diff --git a/src/seo/autofill.py b/src/seo/autofill.py
index 717599c..ccd397d 100644
--- a/src/seo/autofill.py
+++ b/src/seo/autofill.py
@@ -27,12 +27,24 @@ from src.seo import rules as R
 
 logger = structlog.get_logger(__name__)
 
-# Canonical site host per language. The wrapped <a> hrefs use the www host (the
-# site's canonical form); rules.internal_links still counts them because its
-# SITE_HOST ("theexclusionzone.com") is a substring of "www.theexclusionzone.com".
+# Canonical site host per language — and the two are INVERTED, which is exactly
+# the trap: EN canonicalizes on www, ES canonicalizes on the APEX. Same
+# inversion that caused the 522 outage, and the same one ghst-es warns about.
+#
+# ES said "www." until 2026-07-21. Every internal link the generator wrote into
+# a Spanish draft therefore pointed at a non-canonical host and ate a 301 —
+# that is where the stray www link found in los-villares came from. Nothing
+# downstream would have stopped it either: seo_watch only sees it after the
+# post is published.
+#
+# Caveat, unchanged by this fix: rules.internal_links counts EN links because
+# its SITE_HOST ("theexclusionzone.com") is a substring of the www form, but it
+# counts ZERO for ES under either host, since that module is hardcoded to the
+# EN host (it is vendored byte-for-byte and CI enforces the copy). Harmless
+# today because internal_links.too_few is not a blocking rule at draft time.
 SITE_BY_LANG = {
     "en": "www.theexclusionzone.com",
-    "es": "www.zonadeexclusion.com",
+    "es": "zonadeexclusion.com",
 }
 
 # Tag allow-list — the model may ONLY pick from these; invented tags are dropped
diff --git a/tests/test_seo_autofill.py b/tests/test_seo_autofill.py
index da938f3..8ed2a4f 100644
--- a/tests/test_seo_autofill.py
+++ b/tests/test_seo_autofill.py
@@ -1,4 +1,5 @@
-from src.seo.autofill import ALLOWED_TAGS, DEFAULT_TAG, _coerce, _system_prompt
+from src.seo.autofill import (ALLOWED_TAGS, DEFAULT_TAG, _coerce, _system_prompt,
+                              insert_internal_links)
 
 BASE = {
     "meta_title": "t",
@@ -81,3 +82,28 @@ def test_collision_note_is_markdown_safe():
 
 def test_slugify_title():
     assert _slugify_title("USS Russell 2019: The Pyramid UAP!") == "uss-russell-2019-the-pyramid-uap"
+
+
+# --- canonical host per language -------------------------------------------
+# EN canonicalizes on www, ES on the APEX. They are INVERTED, and the ES entry
+# said "www." until 2026-07-21, so every internal link written into a Spanish
+# draft ate a 301. Nothing caught it: no test covered the host, and seo_watch
+# only sees a link once the post is published. These two pin it.
+
+def test_internal_link_uses_es_apex_canonical():
+    html = "<p>El caso de Manises sigue abierto.</p>"
+    out, pairs = insert_internal_links(
+        html, [{"phrase": "Manises", "slug": "manises-1979"}],
+        [{"slug": "manises-1979", "title": "Manises"}], "es")
+    assert 'href="https://zonadeexclusion.com/manises-1979/"' in out
+    assert "www.zonadeexclusion.com" not in out
+    assert len(pairs) == 1
+
+
+def test_internal_link_uses_en_www_canonical():
+    html = "<p>The Roswell debris was recovered.</p>"
+    out, pairs = insert_internal_links(
+        html, [{"phrase": "Roswell", "slug": "roswell-1947"}],
+        [{"slug": "roswell-1947", "title": "Roswell"}], "en")
+    assert 'href="https://www.theexclusionzone.com/roswell-1947/"' in out
+    assert len(pairs) == 1