diff --git a/app/verify/cli.py b/app/verify/cli.py
index 45c3e88..310f92f 100644
--- a/app/verify/cli.py
+++ b/app/verify/cli.py
@@ -409,9 +409,18 @@ def cmd_check_urls(args: argparse.Namespace) -> int:
         max_workers=args.workers,
         min_interval=args.min_interval,
     )
+    # A rate-limited answer is not a verdict — leave it out so the next run asks
+    # again instead of parking the URL as dead for the whole TTL.
+    throttled = sum(1 for r in results if r.transient)
     for r in results:
-        cache[r.url] = http_check.result_to_entry(r, ts)
+        if not r.transient:
+            cache[r.url] = http_check.result_to_entry(r, ts)
     http_check.save_cache(cache)
+    if throttled:
+        print(
+            f"note: {throttled} URL(s) rate-limited after retries; "
+            "not cached, will retry next run"
+        )
     print(f"cache: wrote {len(cache)} URL result(s) to data/_verify/state/url_cache.jsonl")
     _summarize_cache(cache, targets)
     return 0
diff --git a/app/verify/http_check.py b/app/verify/http_check.py
index d56b754..eeb3e19 100644
--- a/app/verify/http_check.py
+++ b/app/verify/http_check.py
@@ -32,6 +32,17 @@
 )
 
 
+# "Come back later" is not "this link is dead". A host that rate-limits us says
+# nothing about whether the page exists, so these answers must never be cached as
+# a verdict — otherwise one impatient run marks a whole host dead for a TTL.
+TRANSIENT_STATUSES = frozenset({429, 503})
+RETRY_ATTEMPTS = 3
+RETRY_BACKOFF_S = (2.0, 6.0)
+MAX_RETRY_AFTER_S = 15.0
+# How much to slow a host down for the rest of the run once it has pushed back.
+RATE_LIMIT_PENALTY = 4.0
+
+
 class CheckResult(NamedTuple):
     url: str
     status: int | None
@@ -39,6 +50,10 @@ class CheckResult(NamedTuple):
     alive: bool
     reason: str
 
+    @property
+    def transient(self) -> bool:
+        return self.status in TRANSIENT_STATUSES
+
 
 # --- opener abstraction (injectable for tests) -----------------------------------
 
@@ -93,10 +108,21 @@ def classify(original_url: str, status: int | None, final_url: str | None) -> tu
     return True, f"http-{status}"
 
 
-def check_one(url: str, opener: Any) -> CheckResult:
-    """HEAD first; fall back to GET when HEAD is rejected (405/403) or errors."""
+def _retry_after_seconds(exc: Exception) -> float | None:
+    """Seconds requested by a ``Retry-After`` header, clamped to something sane."""
+    headers = getattr(exc, "headers", None)
+    raw = headers.get("Retry-After") if headers is not None else None
+    try:
+        return min(float(raw), MAX_RETRY_AFTER_S) if raw is not None else None
+    except (TypeError, ValueError):
+        return None  # HTTP-date form; fall back to our own backoff
+
+
+def _attempt(url: str, opener: Any) -> tuple[int | None, str | None, float | None]:
+    """One HEAD-then-GET pass. Returns (status, final_url, retry_after)."""
     status: int | None = None
     final: str | None = None
+    retry_after: float | None = None
     for method in ("HEAD", "GET"):
         try:
             status, final = opener.open(url, method)
@@ -107,10 +133,31 @@ def check_one(url: str, opener: Any) -> CheckResult:
             code = getattr(exc, "code", None)
             if isinstance(code, int):
                 status, final = code, getattr(exc, "url", None) or url
+                retry_after = _retry_after_seconds(exc)
                 if method == "HEAD" and code in (400, 403, 405, 501):
                     continue
                 break
             status, final = None, None
+    return status, final, retry_after
+
+
+def check_one(
+    url: str, opener: Any, *, on_rate_limit: Callable[[str], None] | None = None
+) -> CheckResult:
+    """HEAD first; fall back to GET when HEAD is rejected (405/403) or errors.
+
+    A rate-limit answer (429/503) is retried with backoff — the host is telling us
+    to wait, not that the page is gone.
+    """
+    status = final = retry_after = None
+    for attempt in range(RETRY_ATTEMPTS):
+        status, final, retry_after = _attempt(url, opener)
+        if status not in TRANSIENT_STATUSES:
+            break
+        if on_rate_limit is not None:
+            on_rate_limit(host_of(url))
+        if attempt < RETRY_ATTEMPTS - 1:
+            time.sleep(retry_after if retry_after is not None else RETRY_BACKOFF_S[attempt])
     alive, reason = classify(url, status, final)
     return CheckResult(url, status, final, alive, reason)
 
@@ -124,13 +171,26 @@ class HostRateLimiter:
     def __init__(self, min_interval: float = 1.0) -> None:
         self.min_interval = min_interval
         self._last: dict[str, float] = {}
+        self._interval: dict[str, float] = {}
         self._lock = threading.Lock()
 
