insert(self):
         arrays = [['a', 'top', 'top', 'routine1', 'routine1', 'routine2'],