diff --git a/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.cpp b/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.cpp
index 634285d530..eba440d84d 100644
--- a/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.cpp
+++ b/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.cpp
@@ -48,12 +48,21 @@ void BK72xxBLETracker::on_ota_global_state(ota::OTAState state, float progress,
                                            ota::OTAComponent *comp) {
   if (state == ota::OTA_STARTED) {
     this->scan_continuous_before_ota_ = this->scan_continuous_;
+    this->scan_requested_before_ota_ = this->scan_requested_;
     this->stop_scan();
-  } else if ((state == ota::OTA_ERROR || state == ota::OTA_ABORT) && this->scan_continuous_before_ota_) {
+  } else if (state == ota::OTA_ERROR || state == ota::OTA_ABORT) {
     // On success the device reboots, so restore only on a failed/aborted update;
     // loop() restarts the scan on its next iteration (continuous idle branch).
-    this->scan_continuous_before_ota_ = false;
-    this->scan_continuous_ = true;
+    if (this->scan_continuous_before_ota_) {
+      this->scan_continuous_before_ota_ = false;
+      this->scan_continuous_ = true;
+    }
+    // A one-shot request that was still pending (latched, retrying) when the
+    // OTA paused scanning is re-latched, not dropped — loop() resumes the retry.
+    if (this->scan_requested_before_ota_) {
+      this->scan_requested_before_ota_ = false;
+      this->scan_requested_ = true;
+    }
   }
 }
 #endif  // USE_OTA_STATE_LISTENER
@@ -62,25 +71,11 @@ void BK72xxBLETracker::loop() {
   const uint32_t now = millis();
   if (this->scan_continuous_) {
     if (!this->scan_running_) {
-      // Rate-limit (re)start attempts. The controller start can fail (no idle activity
-      // handle, WiFi/BLE coexistence) and leave scan_running_ false; retrying every
-      // main-loop iteration would spin the single-core CPU and starve WiFi (device
-      // becomes unresponsive). The interval backs off with consecutive failures so a
-      // controller that never comes up polls slowly and quietly.
-      const uint8_t doublings = std::min<uint8_t>(this->failed_start_count_, SCAN_START_RETRY_MAX_DOUBLINGS);
-      if (now - this->last_scan_start_attempt_ >= (SCAN_START_RETRY_MS << doublings)) {
-        this->last_scan_start_attempt_ = now;
-        this->start_scan_();
-        if (this->scan_running_) {
-          this->failed_start_count_ = 0;
-        } else if (this->failed_start_count_ < SCAN_START_RETRY_MAX_DOUBLINGS) {
-          ++this->failed_start_count_;
-          if (this->failed_start_count_ == SCAN_START_RETRY_MAX_DOUBLINGS) {
-            ESP_LOGW(TAG, "Scan start keeps failing; retrying every %" PRIu32 " s",
-                     (SCAN_START_RETRY_MS << SCAN_START_RETRY_MAX_DOUBLINGS) / 1000);
-          }
-        }
-      }
+      // A start that succeeded re-anchored the period timer from a later millis(),
+      // so the stale `now` below would underflow the comparison and fire
+      // on_scan_end() for a scan that just began. Resume next iteration.
+      if (this->try_start_with_backoff_(now))
+        return;
     }
     // Period timer: fire on_scan_end() once per scan_duration_ window, mirroring
     // esp32_ble_tracker::cleanup_scan_state_(). Gated on scan_started_once_ so a scan
@@ -98,11 +93,51 @@ void BK72xxBLETracker::loop() {
 
   // Non-continuous mode: run for scan_duration_ ms, then stop and fire on_scan_end.
   // Restart is driven externally (e.g. api: on_client_connected:).
+  //
+  // A requested start that failed (same controller failures the continuous branch
+  // absorbs) is retried with the same backoff — otherwise a failed one-shot start
+  // would be silent: the scan never runs, stop_scan_() is never reached and
+  // on_scan_end() never fires, leaving period-keyed consumers waiting forever.
+  if (this->scan_requested_ && !this->scan_running_) {
+    // Same stale-`now` hazard as the continuous branch: start_scan_() stamps
+    // scan_start_time_ from a later millis(), so the duration check below would
+    // underflow and stop the scan in the iteration that started it.
+    if (this->try_start_with_backoff_(now))
+      return;
+  }
   if (this->scan_running_ && now - this->scan_start_time_ >= this->scan_duration_) {
     this->stop_scan_();
   }
 }
 