+    def interval_for(self, host: str) -> float:
+        return self._interval.get(host, self.min_interval)
+
+    def back_off(self, host: str, factor: float = RATE_LIMIT_PENALTY) -> None:
+        """A host pushed back — slow it down for the rest of the run.
+
+        Without this, one rate-limited host keeps being hammered at the global
+        pace and every subsequent URL on it comes back 429.
+        """
+        with self._lock:
+            self._interval[host] = self.interval_for(host) * factor
+
     def wait(self, host: str) -> None:
         with self._lock:
             now = time.time()
             prev = self._last.get(host, 0.0)
-            sleep_for = max(0.0, self.min_interval - (now - prev))
+            sleep_for = max(0.0, self.interval_for(host) - (now - prev))
             self._last[host] = now + sleep_for
         if sleep_for > 0:
             time.sleep(sleep_for)
@@ -172,7 +232,7 @@ def _get_opener() -> Any:
 
     def _task(url: str) -> CheckResult:
         limiter.wait(host_of(url))
-        return check_one(url, _get_opener())
+        return check_one(url, _get_opener(), on_rate_limit=limiter.back_off)
 
     if not urls:
         return []
@@ -184,7 +244,18 @@ def _task(url: str) -> CheckResult:
 
 
 def load_cache(path: Path = URL_CACHE_PATH) -> dict[str, dict[str, Any]]:
-    return {e["url"]: e for e in ledger.iter_entries(path) if isinstance(e.get("url"), str)}
+    """Load the cache, dropping rate-limit answers written by older runs.
+
+    A 429/503 is not a verdict, so an entry holding one is not a cache hit —
+    it is a URL we still have to check. Filtering on load heals a cache that a
+    previous run poisoned (3,998 GSMArena pages were parked as dead this way,
+    all of which answer 200 when asked at a civil pace).
+    """
+    return {
+        e["url"]: e
+        for e in ledger.iter_entries(path)
+        if isinstance(e.get("url"), str) and e.get("status") not in TRANSIENT_STATUSES
+    }
 
 
 def _parse_ts(ts: str) -> datetime | None:
diff --git a/tests/verify/test_http_check.py b/tests/verify/test_http_check.py
index a2852b5..315ef47 100644
--- a/tests/verify/test_http_check.py
+++ b/tests/verify/test_http_check.py
@@ -103,3 +103,67 @@ def test_cache_roundtrip():
         assert loaded["https://x.com/y"]["alive"] is True
     finally:
         path.unlink(missing_ok=True)
+
+
+class _Http429(Exception):
+    """urllib-shaped 429, optionally carrying a Retry-After header."""
+
+    def __init__(self, url, retry_after=None):
+        super().__init__("Too Many Requests")
+        self.code = 429
+        self.url = url
+        self.headers = {"Retry-After": retry_after} if retry_after else {}
+
+
+class FlakyOpener(FakeOpener):
+    """Rate-limits the first ``fail_times`` calls, then answers normally."""
+
+    def __init__(self, table, fail_times):
+        super().__init__(table)
+        self.remaining = fail_times
+
+    def open(self, url, method):
+        self.calls.append((url, method))
+        if self.remaining > 0:
+            self.remaining -= 1
+            raise _Http429(url, retry_after="0")
+        return self.table[url]
+
+
+def test_rate_limit_is_retried_then_succeeds(monkeypatch):
+    monkeypatch.setattr(http_check.time, "sleep", lambda _s: None)
+    url = "https://www.gsmarena.com/x-1.php"
+    op = FlakyOpener({url: (200, url)}, fail_times=2)
+    [res] = http_check.check_urls([url], opener_factory=lambda: op, min_interval=0)
+    assert res.alive and res.status == 200 and not res.transient
+
+
+def test_persistent_rate_limit_is_transient_not_dead(monkeypatch):
+    monkeypatch.setattr(http_check.time, "sleep", lambda _s: None)
+    url = "https://www.gsmarena.com/y-2.php"
+    op = FlakyOpener({url: (200, url)}, fail_times=99)
+    [res] = http_check.check_urls([url], opener_factory=lambda: op, min_interval=0)
+    assert res.status == 429 and res.transient  # caller must not cache this as a verdict
+
+
+def test_rate_limit_slows_the_host_down(monkeypatch):
+    monkeypatch.setattr(http_check.time, "sleep", lambda _s: None)
+    url = "https://www.gsmarena.com/z-3.php"
+    limiter = http_check.HostRateLimiter(min_interval=1.0)
+    op = FlakyOpener({url: (200, url)}, fail_times=1)
+    http_check.check_urls([url], opener_factory=lambda: op, limiter=limiter)
+    assert limiter.interval_for("gsmarena.com") > 1.0
+
+
+def test_cached_rate_limit_entries_are_not_cache_hits(tmp_path):
+    path = tmp_path / "url_cache.jsonl"
+    path.write_text(
+        '{"url": "https://www.gsmarena.com/a-1.php", "status": 429, "alive": false,'
+        ' "reason": "http-429", "checked_at": "2026-08-03T00:00:00Z"}\n'
+        '{"url": "https://en.wikipedia.org/wiki/X", "status": 200, "alive": true,'
+        ' "reason": "http-200", "checked_at": "2026-08-03T00:00:00Z"}\n',
+        encoding="utf-8",
+    )
+    cache = http_check.load_cache(path)
+    assert "https://en.wikipedia.org/wiki/X" in cache
+    assert "https://www.gsmarena.com/a-1.php" not in cache  # a 429 is not an answer