diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
index 72c0249d1..3293c9bbb 100644
--- a/benchmarks/CMakeLists.txt
+++ b/benchmarks/CMakeLists.txt
@@ -68,6 +68,19 @@ if(MSVC AND BUILD_SHARED_LIBS)
         "$<TARGET_FILE_DIR:percent_encode>")                 # <--this is out-file path
 endif()
 
+# Percent Decode
+add_executable(percent_decode percent_decode.cpp)
+target_link_libraries(percent_decode PRIVATE ada counters::counters)
+target_include_directories(percent_decode PUBLIC "$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>")
+target_include_directories(percent_decode PUBLIC "$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/benchmarks>")
+if(MSVC AND BUILD_SHARED_LIBS)
+  # Copy the ada dll into the directory
+  add_custom_command(TARGET percent_decode POST_BUILD        # Adds a post-build event
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different  # which executes "cmake -E copy_if_different..."
+        "$<TARGET_FILE:ada>"      # <--this is in-file
+        "$<TARGET_FILE_DIR:percent_decode>")                 # <--this is out-file path
+endif()
+
 add_executable(model_bench model_bench.cpp)
 target_link_libraries(model_bench PRIVATE ada counters::counters)
 target_compile_definitions(model_bench PRIVATE ADA_URL_FILE="${url-dataset_SOURCE_DIR}/out.txt")
@@ -78,10 +91,11 @@ target_link_libraries(benchdata PRIVATE benchmark::benchmark)
 target_link_libraries(bbc_bench PRIVATE benchmark::benchmark)
 target_link_libraries(bench_ipv4 PRIVATE benchmark::benchmark)
 target_link_libraries(percent_encode PRIVATE benchmark::benchmark)
+target_link_libraries(percent_decode PRIVATE benchmark::benchmark)
 target_link_libraries(bench_search_params PRIVATE benchmark::benchmark)
 target_link_libraries(urlpattern PRIVATE benchmark::benchmark)
 
-set(BENCHMARKS wpt_bench bench benchdata bbc_bench bench_ipv4 percent_encode bench_search_params urlpattern)
+set(BENCHMARKS wpt_bench bench benchdata bbc_bench bench_ipv4 percent_encode percent_decode bench_search_params urlpattern)
 
