diff --git a/misc/management/commands/retranslate_html_links.py b/misc/management/commands/retranslate_html_links.py
new file mode 100644
index 0000000000..3674f77041
--- /dev/null
+++ b/misc/management/commands/retranslate_html_links.py
@@ -0,0 +1,103 @@
+import logging
+from typing import Any
+
+from django.core.management.base import BaseCommand, CommandParser
+from django.db.models import Q
+from modeltranslation.translator import translator
+from modeltranslation.utils import build_localized_fieldname
+from django.conf import settings
+
+from utils.translation import (
+    detect_and_update_content_language,
+    get_translation_fields_for_model,
+    update_translations_for_model,
+)
+
+logging.basicConfig(level=logging.DEBUG)
+
+# Markdown links whose label contains an inline HTML tag - e.g. [<u>text</u>](url).
+# Those were destroyed by translating markdown with the (previously defaulted)
+# "text/html" mimeType, which dropped the "[" and "](url)" around the tag.
+BROKEN_LINK_REGEX = r"\[[^\]]*<[a-zA-Z]"
+
+
+def get_models(app_label, model_name):
+    models = translator.get_registered_models(abstract=False)
+    models = [m for m in models if not m._meta.proxy and m._meta.managed]
+
+    if app_label:
+        models = [m for m in models if m._meta.app_label == app_label]
+
+    if model_name:
+        model_name = model_name.lower()
+        models = [m for m in models if m._meta.model_name == model_name]
+
+    return models
+
+
+def filter_affected(queryset):
+    model = queryset.model
+    original_lang = settings.ORIGINAL_LANGUAGE_CODE
+
+    match_any_field = Q()
+    for field_name in get_translation_fields_for_model(model):
+        original_field = build_localized_fieldname(field_name, original_lang)
+        match_any_field |= Q(**{f"{original_field}__regex": BROKEN_LINK_REGEX})
+
+    return queryset.filter(
+        match_any_field,
+        is_automatically_translated=True,
+        # Objects without a detected language were never translated, so they have
+        # nothing broken to repair. Leave them to the regular update_translations run
+        # rather than translating them for the first time as a side effect of this.
+        content_original_lang__isnull=False,
+    )
+
+
+class Command(BaseCommand):
+    help = (
+        "Re-translate content whose markdown links were destroyed by translating"
+        " with the html mimeType. Dry run by default."
+    )
+
+    def add_arguments(self, parser: CommandParser) -> None:
+        parser.add_argument("app_label", nargs="?")
+        parser.add_argument("model_name", nargs="?")
+        parser.add_argument("--batch_size", default=5, type=int)
+        parser.add_argument(
+            "--apply",
+            action="store_true",
+            help="Actually re-translate. Without it the command only reports counts.",
+        )
+
+    def handle(self, *args: Any, **options: Any) -> None:
+        batch_size = options["batch_size"]
+        apply = options["apply"]
+
+        for model in get_models(options["app_label"], options["model_name"]):
+            pks = list(
+                filter_affected(model._default_manager.all()).values_list(
+                    "pk", flat=True
+                )
+            )
+
+            if not pks:
+                continue
+
+            logging.info(
+                f"{model.__name__}: {len(pks)} affected objects, e.g. {pks[:10]}"
+            )
+
+            if not apply:
+                continue
+
+            queryset = model._default_manager.filter(pk__in=pks)
+            # Nulling the hash is what makes queryset_filter_outdated_translations
+            # pick these up again - the original content itself has not changed.
+            queryset.update(content_last_md5=None)
+
+            detect_and_update_content_language(queryset, batch_size)
+            update_translations_for_model(queryset, batch_size)
+
+        if not apply:
+            logging.info("Dry run - pass --apply to re-translate.")
diff --git a/tests/unit/test_utils/test_translation.py b/tests/unit/test_utils/test_translation.py
new file mode 100644
index 0000000000..91ceedc919
--- /dev/null
+++ b/tests/unit/test_utils/test_translation.py
@@ -0,0 +1,164 @@
+import asyncio
+from contextlib import asynccontextmanager
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from utils.translation import (
+    agoogle_translate_text,
+    rejoin_translated,
+    split_for_translation,
+)
+
+# The Spanish original of question 35557, which regressed: translating it as HTML
+# dropped the "[" and "](url)" around the <u> tag, leaving unclickable text.
+GEORGESCU_ES = (
+    "En noviembre de 2024, Călin Georgescu, un [<u>candidato ultranacionalista con"
+    " sentimientos pro-rusos</u>](https://ro.wikipedia.org/wiki/C%C4%83lin_Georgescu),"
+    " ganó la primera vuelta."
+)
+
+
+def roundtrip(text, translate=lambda line: line):
+    translatable, layout = split_for_translation(text)
+    return rejoin_translated(layout, [translate(line) for line in translatable])
+
+
+class TestSplitForTranslation:
+    def test_only_non_blank_lines_are_sent(self):
