diff --git a/CHANGELOG.md b/CHANGELOG.md
index cee3d83..9a638e3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -37,6 +37,14 @@ All notable changes to kage are recorded here. The format follows
   directly on its article and keeps that article's title metadata ([#62](https://github.com/tamnd/kage/issues/62)).
 - Non-UTF-8 `<meta charset>` and Content-Type charset declarations are
   rewritten to `utf-8`, matching the encoding kage writes to disk ([#16](https://github.com/tamnd/kage/issues/16)).
+- Relative links on redirected pages resolve against the browser's final URL
+  and the document's first `<base href>`, while the page remains saved under
+  the URL that was originally discovered. Consumed `href` attributes are
+  removed from every `<base>` so they cannot re-root rewritten links when the
+  saved page opens, while `target` behavior is preserved.
+  After a cross-host redirect,
+  relative references resolve to the other host and remain absolute when that
+  host is outside the crawl scope, so those resources are not localised.
 - `--resume` picks an interrupted crawl back up instead of doing nothing ([#36](https://github.com/tamnd/kage/issues/36)).
   `state.json` persisted only the visited set, and the frontier was rebuilt purely by re-rendering pages and following their links, which resume exists to avoid.
   So a resumed run found its seed already visited, `enqueuePage` turned it down, nothing was queued, and the run printed `pages 0` and exited successfully with most of the site still missing.
diff --git a/clone/cloner.go b/clone/cloner.go
index f597ec4..bc2ad7b 100644
--- a/clone/cloner.go
+++ b/clone/cloner.go
@@ -19,6 +19,7 @@ import (
 	"github.com/tamnd/kage/sanitize"
 	"github.com/tamnd/kage/urlx"
 	"golang.org/x/net/html"
+	"golang.org/x/net/html/atom"
 	"golang.org/x/time/rate"
 )
 
@@ -315,6 +316,13 @@ func (c *Cloner) processPage(ctx context.Context, j pageItem) {
 		return
 	}
 
+	// Resolve references against the post-redirect URL (and any <base href>),
+	// but keep writing the page under the discovered URL so existing offline
+	// links that pointed at /old still resolve. Cross-host redirects leave the
+	// resolve base as the final location for relative refs; scope checks still
+	// use that absolute URL.
+	resolveBase := pageResolveBase(j.u, res.FinalURL, root)
+
 	localFile := urlx.LocalPath(c.seedHost, j.u, urlx.Page, c.cfg.Reserved)
 	fileDir := urlx.Dir(localFile)
 
@@ -337,7 +345,7 @@ func (c *Cloner) processPage(ctx context.Context, j pageItem) {
 		}
 	}
 
-	asset.RewriteHTML(root, j.u, sink)
+	asset.RewriteHTML(root, resolveBase, sink)
 	sanitize.CleanTree(root, sanitize.Options{
 		KeepNoscript:   c.cfg.KeepNoscript,
 		MobileReadable: c.cfg.MobileReadable,
@@ -367,6 +375,55 @@ func (c *Cloner) waitForCrawlDelay(ctx context.Context) bool {
 	return c.crawlLimiter.Wait(ctx) == nil
 }
 
