The method
A verification number is worth exactly as much as the method behind it. Ours is deliberately boring, and deliberately checkable — this page is the whole of it, including the parts that are inconvenient for us.
One impression, one gate
Start at the DSP's own count. Every impression it billed walks the gates below and stops at the first one it fails. It gets one row, at that stage, with that reason.
An impression that clears every gate is valid — the thing you were actually buying. Add all the stages back up and you are at the DSP's count again.
This is the invariant, and it is enforced, not aspired to. A fixture with known totals runs on every change. If the rows stop summing to the DSP's count — including for impressions whose beacons straddle an hour boundary, which is a real bug we shipped once — the build fails and the change does not merge.
When a partner cannot pass our ID
Some partners cannot put our impression ID in the tag. They send their own auction ID and we resolve it through the logs. That is a different quality of evidence, so it is reported as a different tier and never folded into an exact total.
| Tier | How it was matched | Summable? |
|---|---|---|
exact | Our impression ID, present on both sides. | Yes |
partner_id | The partner's own auction ID, resolved through the logs. Still an exact join. | Yes, when labelled |
fuzzy | Matched on time, campaign and creative within a window. | No |
none | Evidence with no usable key at all. | No |
A match rate that quietly includes guesses is worse than no match rate, because you would put it in a client report.
What the DSP never billed
Some impressions have SSP or beacon evidence but no DSP record at all. They are handled by a separate daily pass and carry stage = mismatch, so they are visible without ever entering a billed total.
ssp_only— the SSP served it and the DSP has no record.tracker_only— a beacon for an impression neither log contains.unattributed— evidence with no usable key.
These are partner logging gaps, not wasted spend, and the reports treat them that way.
What we will not do
- Fill gaps with estimates. An unmeasured impression is reported as unmeasured. Modelled numbers are indistinguishable from real ones the moment they enter a spreadsheet.
- Trust the tag on anything that controls access. The advertiser an impression belongs to comes from the DSP log, never from the page — the tag's attributes are not covered by its signature.
- Present preliminary numbers as settled. Every response says which it is.
- Silently change history. A correction is a replay of a whole day, and reruns are idempotent by construction.