+        translatable, _ = split_for_translation("uno\n\n\ndos")
+        assert translatable == ["uno", "dos"]
+
+    def test_blank_lines_are_preserved(self):
+        assert roundtrip("uno\n\n\ndos") == "uno\n\n\ndos"
+
+    def test_leading_and_trailing_newlines_are_preserved(self):
+        assert roundtrip("\n\nuno\n\n") == "\n\nuno\n\n"
+
+    def test_carriage_returns_are_preserved(self):
+        assert roundtrip("uno\r\n\r\ndos") == "uno\r\n\r\ndos"
+
+    def test_single_line(self):
+        assert roundtrip("uno") == "uno"
+
+    def test_empty_text_has_nothing_to_translate(self):
+        translatable, layout = split_for_translation("")
+        assert translatable == []
+        assert rejoin_translated(layout, []) == ""
+
+    def test_whitespace_only_text_is_untouched(self):
+        translatable, _ = split_for_translation("\n \n")
+        assert translatable == []
+        assert roundtrip("\n \n") == "\n \n"
+
+    def test_markdown_structure_survives_a_translation(self):
+        text = "# Título\n\nUn [enlace](https://example.com) aquí.\n\n* uno\n* dos"
+        assert roundtrip(text, translate=str.upper) == (
+            "# TÍTULO\n\nUN [ENLACE](HTTPS://EXAMPLE.COM) AQUÍ.\n\n* UNO\n* DOS"
+        )
+
+    def test_link_with_inline_html_label_is_left_intact(self):
+        # The regression from #2299: the link must reach the API in one piece.
+        translatable, _ = split_for_translation(GEORGESCU_ES)
+        assert translatable == [GEORGESCU_ES]
+        assert roundtrip(GEORGESCU_ES) == GEORGESCU_ES
+
+
+class TestRejoinTranslated:
+    def test_rejects_a_short_response(self):
+        _, layout = split_for_translation("uno\ndos")
+        with pytest.raises(ValueError):
+            rejoin_translated(layout, ["one"])
+
+
+def mock_translate_session(translations):
+    response = MagicMock()
+    response.status = 200
+    response.json = AsyncMock(
+        return_value={"translations": [{"translatedText": t} for t in translations]}
+    )
+
+    post_calls = []
+
+    @asynccontextmanager
+    async def post(url, headers, json):
+        post_calls.append({"url": url, "headers": headers, "json": json})
+        yield response
+
+    session = MagicMock()
+    session.post = post
+
+    @asynccontextmanager
+    async def client_session():
+        yield session
+
+    return client_session, post_calls
+
+
+class TestAgoogleTranslateText:
+    def _translate(self, text, translations):
+        client_session, post_calls = mock_translate_session(translations)
+
+        with (
+            patch(
+                "utils.translation.get_and_cache_sa_info",
+                return_value=("token", "project"),
+            ),
+            patch("aiohttp.ClientSession", client_session),
+        ):
+            result = asyncio.run(agoogle_translate_text("es", "en", text))
+
+        return result, post_calls
+
+    def test_sends_plain_text_mimetype_in_the_payload(self):
+        _, post_calls = self._translate("hola", ["hello"])
+
+        payload = post_calls[0]["json"]
+        assert payload["mimeType"] == "text/plain"
+        # mimeType in the headers is what caused #2299 - it was silently ignored.
+        assert "mimeType" not in post_calls[0]["headers"]
+
+    def test_sends_contents_as_a_list_of_lines(self):
+        _, post_calls = self._translate("hola\n\nadiós", ["hello", "goodbye"])
+
+        assert post_calls[0]["json"]["contents"] == ["hola", "adiós"]
+
+    def test_reassembles_the_translated_lines(self):
+        result, _ = self._translate("hola\n\nadiós", ["hello", "goodbye"])
+
+        assert result == "hello\n\ngoodbye"
+
+    def test_unescapes_html_entities_in_the_response(self):
+        result, _ = self._translate("Tom y Jerry", ["Tom &amp; Jerry"])
+
+        assert result == "Tom & Jerry"
+
+    def test_same_language_skips_the_api(self):
+        client_session, post_calls = mock_translate_session([])
+
+        with (
+            patch(
+                "utils.translation.get_and_cache_sa_info",
+                return_value=("token", "project"),
+            ),
+            patch("aiohttp.ClientSession", client_session),
+        ):
+            result = asyncio.run(agoogle_translate_text("en", "en", "unchanged"))
+
+        assert result == "unchanged"
+        assert post_calls == []
+
+    def test_text_without_anything_translatable_skips_the_api(self):
+        client_session, post_calls = mock_translate_session([])
+
+        with (
+            patch(
+                "utils.translation.get_and_cache_sa_info",
+                return_value=("token", "project"),
+            ),
+            patch("aiohttp.ClientSession", client_session),
+        ):
+            result = asyncio.run(agoogle_translate_text("es", "en", "\n\n"))
+
+        assert result == "\n\n"
+        assert post_calls == []
diff --git a/utils/translation.py b/utils/translation.py
index 354527ebae..0895e47d63 100644
--- a/utils/translation.py
+++ b/utils/translation.py
@@ -76,41 +76,77 @@ async def agoogle_translate_detect_language(text):
                 raise Exception(f"Error detecting language: {error}")
 
