Myth: Log-level data is the unfiltered ground truth

Myth: Log-level data is the unfiltered ground truth

Buyers treat raw log-level data — per-impression records of bids, wins, prices, and outcomes — as the bedrock reality, the thing you reconcile everything else against. "The logs don't lie."

The logs do, in specific and knowable ways.

1. Loss logs are systematically incomplete. SSPs report loss reasons (loss reason codes) inconsistently; many auctions you lost return no clearing price or a generic code, so your view of the bid landscape is censored at exactly the points you most need.

2. Timestamps drift across systems. The DSP's bid log and the SSP's win notification can disagree by hundreds of milliseconds, breaking naive joins and double-counting or dropping events.

3. Win notifications (the win notice fired when you win) get lost in transit or fire late, so won-impression logs under-report. Some impressions you won never render, yet appear as wins.

4. Sampled logs are common at scale — vendors deliver 1-in-N rows to control volume — and buyers forget the sampling rate when computing rates, skewing every downstream ratio.

The right posture is to treat logs as instrument readings with known biases: model the censoring on losses, reconcile timestamps to a tolerance window, and confirm sampling rates before trusting any ratio.

Why it matters: log-level data is the best evidence you have, not raw truth. Analysis that assumes completeness inherits every gap in the bidstream's plumbing.