diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6ebf9d0f..78da8414 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -50,6 +50,12 @@
   terms are no longer indexed (#322). SQLite property values stored by earlier
   versions were only lowercased, so values with extra whitespace or decomposed
   Unicode may no longer match; re-ingest such databases to rebuild the indexes.
+- Term normalization now lowercases before applying NFC, so normalizing an
+  already-normalized term no longer changes it. This affects only terms where
+  lowercasing yields a sequence that composes differently (e.g. an uppercase
+  letter followed by a combining mark, such as `T̈`). Indexes stored by
+  earlier versions keep such terms in the old form, so they may no longer
+  match; re-ingest to rebuild the indexes.
 
 ### 0.5.0 (May 26)
 
diff --git a/src/typeagent/knowpro/common.py b/src/typeagent/knowpro/common.py
index 0eebfc48..7f85788d 100644
--- a/src/typeagent/knowpro/common.py
+++ b/src/typeagent/knowpro/common.py
@@ -10,12 +10,17 @@
 def normalize_term(term: str) -> str:
     """Canonical form of an index term, shared by all storage backends.
 
-    Strips surrounding whitespace, applies NFC normalization, collapses runs
-    of whitespace to a single space, and lowercases. Every term index must
+    Strips surrounding whitespace, lowercases, applies NFC normalization, and
+    collapses runs of whitespace to a single space. Every term index must
     apply this on both write and lookup so backends agree on term identity.
+
+    Lowercasing comes before NFC because lowercasing can produce a sequence
+    that composes differently (e.g. "T\u0308" -> "t\u0308" -> "\u1e97"); in the
+    reverse order the result isn't idempotent, so re-normalizing a stored term
+    (as deserialization does) would change it.
     """
-    term = unicodedata.normalize("NFC", term.strip())
-    return re.sub(r"\s+", " ", term).lower()
+    term = unicodedata.normalize("NFC", term.strip().lower())
+    return re.sub(r"\s+", " ", term)
 
 
 def is_search_term_wildcard(search_term: SearchTerm) -> bool:
diff --git a/tests/test_term_normalization.py b/tests/test_term_normalization.py
index 9dda2847..cfb2c6d7 100644
--- a/tests/test_term_normalization.py
+++ b/tests/test_term_normalization.py
@@ -35,6 +35,22 @@ def test_normalize_term(raw: str, expected: str) -> None:
     assert normalize_term(raw) == expected
 
 
+# Terms whose lowercase form composes differently under NFC, so normalizing
+# NFC-then-lowercase isn't idempotent.
+NON_IDEMPOTENT_TERMS = [
+    "T\u0308",  # Uppercase base + combining mark: lowercase composes to \u1e97
+    "J\u030c",  # \u01f0 has no uppercase precomposed form
+    "\u0391\u0342",  # Greek polytonic: Alpha + perispomeni lowercases to \u1fb6
+]
+
+
+@pytest.mark.parametrize("raw", NON_IDEMPOTENT_TERMS)
+def test_normalize_term_is_idempotent(raw: str) -> None:
+    once = normalize_term(raw)
+    assert normalize_term(once) == once
+    assert normalize_term(raw.lower()) == once
+
+
 @pytest.fixture(params=["memory", "sqlite"])
 def term_index(
     request: pytest.FixtureRequest,
@@ -179,3 +195,26 @@ async def test_term_index_deserialize_dedupes_merged_ordinals(
     await term_index.deserialize(data)  # type: ignore[arg-type]
     found = await term_index.lookup_term("foo")
     assert found is not None and [r.semantic_ref_ordinal for r in found] == [1]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("term", NON_IDEMPOTENT_TERMS)
+async def test_term_index_normalization_round_trips(
+    term_index: ITermToSemanticRefIndex, term: str
+) -> None:
+    """A stored term must normalize to itself, or it can't be looked up."""
+    await term_index.add_term(term, 1)
+
+    [stored] = await term_index.get_terms()
+    found = await term_index.lookup_term(stored)
+    assert found is not None and [r.semantic_ref_ordinal for r in found] == [1]
+    found = await term_index.lookup_term(term.lower())
+    assert found is not None and [r.semantic_ref_ordinal for r in found] == [1]
+
+    # deserialize re-normalizes the already-normalized stored terms.
+    data = await term_index.serialize()
+    await term_index.clear()
+    await term_index.deserialize(data)
+    assert await term_index.get_terms() == [stored]
+    found = await term_index.lookup_term(term)
+    assert found is not None and [r.semantic_ref_ordinal for r in found] == [1]