diff --git a/internal/scraper/scraper.go b/internal/scraper/scraper.go
index c491e61b..b56c47b3 100644
--- a/internal/scraper/scraper.go
+++ b/internal/scraper/scraper.go
@@ -216,14 +216,8 @@ func (s *Scraper) Run(ctx context.Context) {
 	// TODO(mem): keep count of the number of successive errors and
 	// collect logs if threshold is reached.
 
-	var (
-		frequency = ms(s.check.Frequency)
-		offset    = ms(s.check.Offset)
-	)
-
-	if offset == 0 {
-		offset = randDuration(min(frequency, maxPublishInterval))
-	}
+	frequency := ms(s.check.Frequency)
+	offset := computeOffset(ms(s.check.Offset), frequency, timeFromNs(s.check.Created), time.Now())
 
 	scrapeHandler := scrapeHandler{scraper: s}
 
@@ -334,6 +328,48 @@ func ms(n int64) time.Duration {
 	return time.Duration(n) * time.Millisecond
 }
 
+func timeFromNs(ns float64) time.Time {
+	sec := int64(math.Floor(ns / 1e9))
+	nsec := int64(math.Mod(ns, 1e9))
+	return time.Unix(sec, nsec)
+}
+
+func computeOffset(offset, frequency time.Duration, t0, now time.Time) time.Duration {
+	if now.Sub(t0) < frequency {
+		// The check was created less than the frequency ago, we should
+		// starting running it right away.
+		if offset != 0 {
+			return offset
+		}
+
+		return randDuration(min(frequency, maxPublishInterval))
+	}
+
+	// The check was created more than the frequency ago, so we need to
+	// compute the time until the next time the check should run.
+	//
+	// Compute the number of runs since t0, add one for the next run and
+	// multiply by the frequency in order to obtain its timestamp. Finally,
+	// compute the remaining time until that timestamp.
+
+	runs := (now.UnixMilli() - t0.UnixMilli()) / frequency.Milliseconds()
+
+	timeUntilNextRun := t0.Add(time.Duration(runs+1) * frequency).Sub(now)
+
+	if timeUntilNextRun <= maxPublishInterval {
+		return timeUntilNextRun
+	}
+
+	// The reason why we need to ignore the computed offset is that the
+	// check ran in the past, and it's possible that it was filling the
+	// data with repeated samples that we no longer have access to. We
+	// cannot wait until the next run because that might be a long time
+	// from now, creating a gap in the data. Instead we wait for a random
+	// value that avoids creating gaps (assuming the last published sample
+	// was recent).
+	return randDuration(maxPublishInterval)
+}
+
 func randDuration(d time.Duration) time.Duration {
 	return time.Duration(rand.Int63n(int64(d)))
 }
diff --git a/internal/scraper/scraper_test.go b/internal/scraper/scraper_test.go
index 9734a171..c45f1f60 100644
--- a/internal/scraper/scraper_test.go
+++ b/internal/scraper/scraper_test.go
@@ -1991,3 +1991,113 @@ func TestTickWithOffset(t *testing.T) {
 		})
 	}
 }
+
+func TestTimeFromNs(t *testing.T) {
+	testcases := map[string]struct {
+		ns       float64
+		expected time.Time
+	}{
+		"zero": {
+			ns:       0,
+			expected: time.Unix(0, 0),
+		},
+		"one": {
+			ns:       1,
+			expected: time.Unix(0, 1),
+		},
+		"2020-01-01 00:00:00.000000000": {
+			ns:       1577836800 * 1e9,
+			expected: time.Unix(1577836800, 0),
+		},
+		"2024-07-02 21:21:50.123456768": {
+			ns:       1719955310*1e9 + 123456789,
+			expected: time.Unix(1719955310, 123456789),
+		},
+		"2262-04-11 23:47:15.999999999": {
+			ns:       9223372035*1e9 + 999999999, // This is close to the maximum value that can be represented by a time.Time
+			expected: time.Unix(9223372035, 999999999),
+		},
+	}
+
+	for name, tc := range testcases {
+		t.Run(name, func(t *testing.T) {
+			actual := timeFromNs(tc.ns)
+			// Why UnixMicro instead of UnixNano? Because some
+			// precision is lost during the conversion from int64
+			// to float64, and getting this right at the microsecond
+			// level is good enough.
+			require.InDelta(t, tc.expected.UnixMicro(), actual.UnixMicro(), 1)
+		})
+	}
+}
+
+func TestComputeOffset(t *testing.T) {
+	t0 := time.Unix(1_000_000, 0)
+
+	testcases := map[string]struct {
+		frequency time.Duration
+		offset    time.Duration
+		now       time.Time
+		expected  time.Duration
+	}{
+		"zero": {
+			offset:    0,
+			frequency: 60 * time.Second,
+			now:       t0.Add(0),
+			expected:  0,
+		},
+		"1s": {
+			offset:    1 * time.Second,
+			frequency: 60 * time.Second,
+			now:       t0.Add(0),
+			expected:  1 * time.Second,
+		},
+		"30s": {
+			offset:    30 * time.Second,
+			frequency: 60 * time.Second,
+			now:       t0.Add(0),
+			expected:  30 * time.Second,
+		},
+		"created 100 seconds ago": {
+			offset:    0 * time.Second,
+			frequency: 60 * time.Second,
+			now:       t0.Add(100 * time.Second),
+			expected:  20 * time.Second, // 100 - 60 = 40 -> 60 - 40 = 20
+		},
+		"created 1000 seconds ago": {
+			offset:    0 * time.Second,
+			frequency: 60 * time.Second,
+			now:       t0.Add(1000 * time.Second),
+			expected:  20 * time.Second, // 1000 / 60 = 16 -> 1000 - 60 * 16 = 40 -> 60 - 40 = 20
+		},
+		"slow check": {
+			offset:    0 * time.Second,
+			frequency: 5 * time.Minute,
+			now:       t0.Add(1000 * time.Minute),
+			expected:  0,
+		},
+		"slow check close to next run": {
+			offset:    0 * time.Second,
+			frequency: 5 * time.Minute,
+			now:       t0.Add(999 * time.Minute),
+			expected:  1 * time.Minute,
+		},
+		"slow check just ran": {
+			offset:    0 * time.Second,
+			frequency: 5 * time.Minute,
+			now:       t0.Add(1001 * time.Minute),
+			expected:  0,
+		},
+	}
+
+	for name, tc := range testcases {
+		t.Run(name, func(t *testing.T) {
+			actual := computeOffset(tc.offset, tc.frequency, t0, tc.now)
+			if tc.expected != 0 {
+				require.Equal(t, tc.expected, actual)
+			} else {
+				require.LessOrEqual(t, actual, maxPublishInterval)
+			}
+		})
+	}
+}