 add_custom_target(run_all_benchmarks
         COMMAND ${CMAKE_COMMAND} -E echo "Running all benchmarks..."
diff --git a/benchmarks/percent_decode.cpp b/benchmarks/percent_decode.cpp
new file mode 100644
index 000000000..3fa873cf8
--- /dev/null
+++ b/benchmarks/percent_decode.cpp
@@ -0,0 +1,92 @@
+#include <memory>
+
+#include "ada.h"
+#include "ada/character_sets.h"
+#include "ada/unicode.h"
+#include "counters/event_counter.h"
+counters::event_collector collector;
+size_t N = 1000;
+
+#include <benchmark/benchmark.h>
+
+std::string examples[] = {
+    "utm_source=twc%20mobile&name=John%20Doe&ref=web-twc-ao-gbl",
+    "https%3A%2F%2Fexample.com%2Fpath%3Fa%3D1%26b%3D2%26c%3D3%26d%3D4",
+    "the/quick/brown/fox/jumps/over/the/lazy/dog%2Ehtml",
+    "caf%C3%A9%20%E4%BD%A0%E5%A5%BD%20%F0%9F%98%80%20done"};
+
+void init_data() {}
+
+double examples_bytes = []() -> double {
+  size_t bytes{0};
+  for (std::string& url_string : examples) {
+    bytes += url_string.size();
+  }
+  return double(bytes);
+}();
+
+static void Decode(benchmark::State& state) {
+  for (auto _ : state) {
+    for (std::string& url_string : examples) {
+      benchmark::DoNotOptimize(
+          ada::unicode::percent_decode(url_string, url_string.find('%')));
+    }
+  }
+  if (collector.has_events()) {
+    counters::event_aggregate aggregate{};
+    for (size_t i = 0; i < N; i++) {
+      std::atomic_thread_fence(std::memory_order_acquire);
+      collector.start();
+      for (std::string& url_string : examples) {
+        benchmark::DoNotOptimize(
+            ada::unicode::percent_decode(url_string, url_string.find('%')));
+      }
+      std::atomic_thread_fence(std::memory_order_release);
+      counters::event_count allocate_count = collector.end();
+      aggregate << allocate_count;
+    }
+    state.counters["instructions/url"] =
+        aggregate.best.instructions() / std::size(examples);
+    state.counters["instructions/cycle"] =
+        aggregate.total.instructions() / aggregate.total.cycles();
+    state.counters["instructions/byte"] =
+        aggregate.best.instructions() / examples_bytes;
+    state.counters["GHz"] =
+        aggregate.total.cycles() / aggregate.total.elapsed_ns();
+  }
+  state.counters["time/byte"] = benchmark::Counter(
+      examples_bytes, benchmark::Counter::kIsIterationInvariantRate |
+                          benchmark::Counter::kInvert);
+  state.counters["time/url"] =
+      benchmark::Counter(double(std::size(examples)),
+                         benchmark::Counter::kIsIterationInvariantRate |
+                             benchmark::Counter::kInvert);
+  state.counters["speed"] = benchmark::Counter(
+      examples_bytes, benchmark::Counter::kIsIterationInvariantRate);
+  state.counters["url/s"] =
+      benchmark::Counter(double(std::size(examples)),
+                         benchmark::Counter::kIsIterationInvariantRate);
+}
+BENCHMARK(Decode);
+
+int main(int argc, char** argv) {
+#if defined(ADA_RUST_VERSION)
+  benchmark::AddCustomContext("rust version ", ADA_RUST_VERSION);
+#endif
+#if (__APPLE__ && __aarch64__) || defined(__linux__)
+  if (!collector.has_events()) {
+    benchmark::AddCustomContext("performance counters",
+                                "No privileged access (sudo may help).");
+  }
+#else
+  if (!collector.has_events()) {
+    benchmark::AddCustomContext("performance counters", "Unsupported system.");
+  }
+#endif
+  if (collector.has_events()) {
+    benchmark::AddCustomContext("performance counters", "Enabled");
+  }
+  benchmark::Initialize(&argc, argv);
+  benchmark::RunSpecifiedBenchmarks();
+  benchmark::Shutdown();
+}
diff --git a/src/unicode.cpp b/src/unicode.cpp
index 5899a4640..7389fc983 100644
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@@ -461,31 +461,51 @@ std::string percent_decode(const std::string_view input, size_t first_percent) {
   if (first_percent == std::string_view::npos) {
     return std::string(input);
   }
-  std::string dest;
-  dest.reserve(input.length());
-  dest.append(input.substr(0, first_percent));
-  const char* pointer = input.data() + first_percent;
-  const char* end = input.data() + input.size();
-  // Optimization opportunity: if the following code gets
-  // called often, it can be optimized quite a bit.
-  while (pointer < end) {
-    const char ch = pointer[0];
-    size_t remaining = end - pointer - 1;
-    if (ch != '%' || remaining < 2 ||
-        (  // ch == '%' && // It is unnecessary to check that ch == '%'.
-            (!is_ascii_hex_digit(pointer[1]) ||
-             !is_ascii_hex_digit(pointer[2])))) {
-      dest += ch;
-      pointer++;
+
+  // NOLINTNEXTLINE(bugprone-suspicious-stringview-data-usage)
+  const char* const src = input.data();
+  const char* const end = src + input.size();
+
+  // Decoding never grows the string, so a single pre-sized buffer written via
+  // bulk memcpy of the plain runs (then shrunk to the final length) avoids the
+  // byte-at-a-time appends of the naive version.
+  std::string out(input.size(), '\0');
+  char* d = out.data();
+  char* const d0 = d;
+
+  std::memcpy(d, src, first_percent);
+  d += first_percent;
+
+  const char* p = src + first_percent;
+  while (p < end) {
+    if (*p == '%') {
+      // Decode runs of valid %XX tightly (common for nested/encoded URLs).
+      while (p + 2 < end && *p == '%') {
+        if (!is_ascii_hex_digit(p[1]) || !is_ascii_hex_digit(p[2])) {
+          break;
+        }
+        *d++ = static_cast<char>(convert_hex_to_binary(p[1]) * 16 +
+                                 convert_hex_to_binary(p[2]));
+        p += 3;
+      }
+      if (p < end && *p == '%') {
+        // Not a valid escape (too few chars left or bad hex): copy '%'
+        // literally and keep scanning after it.
+        *d++ = *p++;
+      }
     } else {
-      unsigned a = convert_hex_to_binary(pointer[1]);
-      unsigned b = convert_hex_to_binary(pointer[2]);
-      char c = static_cast<char>(a * 16 + b);
-      dest += c;
-      pointer += 3;
+      const char* q = static_cast<const char*>(
+          std::memchr(p, '%', static_cast<size_t>(end - p)));
+      const char* run_end = q ? q : end;
+      const size_t n = static_cast<size_t>(run_end - p);
+      std::memcpy(d, p, n);
+      d += n;
+      p = run_end;
     }
   }
-  return dest;
+
+  out.resize(static_cast<size_t>(d - d0));
+  return out;
 }
 
 // 0..15 for hex digits, 0xFF otherwise - validate and decode with two loads.
diff --git a/tests/basic_tests.cpp b/tests/basic_tests.cpp
index 0e2025fbb..e65a0f751 100644
--- a/tests/basic_tests.cpp
+++ b/tests/basic_tests.cpp
@@ -572,6 +572,26 @@ TEST(basic_tests, can_parse_consistency_percent_encoded_host) {
   }
 }
 
+// ada::unicode::percent_decode
+TEST(basic_tests, percent_decode_direct) {
+  using ada::unicode::percent_decode;
+  constexpr auto npos = std::string_view::npos;
+  // No percent sign: first_percent == npos returns the input unchanged.
+  ASSERT_EQ(percent_decode("no percent here", npos), "no percent here");
+  ASSERT_EQ(percent_decode("", npos), "");
+  // Valid escapes are decoded.
+  ASSERT_EQ(percent_decode("a%2Eb", 1), "a.b");
+  ASSERT_EQ(percent_decode("%41%42%43", 0), "ABC");
+  ASSERT_EQ(percent_decode("caf%C3%A9", 3), std::string("caf\xc3\xa9"));
+  // A plain run followed by an escape exercises the memchr run-copy.
+  ASSERT_EQ(percent_decode("hello%20world", 5), "hello world");
+  // Invalid escapes are copied literally.
+  ASSERT_EQ(percent_decode("%zz", 0), "%zz");    // non-hex digits
+  ASSERT_EQ(percent_decode("x%2", 1), "x%2");    // truncated escape at end
+  ASSERT_EQ(percent_decode("100%", 3), "100%");  // trailing '%'
+  ASSERT_EQ(percent_decode("%%41", 0), "%A");    // '%' then a valid escape
+}
+
 // Regression: try_can_parse_absolute_fast returned true for a valid IPv4 host
 // without validating the port. For "wS://1.3.3.51.:+" the host "1.3.3.51."
 // passes the IPv4 fast path, but the port "+" is not a valid digit, so the