The rule that keeps you honest
Any video CPM change should be measured with reach, VTR and outcome rate held flat. If CPM fell 25% and completed views fell 25%, nothing was saved, the media was just downgraded. Freeze a baseline first, then judge every lever against it.
1. Buy on the bid cycle, not the daily average
Auction prices move intraday and by day of week. A flat target CPM pays the average all week. Re-pricing bids against the actual cycle is usually the single largest reduction available, and it changes nothing a viewer sees.
2. Collapse duplicate supply paths
The same impression is often reachable through several exchanges at different clearing prices, each with its own fee. Pick the shortest path per publisher and the fee stack shrinks without touching reach.
3. Cap frequency where returns fall off
Impressions eight through twenty to the same user cost full price and convert far below the first three. Capping there frees budget for unreached users, which lowers effective CPM and raises reach at once.
4. Remove audience overlap between line items
Prospecting and retargeting pods that share members bid against each other in the same auction. Mutual exclusion lists stop you from raising your own clearing price.
5. Daypart and geo-price deliberately
A view at 3pm in a dense metro is not priced like a view at 1am elsewhere. Setting modifiers by measured elasticity, not by intuition, moves CPM without cutting volume.
6. Prune inventory that only looks cheap
Low-quality app and made-for-advertising inventory drops your CPM and your completion rate together. Removing it raises headline CPM and lowers cost per outcome, track both so the trade is visible.
7. Rotate creative before fatigue prices you up
As relevance drops, platforms charge more for the same placement. Refreshing creative on a schedule keeps quality signals up and the price down.
How to sequence them
- Week 1: frequency caps and audience overlap: fastest, lowest risk.
- Week 2: supply-path cleanup and inventory pruning.
- Week 3+: bid-cycle pricing and dayparting, which need enough data to model elasticity.
Ranges and sequencing here reflect patterns we see across accounts, not guaranteed outcomes. The size of each lever depends on how your account is already structured.