diff --git a/readers/multishard.cc b/readers/multishard.cc
index 4c65b71e014a..3abfecc6318b 100644
--- a/readers/multishard.cc
+++ b/readers/multishard.cc
@@ -607,7 +607,7 @@ future<> evictable_reader_v2::fill_buffer() {
         // First make sure we've made progress w.r.t. _next_position_in_partition.
         // This loop becomes inifinite when next pos is a partition start.
         // In that case progress is guranteed anyway, so skip this loop entirely.
-        while (!_next_position_in_partition.is_partition_start() && next_mf && _tri_cmp(_next_position_in_partition, buffer().back().position()) <= 0) {
+        while (!_next_position_in_partition.is_partition_start() && next_mf && _tri_cmp(buffer().back().position(), _next_position_in_partition) <= 0) {
             push_mutation_fragment(_reader->pop_mutation_fragment());
             next_mf = co_await _reader->peek();
         }
diff --git a/test/boost/mutation_reader_test.cc b/test/boost/mutation_reader_test.cc
index 1773392d814d..be9b7cb736b8 100644
--- a/test/boost/mutation_reader_test.cc
+++ b/test/boost/mutation_reader_test.cc
@@ -3648,9 +3648,13 @@ SEASTAR_THREAD_TEST_CASE(test_evictable_reader_next_pos_is_partition_start) {
     auto stop_rd = deferred_close(rd);
     rd.set_max_buffer_size(max_buf_size);
 
+    // #13491 - the reader must not consume the entire partition but a small batch of fragments based on the buffer size.
+    rd.fill_buffer().get();
     rd.fill_buffer().get();
     auto buf1 = rd.detach_buffer();
-    BOOST_REQUIRE_EQUAL(buf1.size(), 3);
+    // There should be 6-7 fragments, but to avoid computing the exact number of fragments that should fit in `max_buf_size`,
+    // just ensure that there are <= 10 (consuming the whole partition would give ~1000 fragments).
+    BOOST_REQUIRE_LE(buf1.size(), 10);
 }
 
 struct mutation_bounds {