diff --git a/views/ndt/referenced-by/referenced-by/downloads.sql b/views/ndt/referenced-by/referenced-by/downloads.sql
index 08dc4adc..9214a01d 100644
--- a/views/ndt/referenced-by/referenced-by/downloads.sql
+++ b/views/ndt/referenced-by/referenced-by/downloads.sql
@@ -1,6 +1,24 @@
 #standardSQL
--- All good quality download tests
-SELECT * FROM `%s.ndt.recommended`
+SELECT
+      -- All good quality download tests
+       *,
+      -- A synthetic column that conforms to our best practices.
+      -- The times are in microseconds, and dividing total throughput by total
+      -- microseconds yields mean throughput in Megabits per second.  This
+      -- query excludes any time during which the connection was quiescent,
+      -- and as such is a fairer representation of how fast we were able to
+      -- push data down the pipe as compared to just dividing by
+      -- web100_log_entry.snap.Duration
+      --
+      -- TODO: Delete this once https://github.com/m-lab/etl/issues/663 is
+      -- resolved.
+      STRUCT(
+        8 * (web100_log_entry.snap.HCThruOctetsAcked /
+        (web100_log_entry.snap.SndLimTimeRwin +
+         web100_log_entry.snap.SndLimTimeCwnd +
+         web100_log_entry.snap.SndLimTimeSnd)) AS mean_download_throughput_mbps
+      ) as alpha
+FROM `%s.ndt.recommended`
 WHERE
   -- download direction, and at least 8KB transfered
   connection_spec.data_direction IS NOT NULL