diff --git a/changedetectionio/html_tools.py b/changedetectionio/html_tools.py
index 0a3cd108f..14414c19d 100644
--- a/changedetectionio/html_tools.py
+++ b/changedetectionio/html_tools.py
@@ -182,8 +182,10 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
     if is_rss:
         # So that we can keep CDATA for cdata_in_document_to_text() to process
         parser = etree.XMLParser(strip_cdata=False)
-
-    tree = html.fromstring(bytes(html_content, encoding='utf-8'), parser=parser)
+        # For XML/RSS content, use etree.fromstring to properly handle XML declarations
+        tree = etree.fromstring(html_content.encode('utf-8') if isinstance(html_content, str) else html_content, parser=parser)
+    else:
+        tree = html.fromstring(html_content, parser=parser)
     html_block = ""
 
     # Build namespace map for XPath queries
@@ -216,7 +218,10 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
         if type(element) == str:
             html_block += element
         elif issubclass(type(element), etree._Element) or issubclass(type(element), etree._ElementTree):
-            html_block += etree.tostring(element, pretty_print=True).decode('utf-8')
+            # Use 'xml' method for RSS/XML content, 'html' for HTML content
+            # parser will be XMLParser if we detected XML content
+            method = 'xml' if (is_rss or isinstance(parser, etree.XMLParser)) else 'html'
+            html_block += etree.tostring(element, pretty_print=True, method=method, encoding='unicode')
         else:
             html_block += elementpath_tostring(element)
 
@@ -226,13 +231,14 @@ def xpath_filter(xpath_filter, html_content, append_pretty_line_formatting=False
 # 'xpath1:'
 def xpath1_filter(xpath_filter, html_content, append_pretty_line_formatting=False, is_rss=False):
     from lxml import etree, html
-
     parser = None
     if is_rss:
         # So that we can keep CDATA for cdata_in_document_to_text() to process
         parser = etree.XMLParser(strip_cdata=False)
-
-    tree = html.fromstring(bytes(html_content, encoding='utf-8'), parser=parser)
+        # For XML/RSS content, use etree.fromstring to properly handle XML declarations
+        tree = etree.fromstring(html_content.encode('utf-8') if isinstance(html_content, str) else html_content, parser=parser)
+    else:
+        tree = html.fromstring(html_content, parser=parser)
     html_block = ""
 
     # Build namespace map for XPath queries
@@ -261,8 +267,11 @@ def xpath1_filter(xpath_filter, html_content, append_pretty_line_formatting=Fals
         if isinstance(element, (str, bytes)):
             html_block += element
         else:
-            # Return the HTML which will get parsed as text
-            html_block += etree.tostring(element, pretty_print=True).decode('utf-8')
+            # Return the HTML/XML which will get parsed as text
+            # Use 'xml' method for RSS/XML content, 'html' for HTML content
+            # parser will be XMLParser if we detected XML content
+            method = 'xml' if (is_rss or isinstance(parser, etree.XMLParser)) else 'html'
+            html_block += etree.tostring(element, pretty_print=True, method=method, encoding='unicode')
 
