In-app bidding, without the sales pitch.

01Definition

In-app bidding is an auction model where every demand source bids on every impression at the same moment, and the highest bid wins. There is no ordered asking sequence: the decision is one comparison across all sources, made in the milliseconds between the ad request and the render. On the platform side it is usually called header bidding or real-time bidding; on the app side, in-app bidding; the mechanism is the same.

02The waterfall, explained

The waterfall is the legacy sequencing model. Demand sources are ordered by an assumed price, usually a manually set CPM target. The impression is offered to the first source; if it declines or misses a timeout, the second is asked, then the third, until someone accepts. The model makes two wrong assumptions: that yesterday's price ranking is today's truth, and that every source answers within its slot. What it leaves undiscovered is any impression worth more to a lower rung than the price target above it. Multiply that gap across a year and it is not a rounding error.

03Why bidding is not automatically more revenue

Bidding replaces the waterfall's sequencing with three new problems. Latency: every source is invited to every auction, and the auction must finish inside the render budget, so slow sources either get faster or get filtered, sometimes taking bids with them. Bid density: in geos where only two sources bid, simultaneous competition barely competes, and the waterfall's negotiated rates can outperform thin parallel bidding. Floor dynamics: a floor set for a waterfall rung can suppress bids in an open auction, because bidders see the floor and adjust. Switching is a re-optimization, not a toggle.

04Unified auction

A unified auction combines bidding sources and non-bidding networks in one decision: RTB demand bids in parallel while non-bidding networks run automated waterfall targets evaluated in the same comparison. This resolves the mediation problem that pure bidding creates, where networks without RTB endpoints would fall out of monetization entirely. The mediation problem the unified model creates is subtler: waterfall targets are estimates, and mixing estimated values with real bids means the optimizer, not the market, sets the price of the non-bidding share. That is acceptable, provided the targets are auditable and the bidding share is genuinely at par.

05Price and priority logic

Three interactions matter in practice. Ties: when bids match, the tiebreaker should be published, deterministic and free of self-preference.Floors: the floor is a private minimum, applied identically to every source including house demand. Direct campaigns: guaranteed and preferred deals carry their deal terms into the same auction and win on price against open demand, and competitive separation applies only after the price comparison.

06Switching without losing revenue

1. Map the current waterfall: every network, every placement, every target price.
2. Integrate the unified SDK and recreate the placement map in the new system.
3. Run parallel on a cohort: a percentage of traffic on the new stack, the rest on the old, chosen randomly and held stable.
4. Compare per-placement revenue, fill and latency between cohorts over at least one full optimization cycle.
5. Move placements over as they prove out, and keep the old targets as fallbacks during the transition window.

07Measurement: how to actually verify bidding won

Before-and-after eyeballing is not measurement: seasonality, app releases and audience shifts all happen at the same time as a bidding switch. The credible design is a cohort A/B: randomized traffic split, both stacks live, comparing revenue per impression, fill and retention per cohort over the test window. Read revenue per impression rather than total revenue, because the cohorts differ in size. Read retention alongside revenue, because monetization that costs retention is not a win.

08Pitfalls

• Thin bid density per geo. Two bidders do not make an auction. In thin geos, keep negotiated waterfall targets live inside the unified decision.
• Over-refreshed test conclusions. Testing on a handful of placements and extrapolating across the app repeats the waterfall's original sin at smaller scale.
• Discrepancy confusion. Networks count impressions by their own definitions. A revenue gap between dashboards is usually definition drift, not theft; the reporting page explains the reconciliation.
• Floor drift. Floors imported from the waterfall and never revisited suppress auction bids quietly. Review floors per placement after the switch stabilizes.