adfire.digital Reading the machinery behind a display ad slot — one bid request at a time
adfire.digital  /  Reconciling two totals

Reconciling two totals without guessing

When two systems report different impression counts, the useful question is not which is right but where the two definitions part company. This is a procedure for locating that point, and it ends either with a named cause or with an explicit statement that the cause is not determinable from outside.

The procedure

Each step either explains part of the gap or eliminates a candidate cause. Do them in order; later steps assume earlier ones are settled.

  1. Align the window before comparing anything.Re-pull both totals over a period long enough that a day-boundary offset cannot matter — a full week, ending at least one full day in the past. If the gap collapses, the cause was the boundary and you are finished.

  2. Write down each system's counting moment.Dispatch and render are different events. Two systems counting at different events must disagree, and the direction of the disagreement is predictable: the dispatch-side total is the larger one.

  3. Separate delivery from viewability.Confirm whether each total is a delivery count or a viewable count. Comparing one against the other produces a gap that no investigation will close, because the two are measuring different things by design.

  4. Check whether the measurement call is being prevented from loading.Where it is, the observation is missing while the delivery is not. This asymmetry only ever reduces the observing system's total, never the serving one's.

  5. State what remains unexplained.After the four steps above, any residual gap is either inside a system you cannot inspect or below the level worth chasing. Writing "unexplained residual, N%" is a finding; distributing it across plausible causes is not.

What this procedure cannot do

It can

  • Establish whether the two totals were ever comparable in principle.
  • Attribute a gap to a documented definitional difference.
  • Bound the size of the genuinely unexplained portion.

It cannot

  • Adjudicate between two systems' internal filtering decisions from outside.
  • Recover events that were never observed by either party.
  • Convert a viewable count into a delivery count, or the reverse.

The specifications behind this page