diff --git a/src/modelskill/matching.py b/src/modelskill/matching.py
index 38e5c913a..30e0e3709 100644
--- a/src/modelskill/matching.py
+++ b/src/modelskill/matching.py
@@ -378,6 +378,16 @@ def _match_single_obs(
         return None
     matched_data.attrs["weight"] = obs.weight
 
+    # The dummy is extracted from the whole observation, but its mean must be the
+    # mean of what is compared, which is known only after matching.
+    obs_mean = float(matched_data["Observation"].mean())
+    for m in models:
+        if isinstance(m, DummyModelResult) and m.strategy == "mean":
+            matched_data[m.name] = matched_data[m.name].copy(
+                data=np.full(matched_data[m.name].shape, obs_mean)
+            )
+            raw_mod_data[m.name].data[m.name][:] = obs_mean
+
     return Comparer(matched_data=matched_data, raw_mod_data=raw_mod_data)
 
 
diff --git a/src/modelskill/model/dummy.py b/src/modelskill/model/dummy.py
index 58a33c790..ddc99dcb6 100644
--- a/src/modelskill/model/dummy.py
+++ b/src/modelskill/model/dummy.py
@@ -22,7 +22,12 @@ class DummyModelResult:
     name : str, optional
         The name of the model result, by default 'dummy'
     strategy : str, optional
-        The strategy to use, 'mean' uses the mean of the observation, 'constant' uses the value given in data, by default 'constant'
+        The strategy to use, by default 'constant'
+
+        - 'mean': the mean of the observation values that are compared. In `match()`
+          this is the mean over the time steps shared with the other models, so the
+          dummy has r2 = 0. For a long-term mean, pass it as `data`.
+        - 'constant': the value given in `data`
 
     Examples
     --------
diff --git a/tests/test_match.py b/tests/test_match.py
index 0d0032269..7ff7f7089 100644
--- a/tests/test_match.py
+++ b/tests/test_match.py
@@ -918,13 +918,32 @@ def test_compare_model_vs_dummy_for_track(mr1, o3):
 
     cmp2 = ms.match(obs=o3, mod=[mr1, mr])
 
-    # not identical to above since it is evaluated on a subset of the data
-    assert cmp2.score()["dummy"] == pytest.approx(1.225945)
+    # the mean is taken over the subset compared with SW_1, so r2 is still 0
+    assert cmp2.score(metric="r2")["dummy"] == pytest.approx(0.0)
 
     # better than dummy 🙂
     assert cmp2.score()["SW_1"] == pytest.approx(0.3524703)
 
 
+def test_dummy_mean_is_mean_of_compared_observations():
+    time = pd.date_range("2020-01-01", periods=10, freq="h")
+    obs = ms.PointObservation(
+        pd.Series(np.arange(10.0), index=time), name="obs", x=0.0, y=0.0
+    )
+    # model covers only the last 4 time steps
+    mod = ms.PointModelResult(
+        pd.Series(np.arange(6.0, 10.0) + 0.5, index=time[6:]), name="mod"
+    )
+    dummy = ms.DummyModelResult(name="dummy", strategy="mean")
+
+    cmp = ms.match(obs=obs, mod=[mod, dummy])
+
+    assert cmp.n_points == 4
+    assert cmp.data["dummy"].values == pytest.approx([7.5] * 4)
+    assert cmp.score(metric="r2")["dummy"] == pytest.approx(0.0)
+    assert cmp.raw_mod_data["dummy"].values == pytest.approx([7.5] * 10)
+
+
 def test_match_obs_model_pos_args_wrong_order_helpful_error_message():
     # match is pretty helpful in converting strings or dataset
     # so we need to use a ModelResult to trigger the error