+// pageResolveBase picks the URL against which relative references on a rendered
+// page should resolve. Preference order:
+//  1. A document <base href> (the live page's own base);
+//  2. The browser's final URL after redirects;
+//  3. The URL that was enqueued.
+//
+// The page is still written under the enqueued URL so offline links discovered
+// as /old keep working when the server redirected /old → /new.
+func pageResolveBase(enqueued *url.URL, finalURL string, root *html.Node) *url.URL {
+	base := enqueued
+	if finalURL != "" {
+		if u, err := url.Parse(finalURL); err == nil && u.Scheme != "" && u.Host != "" {
+			// Drop fragment; keep query/path as the browser shows them.
+			u.Fragment = ""
+			base = u
+		}
+	}
+	if href := documentBaseHref(root); href != "" {
+		if u, err := urlx.Normalize(base, href); err == nil {
+			return u
+		}
+	}
+	return base
+}
+
+// documentBaseHref returns the first <base href> in document order, or "".
+func documentBaseHref(root *html.Node) string {
+	var found string
+	var walk func(*html.Node)
+	walk = func(n *html.Node) {
+		if found != "" || n == nil {
+			return
+		}
+		if n.Type == html.ElementNode && n.DataAtom == atom.Base {
+			for _, a := range n.Attr {
+				if strings.EqualFold(a.Key, "href") && strings.TrimSpace(a.Val) != "" {
+					found = strings.TrimSpace(a.Val)
+					return
+				}
+			}
+		}
+		for c := n.FirstChild; c != nil && found == ""; c = c.NextSibling {
+			walk(c)
+		}
+	}
+	walk(root)
+	return found
+}
+
 // processAsset downloads one asset, rewriting CSS references on the way, and
 // writes it to its deterministic local path.
 func (c *Cloner) processAsset(ctx context.Context, j assetItem) {
diff --git a/clone/resolve_test.go b/clone/resolve_test.go
new file mode 100644
index 0000000..42d59dc
--- /dev/null
+++ b/clone/resolve_test.go
@@ -0,0 +1,45 @@
+package clone
+
+import (
+	"net/url"
+	"strings"
+	"testing"
+
+	"golang.org/x/net/html"
+)
+
+func TestPageResolveBaseUsesFinalURL(t *testing.T) {
+	enqueued, _ := url.Parse("https://ex.com/old")
+	root, err := html.Parse(strings.NewReader(`<html><head></head><body><a href="next">n</a></body></html>`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	base := pageResolveBase(enqueued, "https://ex.com/new/", root)
+	if base.String() != "https://ex.com/new/" {
+		t.Fatalf("resolve base = %q, want final URL", base)
+	}
+}
+
+func TestPageResolveBasePrefersDocumentBase(t *testing.T) {
+	enqueued, _ := url.Parse("https://ex.com/page")
+	root, err := html.Parse(strings.NewReader(
+		`<html><head><base href="https://ex.com/dir/"></head><body></body></html>`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	base := pageResolveBase(enqueued, "https://ex.com/page", root)
+	if base.String() != "https://ex.com/dir/" {
+		t.Fatalf("resolve base = %q, want document <base>", base)
+	}
+}
+
+func TestDocumentBaseHref(t *testing.T) {
+	root, err := html.Parse(strings.NewReader(
+		`<html><head><base href="/subdir/"><base href="https://ignored.example/"></head></html>`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := documentBaseHref(root); got != "/subdir/" {
+		t.Fatalf("documentBaseHref = %q, want first base", got)
+	}
+}
diff --git a/docs/content/reference/release-notes.md b/docs/content/reference/release-notes.md
index f408934..a2976e3 100644
--- a/docs/content/reference/release-notes.md
+++ b/docs/content/reference/release-notes.md
@@ -18,6 +18,11 @@ The authoritative, commit-level history lives in [`CHANGELOG.md`](https://github
   single-page archives still open directly on their article ([#62](https://github.com/tamnd/kage/issues/62)).
 - **Saved pages declare their real encoding.** Non-UTF-8 charset metadata is
   rewritten to UTF-8, matching the bytes kage writes to disk ([#16](https://github.com/tamnd/kage/issues/16)).
+- **Redirects resolve correctly.** Relative links use the post-redirect URL and
+  the document's first `<base href>`. After rewriting, every base `href` is
+  removed so it cannot affect the saved page, while a base `target` is
+  preserved. On a cross-host redirect, references to an out-of-scope
+  destination remain absolute rather than being localised.
 
