diff --git a/jetstream/analysis.py b/jetstream/analysis.py
index 4982ab7c..aa432934 100644
--- a/jetstream/analysis.py
+++ b/jetstream/analysis.py
@@ -1380,11 +1380,13 @@ def run(
                 continue
 
             segment_labels = ["all"] + [s.name for s in self.config.experiment.segments]
-            analysis_length_dates = 1
-            if period == AnalysisPeriod.OVERALL:
-                analysis_length_dates = time_limits.analysis_length_dates
-            elif period == AnalysisPeriod.WEEK:
-                analysis_length_dates = 7
+            # Derive the analysis window length (in days) from TimeLimits so that
+            # normalize_over_analysis_period scales correctly for every period (day=1, week=7,
+            # days28=28, overall=N, preenrollment_week=7, preenrollment_days28=28). All windows
+            # in a period share the same length; start/end are inclusive days since enrollment,
+            # so length = end - start + 1. (TimeLimits has no `analysis_length_dates` attribute.)
+            last_window = time_limits.analysis_windows[-1]
+            analysis_length_dates = last_window.end - last_window.start + 1
 
             for analysis_basis in analysis_bases:
                 segment_data: DataFrame
diff --git a/jetstream/tests/test_analysis.py b/jetstream/tests/test_analysis.py
index 05ecad79..ca553cb0 100644
--- a/jetstream/tests/test_analysis.py
+++ b/jetstream/tests/test_analysis.py
@@ -1791,3 +1791,86 @@ def capturing_bind(thing, deps):
         "Expected at least one subset_metric_table bind call with prerequisites, "
         "but all bind calls had empty deps. The covariate cross-period edge is missing."
     )
+
+
+@pytest.mark.parametrize(
+    ("period", "time_limits_factory_kwargs", "expected_length"),
+    [
+        # OVERALL: single window. TimeLimits has no `analysis_length_dates` attribute, so the
+        # original code raised AttributeError and killed the run entirely.
+        (
+            AnalysisPeriod.OVERALL,
+            {"single_window": True, "analysis_start_days": 0, "analysis_length_dates": 30},
+            30,
+        ),
+        # DAYS_28: time series. The original code fell through to the default of 1, so
+        # normalize_over_analysis_period silently divided by 1 instead of 28.
+        (
+            AnalysisPeriod.DAYS_28,
+            {"single_window": False, "time_series_period": "28_day"},
+            28,
+        ),
+    ],
+)
+def test_run_derives_analysis_length_dates(
+    experiments, monkeypatch, period, time_limits_factory_kwargs, expected_length
+):
+    """analysis_length_dates must be derived from the TimeLimits window for every period so that
+    normalize_over_analysis_period scales correctly. Regression test for both the OVERALL
+    AttributeError and the default-of-1 fall-through for non-WEEK/OVERALL periods."""
+    from mozanalysis.experiment import TimeLimits
+
+    from jetstream.statistics import Summary as JetstreamSummary
+
+    config = AnalysisSpec.default_for_experiment(experiments[0], ConfigLoader.configs).resolve(
+        experiments[0], ConfigLoader.configs
+    )
+
+    metric = Metric(
+        name="active_hours",
+        data_source=DataSource(name="clients_daily", from_expression="test.test"),
+        select_expression="SUM(ah)",
+        analysis_bases=[AnalysisBasis.ENROLLMENTS],
+    )
+    plain_statistic = MagicMock()
+    plain_statistic.params = {}
+    config.metrics = {period: [JetstreamSummary(metric, plain_statistic, [])]}
+
+    common = {"first_enrollment_date": "2024-01-01", "num_dates_enrollment": 8}
+    if time_limits_factory_kwargs.pop("single_window"):
+        time_limits = TimeLimits.for_single_analysis_window(
+            last_date_full_data="2024-04-01", **time_limits_factory_kwargs, **common
+        )
+    else:
+        time_limits = TimeLimits.for_ts(
+            last_date_full_data="2024-04-01", **time_limits_factory_kwargs, **common
+        )
+    assert not hasattr(time_limits, "analysis_length_dates")
+
+    stats_mock = MagicMock()
+    stats_mock.model_dump.return_value = []
+    calc_stats = MagicMock(return_value=stats_mock)
+
+    monkeypatch.setattr("jetstream.analysis.bind", lambda thing, deps: thing)
+    monkeypatch.setattr("jetstream.analysis.Analysis.ensure_enrollments", Mock())
+    monkeypatch.setattr(
+        "jetstream.analysis.Analysis._get_timelimits_if_ready",
+        MagicMock(return_value=time_limits),
+    )
+    monkeypatch.setattr("jetstream.analysis.Analysis.calculate_metrics", MagicMock())
+    monkeypatch.setattr("jetstream.analysis.Analysis.calculate_statistics", calc_stats)
+    monkeypatch.setattr("jetstream.analysis.Analysis.counts", MagicMock(return_value=stats_mock))
+    monkeypatch.setattr("jetstream.analysis.Analysis.save_statistics", MagicMock())
+    monkeypatch.setattr("jetstream.analysis.Analysis.publish_view", MagicMock())
+    monkeypatch.setattr("jetstream.analysis.LocalCluster", MagicMock())
+    monkeypatch.setattr("jetstream.analysis.Client", MagicMock())
+    monkeypatch.setattr("jetstream.analysis.as_completed", Mock(return_value=[]))
+
+    # must not raise AttributeError on time_limits.analysis_length_dates
+    Analysis("test", "test", config).run(current_date=dt.datetime(2020, 1, 1, tzinfo=pytz.utc))
+
+    assert calc_stats.called, f"calculate_statistics was never called for the {period} period"
+    # analysis_length_dates is the 5th positional arg and must be the derived window length,
+    # not the default of 1
+    analysis_length_dates_args = {call.args[4] for call in calc_stats.call_args_list}
+    assert analysis_length_dates_args == {expected_length}