+bool BK72xxBLETracker::try_start_with_backoff_(uint32_t now, bool force) {
+  // Rate-limit (re)start attempts. The controller start can fail (no idle activity
+  // handle, WiFi/BLE coexistence) and leave scan_running_ false; retrying every
+  // main-loop iteration would spin the single-core CPU and starve WiFi (device
+  // becomes unresponsive). The interval backs off with consecutive failures so a
+  // controller that never comes up polls slowly and quietly.
+  //
+  // force bypasses the gate for an explicit user start (start_scan()) — but
+  // only while the failure streak is clean. Once the controller is failing,
+  // even user-initiated attempts respect the backoff, so a start_scan() action
+  // on a short cadence cannot hammer a failing controller; the attempt stays
+  // inside the failure accounting below either way.
+  const uint8_t doublings = std::min<uint8_t>(this->failed_start_count_, SCAN_START_RETRY_MAX_DOUBLINGS);
+  if ((!force || this->failed_start_count_ != 0) &&
+      now - this->last_scan_start_attempt_ < (SCAN_START_RETRY_MS << doublings))
+    return false;
+  this->last_scan_start_attempt_ = now;
+  this->start_scan_();
+  if (!this->scan_running_ && this->failed_start_count_ < SCAN_START_RETRY_MAX_DOUBLINGS) {
+    ++this->failed_start_count_;
+    if (this->failed_start_count_ == SCAN_START_RETRY_MAX_DOUBLINGS) {
+      ESP_LOGW(TAG, "Scan start keeps failing; retrying every %" PRIu32 " s",
+               (SCAN_START_RETRY_MS << SCAN_START_RETRY_MAX_DOUBLINGS) / 1000);
+    }
+  }
+  return this->scan_running_;
+}
+
 void BK72xxBLETracker::dump_config() {
   ESP_LOGCONFIG(TAG,
                 "BK72xx BLE Tracker:\n"
@@ -153,13 +188,29 @@ void BK72xxBLETracker::on_scan_report(const bk72xx_ble::BLEScanReport &report) {
 void BK72xxBLETracker::start_scan() {
   // Mirrors esp32_ble_tracker::start_scan(): caller sets scan_continuous_ via
   // set_scan_continuous() first, then calls start_scan() to begin scanning.
-  if (!this->scan_running_) {
-    this->start_scan_();
-  }
+  //
+  // Nothing to do while a scan is already running: latching here would leave
+  // scan_requested_ set after that scan ends and silently restart a one-shot
+  // scan nobody asked for.
+  if (this->scan_running_)
+    return;
+
+  // The request is latched: if this immediate attempt fails (controller busy,
+  // WiFi/BLE coexistence), loop() keeps retrying it with backoff even in
+  // non-continuous mode, so a one-shot start cannot fail silently.
+  //
+  // Routed through the backoff helper (forced: the user asked for an immediate
+  // attempt) so a failure here still counts toward the backoff escalation and
+  // its WARN. The force bypass only applies while the failure streak is clean —
+  // against a failing controller, repeated start_scan() calls are rate-limited
+  // like any other attempt.
+  this->scan_requested_ = true;
+  this->try_start_with_backoff_(millis(), /* force= */ true);
 }
 
 void BK72xxBLETracker::stop_scan() {
   this->scan_continuous_ = false;
+  this->scan_requested_ = false;  // also cancels a pending (not yet successful) start
   this->stop_scan_();
 }
 
@@ -177,6 +228,8 @@ void BK72xxBLETracker::start_scan_() {
 
   const uint32_t now = millis();
   this->scan_running_ = true;
+  this->scan_requested_ = false;  // the latched one-shot request is satisfied
+  this->failed_start_count_ = 0;  // reset here so direct starts clear the backoff too
   this->scan_start_time_ = now;
   // Log every explicit start at DEBUG — stop_scan_() logs every stop at DEBUG, and
   // in non-continuous mode each period is an explicit start, so asymmetric logging
diff --git a/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.h b/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.h
index 9c70246f05..3d32798622 100644
--- a/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.h
+++ b/esphome/components/bk72xx_ble_tracker/bk72xx_ble_tracker.h
@@ -112,8 +112,15 @@ class BK72xxBLETracker : public Component,
  protected:
   void start_scan_();
   void stop_scan_();
+  /// Attempt a rate-limited (re)start; returns true when the scan is running,
+  /// which means the caller must not compare its cached millis() against the
+  /// timestamps start_scan_() just refreshed. force bypasses the rate gate for
+  /// an explicit user start only while the failure streak is clean; a failing
+  /// controller rate-limits forced attempts too. Failure accounting always runs.
+  bool try_start_with_backoff_(uint32_t now, bool force = false);
 
   bool scan_running_{false};
+  bool scan_requested_{false};  // latched start_scan() request not yet running; loop() retries with backoff
   // Defaults: the BK reference — 30 % duty cycle
   // (interval 100 ms / window 30 ms), in 0.625 ms BLE units.
   uint32_t scan_interval_{160};  // 160 × 0.625 ms = 100 ms
@@ -122,6 +129,7 @@ class BK72xxBLETracker : public Component,
   bool scan_continuous_{true};
 #ifdef USE_OTA_STATE_LISTENER
   bool scan_continuous_before_ota_{false};  // continuous mode saved at OTA start, restored on OTA failure
+  bool scan_requested_before_ota_{false};   // pending one-shot latch saved at OTA start, re-latched on OTA failure
 #endif
   uint32_t scan_start_time_{0};