 
+def split_for_translation(text):
+    """
+    Splits text into the lines that need translating, plus a layout describing how
+    to put it back together. Each layout entry is either None (a placeholder for the
+    next translated line) or a literal string to emit as-is (blank lines).
+
+    Translating line by line keeps the line structure intact no matter how the
+    translation service handles whitespace, and lines map onto markdown paragraphs.
+    """
+    layout = []
+    translatable = []
+
+    for line in text.split("\n"):
+        if line.strip():
+            layout.append(None)
+            translatable.append(line)
+        else:
+            layout.append(line)
+
+    return translatable, layout
+
+
+def rejoin_translated(layout, translated):
+    expected = sum(1 for entry in layout if entry is None)
+    if len(translated) != expected:
+        raise ValueError(f"Expected {expected} translated lines, got {len(translated)}")
+
+    translated_iter = iter(translated)
+    return "\n".join(
+        next(translated_iter) if entry is None else entry for entry in layout
+    )
+
+
 async def agoogle_translate_text(source_language, target_language, text):
     token, project_id = get_and_cache_sa_info()
 
     if source_language == target_language:
         return text
 
-    # Google Translates doesn't preserve new lines, as it translates text as if it was HTML.
-    # so we use a hack by inserting a <br> element for each \n before translating,
-    # and then we replace it back with the new line after translation.
-    # In addition, it also html-escapes the input so we need to unescape it after the translation
-    text = text.replace("\n", "<br>")
+    translatable, layout = split_for_translation(text)
+    if not translatable:
+        return text
+
     url = f"https://translation.googleapis.com/v3/projects/{project_id}:translateText"
 
     headers = {
         "Authorization": f"Bearer {token}",
         "x-goog-user-project": project_id,
         "Content-Type": "application/json; charset=utf-8",
-        "mimeType": "text/plain",
     }
 
+    # mimeType must be part of the payload, not of the headers. It defaults to
+    # "text/html", which makes the service parse our markdown as HTML: inline tags
+    # split it into separate text nodes and the non-prose ones (like the "](url)"
+    # of a markdown link) get dropped, silently destroying links.
     data = {
         "sourceLanguageCode": source_language,
         "targetLanguageCode": target_language,
-        "contents": text,
+        "mimeType": "text/plain",
+        "contents": translatable,
     }
 
     async with aiohttp.ClientSession() as session:
         async with session.post(url, headers=headers, json=data) as response:
             if response.status == 200:
                 result = await response.json()
-                output = result["translations"][0]["translatedText"]
-                # See the comment above with the new lines
-                output = output.replace("<br>", "\n")
-                output = html.unescape(output)
-                return output
+                translated = [t["translatedText"] for t in result["translations"]]
+                # Kept from the text/html days: harmless if the response is already
+                # unescaped, and it still corrects the output if it is not.
+                translated = [html.unescape(t) for t in translated]
+                return rejoin_translated(layout, translated)
             else:
                 error = await response.text()
                 raise Exception(f"Error translating text: {error}")