     return html_block
 
diff --git a/changedetectionio/processors/magic.py b/changedetectionio/processors/magic.py
index 2a0ef68f8..9d9018d74 100644
--- a/changedetectionio/processors/magic.py
+++ b/changedetectionio/processors/magic.py
@@ -103,15 +103,15 @@ class guess_stream_type():
             self.is_json = True
         elif 'pdf' in magic_content_header:
             self.is_pdf = True
-        elif has_html_patterns or http_content_header == 'text/html':
-            self.is_html = True
-        elif any(s in magic_content_header for s in JSON_CONTENT_TYPES):
-            self.is_json = True
         # magic will call a rss document 'xml'
         # Rarely do endpoints give the right header, usually just text/xml, so we check also for <rss
         # This also triggers the automatic CDATA text parser so the RSS goes back a nice content list
         elif '<rss' in test_content_normalized or '<feed' in test_content_normalized or any(s in magic_content_header for s in RSS_XML_CONTENT_TYPES) or '<rdf:' in test_content_normalized:
             self.is_rss = True
+        elif has_html_patterns or http_content_header == 'text/html':
+            self.is_html = True
+        elif any(s in magic_content_header for s in JSON_CONTENT_TYPES):
+            self.is_json = True
         elif any(s in http_content_header for s in XML_CONTENT_TYPES):
             # Only mark as generic XML if not already detected as RSS
             if not self.is_rss:
diff --git a/changedetectionio/tests/test_xpath_selector_unit.py b/changedetectionio/tests/test_xpath_selector_unit.py
index b4dda0806..a09a65393 100644
--- a/changedetectionio/tests/test_xpath_selector_unit.py
+++ b/changedetectionio/tests/test_xpath_selector_unit.py
@@ -201,3 +201,120 @@ def test_trips(html_content, xpath, answer):
     html_content = html_tools.xpath_filter(xpath, html_content, append_pretty_line_formatting=True)
     assert type(html_content) == str
     assert answer in html_content
+
+
+# Test for UTF-8 encoding bug fix (issue #3658)
+# Polish and other UTF-8 characters should be preserved correctly
+polish_html = """<!DOCTYPE html>
+<html>
+<head><meta charset="utf-8"></head>
+<body>
+<div class="index--s-headline-link">
+    <a class="index--s-headline-link" href="#">
+        Naukowcy potwierdzają: oglądanie krótkich filmików prowadzi do "zgnilizny mózgu"
+    </a>
+</div>
+<div>
+    <a class="other-class" href="#">
+        Test with Polish chars: żółć ąę śń
+    </a>
+</div>
+<div>
+    <p class="unicode-test">Cyrillic: Привет мир</p>
+    <p class="unicode-test">Greek: Γειά σου κόσμε</p>
+    <p class="unicode-test">Arabic: مرحبا بالعالم</p>
+    <p class="unicode-test">Chinese: 你好世界</p>
+    <p class="unicode-test">Japanese: こんにちは世界</p>
+    <p class="unicode-test">Emoji: 🌍🎉✨</p>
+</div>
+</body>
+</html>
+"""
+
+
+@pytest.mark.parametrize("html_content", [polish_html])
+@pytest.mark.parametrize("xpath, expected_text", [
+    # Test Polish characters in xpath_filter
+    ('//a[(contains(@class,"index--s-headline-link"))]', 'Naukowcy potwierdzają'),
+    ('//a[(contains(@class,"index--s-headline-link"))]', 'oglądanie krótkich filmików'),
+    ('//a[(contains(@class,"index--s-headline-link"))]', 'zgnilizny mózgu'),
+    ('//a[@class="other-class"]', 'żółć ąę śń'),
+
+    # Test various Unicode scripts
+    ('//p[@class="unicode-test"]', 'Привет мир'),
+    ('//p[@class="unicode-test"]', 'Γειά σου κόσμε'),
+    ('//p[@class="unicode-test"]', 'مرحبا بالعالم'),
+    ('//p[@class="unicode-test"]', '你好世界'),
+    ('//p[@class="unicode-test"]', 'こんにちは世界'),
+    ('//p[@class="unicode-test"]', '🌍🎉✨'),
+
+    # Test with text() extraction
+    ('//a[@class="other-class"]/text()', 'żółć'),
+])
+def test_xpath_utf8_encoding(html_content, xpath, expected_text):
+    """Test that XPath filters preserve UTF-8 characters correctly (issue #3658)"""
+    result = html_tools.xpath_filter(xpath, html_content, append_pretty_line_formatting=False)
+    assert type(result) == str
+    assert expected_text in result
+    # Ensure characters are NOT HTML-entity encoded
+    # For example, 'ą' should NOT become '&#261;'
+    assert '&#' not in result or expected_text in result
+
+
+@pytest.mark.parametrize("html_content", [polish_html])
+@pytest.mark.parametrize("xpath, expected_text", [
+    # Test Polish characters in xpath1_filter
+    ('//a[(contains(@class,"index--s-headline-link"))]', 'Naukowcy potwierdzają'),
+    ('//a[(contains(@class,"index--s-headline-link"))]', 'mózgu'),
+    ('//a[@class="other-class"]', 'żółć ąę śń'),
+
+    # Test various Unicode scripts with xpath1
+    ('//p[@class="unicode-test" and contains(text(), "Cyrillic")]', 'Привет мир'),
+    ('//p[@class="unicode-test" and contains(text(), "Greek")]', 'Γειά σου'),
+    ('//p[@class="unicode-test" and contains(text(), "Chinese")]', '你好世界'),
+])
+def test_xpath1_utf8_encoding(html_content, xpath, expected_text):
+    """Test that XPath1 filters preserve UTF-8 characters correctly"""
+    result = html_tools.xpath1_filter(xpath, html_content, append_pretty_line_formatting=False)
+    assert type(result) == str
+    assert expected_text in result
+    # Ensure characters are NOT HTML-entity encoded
+    assert '&#' not in result or expected_text in result
+
+
+# Test with real-world example from wyborcza.pl (issue #3658)
+wyborcza_style_html = """<!DOCTYPE html>
+<html lang="pl">
+<head><meta charset="utf-8"></head>
+<body>
+<div class="article-list">
+    <a class="index--s-headline-link" href="/article1">
+        Naukowcy potwierdzają: oglądanie krótkich filmików prowadzi do "zgnilizny mózgu"
+    </a>
+    <a class="index--s-headline-link" href="/article2">
+        Zmiany klimatyczne wpływają na życie w miastach
+    </a>
+    <a class="index--s-headline-link" href="/article3">
+        Łódź: Nowe inwestycje w infrastrukturę miejską
+    </a>
+</div>
+</body>
+</html>
+"""
+
+
+def test_wyborcza_real_world_example():
+    """Test real-world case from wyborcza.pl that was failing (issue #3658)"""
+    xpath = '//a[(contains(@class,"index--s-headline-link"))]'
+    result = html_tools.xpath_filter(xpath, wyborcza_style_html, append_pretty_line_formatting=False)
+
+    # These exact strings should appear in the result
+    assert 'Naukowcy potwierdzają' in result
+    assert 'oglądanie krótkich filmików' in result
+    assert 'zgnilizny mózgu' in result
+    assert 'Łódź' in result
+
+    # Make sure they're NOT corrupted to mojibake like "potwierdzajÄ"
+    assert 'potwierdzajÄ' not in result
+    assert 'ogl&#261;danie' not in result
+    assert 'm&#243;zgu' not in result