diff --git a/lib/ads/video/encode.ts b/lib/ads/video/encode.ts
index 1a79b6f..d4c89cd 100644
--- a/lib/ads/video/encode.ts
+++ b/lib/ads/video/encode.ts
@@ -42,6 +42,22 @@ export const GOP_FRAMES = Math.round(
  */
 const MUSIC_BED_GAIN = "-16dB";
 
+/**
+ * Bring the finished mix to a normal listening level.
+ *
+ * A synthesised read comes back quiet and a five second spot has no room to
+ * ride the gain, so renders were landing around -30 dB mean: audible only if
+ * the viewer had already turned everything up for the programme, which is the
+ * one moment an advert must not ask them to. -16 LUFS is the level streaming
+ * platforms normalise to, so a spot arrives at the same loudness as whatever it
+ * interrupted rather than under it.
+ *
+ * The true-peak ceiling matters as much as the target. Without it, normalising
+ * a quiet source lifts its peaks into clipping, and a clipped voice is worse
+ * than a quiet one.
+ */
+const LOUDNESS = "loudnorm=I=-16:TP=-1.5:LRA=11";
+
 export type Mp4EncodeOptions = {
   /** printf-style pattern of the PNG frame sequence, e.g. `/tmp/x/f-%04d.png`. */
   framePattern: string;
@@ -148,7 +164,7 @@ export function mp4Args(o: Mp4EncodeOptions): string[] {
         `[1:a]apad,atrim=0:${seconds},asetpts=N/SR/TB[voice]`,
         `[2:a]atrim=0:${seconds},asetpts=N/SR/TB,volume=${MUSIC_BED_GAIN},` +
           `afade=t=in:st=0:d=${fade},afade=t=out:st=${(seconds - fade).toFixed(2)}:d=${fade}[bed]`,
-        `[voice][bed]amix=inputs=2:duration=first:normalize=0[a]`,
+        `[voice][bed]amix=inputs=2:duration=first:normalize=0,${LOUDNESS}[a]`,
       ].join(";"),
       "-map", "0:v",
       "-map", "[a]",
@@ -172,7 +188,7 @@ export function mp4Args(o: Mp4EncodeOptions): string[] {
     // not extend it and the encode either kept the original 2.3s of speech or,
     // once padded, dropped the stream entirely. An explicit duration is not a
     // workaround — it is the thing actually being asserted.
-    args.push("-af", "apad");
+    args.push("-af", `apad,${LOUDNESS}`);
     args.push("-c:a", "aac", "-profile:a", "aac_low", "-ar", String(AAC_SAMPLE_RATE), "-b:a", "128k", "-ac", "2");
   } else {
     args.push("-an");
diff --git a/tests/ads-video-compose-validate.test.ts b/tests/ads-video-compose-validate.test.ts
index 3949861..afd27bd 100644
--- a/tests/ads-video-compose-validate.test.ts
+++ b/tests/ads-video-compose-validate.test.ts
@@ -473,7 +473,7 @@ describe("a music bed under the narration", () => {
     expect(a).not.toContain("-filter_complex");
   });
 
-  it("no bed leaves the narrated path exactly as it was", () => {
+  it("no bed keeps the narrated path on a simple filter, not the mixer", () => {
     const a = mp4Args({
       framePattern: "/tmp/f-%04d.png",
       audioPath: "/tmp/narration.mp3",
@@ -482,6 +482,60 @@ describe("a music bed under the narration", () => {
       videoKbps: 4000,
     });
     expect(a).not.toContain("-filter_complex");
-    expect(a[a.indexOf("-af") + 1]).toBe("apad");
+    // The pad still comes first; loudness is applied after it, once the track
+    // is the length of the picture.
+    expect(a[a.indexOf("-af") + 1]).toMatch(/^apad(,|$)/);
+  });
+});
+
+describe("loudness", () => {
+  const narrated = () =>
+    mp4Args({
+      framePattern: "/tmp/f-%04d.png",
+      audioPath: "/tmp/narration.mp3",
+      outPath: "/tmp/out.mp4",
+      profile: "master_1080p",
+      videoKbps: 4000,
+    });
+  const withBed = () =>
+    mp4Args({
+      framePattern: "/tmp/f-%04d.png",
+      audioPath: "/tmp/narration.mp3",
+      musicPath: "/tmp/bed.mp3",
+      outPath: "/tmp/out.mp4",
+      profile: "master_1080p",
+      videoKbps: 4000,
+    });
+
+  it("brings a narrated spot up to streaming level", () => {
+    // Renders were landing at -30 dB mean: audible only to someone who had
+    // already turned everything up, which is the one moment an advert must not
+    // ask them to.
+    expect(narrated().join(" ")).toContain("loudnorm=I=-16");
+  });
+
+  it("normalises the mix, not the voice before the bed joins it", () => {
+    const f = withBed().join(" ");
+    // Normalising the voice alone would move it relative to the bed and undo
+    // the -16 dB the bed was placed at.
+    expect(f).toMatch(/amix=[^;]*normalize=0,loudnorm=/);
+  });
+
+  it("caps true peak, so lifting a quiet source cannot clip it", () => {
+    // A clipped voice is worse than a quiet one.
+    expect(narrated().join(" ")).toContain("TP=-1.5");
+    expect(withBed().join(" ")).toContain("TP=-1.5");
+  });
+
+  it("a silent spot gains no audio filter", () => {
+    const a = mp4Args({
+      framePattern: "/tmp/f-%04d.png",
+      audioPath: null,
+      outPath: "/tmp/out.mp4",
+      profile: "master_1080p",
+      videoKbps: 4000,
+    });
+    expect(a).toContain("-an");
+    expect(a.join(" ")).not.toContain("loudnorm");
   });
 });