diff --git a/pipeline/01_cc_index/schema.sql b/pipeline/01_cc_index/schema.sql
index 0387c05..34a0f1d 100644
--- a/pipeline/01_cc_index/schema.sql
+++ b/pipeline/01_cc_index/schema.sql
@@ -13,6 +13,7 @@ CREATE TABLE IF NOT EXISTS hosts (
     html_title TEXT,
     iframe_allowed BOOLEAN,
     best_icon_hash TEXT,
+    icon_downloaded_at TIMESTAMPTZ,
     parsed BOOLEAN DEFAULT FALSE,
     random_order DOUBLE PRECISION DEFAULT random()
 );
diff --git a/pipeline/04_best_icon/select.sql b/pipeline/04_best_icon/select.sql
index ac1dda6..fd0b514 100644
--- a/pipeline/04_best_icon/select.sql
+++ b/pipeline/04_best_icon/select.sql
@@ -21,7 +21,7 @@ SET work_mem = '2GB';
 \timing on
 
 CREATE TEMP TABLE best_icons AS
-SELECT DISTINCT ON (host_id) host_id, icon_hash
+SELECT DISTINCT ON (host_id) host_id, icon_hash, downloaded_at
 FROM icons
 WHERE scan_state = 'completed'
   AND icon_hash IS NOT NULL
@@ -50,7 +50,7 @@ ORDER BY host_id,
 
 \echo 'Step 2: Updating hosts...'
 
-UPDATE hosts h SET best_icon_hash = b.icon_hash
+UPDATE hosts h SET best_icon_hash = b.icon_hash, icon_downloaded_at = b.downloaded_at
 FROM best_icons b WHERE h.id = b.host_id;
 
 \timing off
diff --git a/pipeline/05_bundle_gen/db.go b/pipeline/05_bundle_gen/db.go
index 78aee54..ec07a85 100644
--- a/pipeline/05_bundle_gen/db.go
+++ b/pipeline/05_bundle_gen/db.go
@@ -2,30 +2,35 @@ package main
 
 import (
 	"context"
+	"time"
 
 	"github.com/jackc/pgx/v5/pgxpool"
 )
 
 type HostRow struct {
-	ID            int64
-	Hostname      string
-	Protocol      string
-	HtmlTitle     string
-	IframeAllowed bool
-	BestIconHash  string
-	RandomOrder   float64
+	ID               int64
+	Hostname         string
+	Protocol         string
+	HtmlTitle        string
+	IframeAllowed    bool
+	BestIconHash     string
+	RandomOrder      float64
+	IconDownloadedAt *time.Time
 }
 
-// fetchHostsPage gets a page of hosts with titles, ordered by id for disk locality.
+// fetchHostsPage gets a page of hosts with titles, ordered by icon_downloaded_at for disk locality.
+// Icons written to disk at similar times are physically adjacent — reading in write order
+// maximizes OS readahead cache hits. Hosts without icons come last (no disk reads needed).
 // random_order is included for bundle bucket assignment (randomized bundles).
-func fetchHostsPage(ctx context.Context, pool *pgxpool.Pool, lastID int64, limit int) ([]HostRow, error) {
+func fetchHostsPage(ctx context.Context, pool *pgxpool.Pool, lastDownloaded *time.Time, lastID int64, limit int) ([]HostRow, error) {
 	rows, err := pool.Query(ctx, `
-		SELECT id, hostname, protocol, html_title, COALESCE(iframe_allowed, true), COALESCE(best_icon_hash, ''), random_order
+		SELECT id, hostname, protocol, html_title, COALESCE(iframe_allowed, true), COALESCE(best_icon_hash, ''), random_order, icon_downloaded_at
 		FROM hosts
-		WHERE html_title IS NOT NULL AND id > $1
-		ORDER BY id
-		LIMIT $2
-	`, lastID, limit)
+		WHERE html_title IS NOT NULL
+		  AND (icon_downloaded_at, id) > ($1, $2)
+		ORDER BY icon_downloaded_at NULLS LAST, id
+		LIMIT $3
+	`, lastDownloaded, lastID, limit)
 	if err != nil {
 		return nil, err
 	}
@@ -34,7 +39,7 @@ func fetchHostsPage(ctx context.Context, pool *pgxpool.Pool, lastID int64, limit
 	var hosts []HostRow
 	for rows.Next() {
 		var h HostRow
-		if err := rows.Scan(&h.ID, &h.Hostname, &h.Protocol, &h.HtmlTitle, &h.IframeAllowed, &h.BestIconHash, &h.RandomOrder); err != nil {
+		if err := rows.Scan(&h.ID, &h.Hostname, &h.Protocol, &h.HtmlTitle, &h.IframeAllowed, &h.BestIconHash, &h.RandomOrder, &h.IconDownloadedAt); err != nil {
 			return nil, err
 		}
 		hosts = append(hosts, h)
diff --git a/pipeline/05_bundle_gen/main.go b/pipeline/05_bundle_gen/main.go
index 8f5ea71..b7cfffb 100644
--- a/pipeline/05_bundle_gen/main.go
+++ b/pipeline/05_bundle_gen/main.go
@@ -135,6 +135,7 @@ func main() {
 	// Stage 1: DB fetcher — continuously fetches pages into hostCh
 	go func() {
 		defer close(hostCh)
+		var lastDownloaded *time.Time
 		var lastID int64
 		pageSize := 50000
 		fetched := 0
@@ -150,14 +151,16 @@ func main() {
 				}
 			}
 			fetchStart := time.Now()
-			hosts, err := fetchHostsPage(ctx, pool, lastID, limit)
+			hosts, err := fetchHostsPage(ctx, pool, lastDownloaded, lastID, limit)
 			if err != nil {
 				log.Fatalf("Failed to fetch hosts: %v", err)
 			}
 			if len(hosts) == 0 {
 				break
 			}
-			lastID = hosts[len(hosts)-1].ID
+			last := hosts[len(hosts)-1]
+			lastDownloaded = last.IconDownloadedAt
+			lastID = last.ID
 			fmt.Printf("[fetcher] %d hosts in %dms (hostCh: %d/%d)\n",
 				len(hosts), time.Since(fetchStart).Milliseconds(), len(hostCh), cap(hostCh))
 			for _, h := range hosts {