 ## v0.3.11
 
diff --git a/sanitize/sanitize.go b/sanitize/sanitize.go
index de6adb7..3cf81d1 100644
--- a/sanitize/sanitize.go
+++ b/sanitize/sanitize.go
@@ -49,6 +49,7 @@ type Report struct {
 	MetaRefreshRemoved  int
 	DeadLinksRemoved    int
 	CondCommentsRemoved int
+	BaseHrefsRemoved    int
 	CharsetAdded        bool
 	CharsetRewritten    bool
 }
@@ -110,6 +111,15 @@ func clean(n *html.Node, opts Options, rep *Report) {
 		}
 		if c.Type == html.ElementNode {
 			switch c.DataAtom {
+			case atom.Base:
+				// Link rewriting has already consumed the document base. Remove every
+				// href so a later base cannot become active when an earlier one is
+				// removed, but preserve target because it controls browsing contexts.
+				rep.BaseHrefsRemoved += stripBaseHrefs(c)
+				if len(c.Attr) == 0 {
+					n.RemoveChild(c)
+					continue
+				}
 			case atom.Script:
 				n.RemoveChild(c)
 				rep.ScriptsRemoved++
@@ -143,6 +153,20 @@ func clean(n *html.Node, opts Options, rep *Report) {
 	}
 }
 
+func stripBaseHrefs(n *html.Node) int {
+	kept := n.Attr[:0]
+	removed := 0
+	for _, a := range n.Attr {
+		if strings.EqualFold(a.Key, "href") {
+			removed++
+			continue
+		}
+		kept = append(kept, a)
+	}
+	n.Attr = kept
+	return removed
+}
+
 // stripHandlers removes every on* event-handler attribute from n.
 func stripHandlers(n *html.Node, rep *Report) {
 	kept := n.Attr[:0]
diff --git a/sanitize/sanitize_test.go b/sanitize/sanitize_test.go
index d618eaf..2e41479 100644
--- a/sanitize/sanitize_test.go
+++ b/sanitize/sanitize_test.go
@@ -167,6 +167,43 @@ func TestKeepMetaRefreshPlain(t *testing.T) {
 	}
 }
 
+func TestBaseHrefRemovedAfterRewrite(t *testing.T) {
+	cases := []struct {
+		name        string
+		bases       string
+		wantBase    bool
+		wantTarget  bool
+		wantRemoved int
+	}{
+		{name: "href only", bases: `<base href="https://example.com/live/">`, wantRemoved: 1},
+		{name: "target only", bases: `<base target="_blank">`, wantBase: true, wantTarget: true},
+		{name: "href and target", bases: `<base href="https://example.com/live/" target="_blank">`, wantBase: true, wantTarget: true, wantRemoved: 1},
+		{name: "all hrefs", bases: `<base href="https://example.com/one/"><base href="https://example.com/two/" target="_blank">`, wantBase: true, wantTarget: true, wantRemoved: 2},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			in := `<html><head>` + tc.bases + `</head><body><a href="saved.html">saved</a></body></html>`
+			out, rep, err := Strip([]byte(in), Options{})
+			if err != nil {
+				t.Fatal(err)
+			}
+			s := strings.ToLower(string(out))
+			if got := strings.Contains(s, "<base"); got != tc.wantBase {
+				t.Errorf("base presence = %v, want %v:\n%s", got, tc.wantBase, out)
+			}
+			if got := strings.Contains(s, `target="_blank"`); got != tc.wantTarget {
+				t.Errorf("base target presence = %v, want %v:\n%s", got, tc.wantTarget, out)
+			}
+			if strings.Contains(s, "example.com/") {
+				t.Errorf("base href survived:\n%s", out)
+			}
+			if rep.BaseHrefsRemoved != tc.wantRemoved {
+				t.Errorf("BaseHrefsRemoved = %d, want %d", rep.BaseHrefsRemoved, tc.wantRemoved)
+			}
+		})
+	}
+}
+
 func TestCharsetAddedWhenMissing(t *testing.T) {
 	// A page whose source declared its charset only in the HTTP header has no
 	// <meta charset>. The saved file must gain one so a reader does not fall back