The Programmatic Bidding Process: What a DSP Actually Does Before It Bids

Three decisions happen before the bid is even sent

A single bid is the visible output of a process most explainers skip past. First, the DSP values the specific impression — who the viewer likely is, what the page or app context is, and how well that matches the campaign's goal — and calculates the maximum it would theoretically be willing to pay. Second, it applies bid shading, lowering that maximum toward a predicted clearing price rather than bidding the full value, because bidding the true maximum in a first-price auction routinely means overpaying relative to what was actually needed to win. Third, it checks pacing — whether the campaign's budget for the day, hour, or flight allows a bid on this particular opportunity right now, or whether spending here would exhaust the budget before better opportunities later in the period show up.

All three of those decisions happen inside the same auction window as the rest of the bid request-response cycle — the DSP doesn't get extra time to value, shade, and pace a bid separately from responding to the auction. That's part of why DSP infrastructure is judged as much on computational efficiency as on the sophistication of its models: a valuation model too slow to finish inside the timeout doesn't get to bid at all, no matter how accurate it would have been given more time.

Bid shading, worked through with real numbers

Say a DSP values an impression at $5.00 CPM — the true maximum the campaign's goal justifies paying. Most exchanges today run first-price auctions, meaning the winner pays exactly what it bid, unlike the older second-price model where the winner paid just above the next-highest bid. Bidding the full $5.00 value in a first-price auction routinely means paying more than necessary, because the second-highest bidder might only have offered $3.20. A bid-shading model looks at the historical clearing prices for similar impressions — this ad size, this domain, this time of day, this win-rate pattern — and submits a shaded bid instead, say $3.60: high enough to have a strong chance of winning, low enough to leave real margin between the bid and the true valuation. Modern DSPs run this calculation with reinforcement learning that updates continuously as new clearing-price data comes in, not a fixed discount applied to every bid.

Pacing: the decision that has nothing to do with the auction itself

Pacing is a budget-management layer sitting on top of the bidding logic, and it's easy to underestimate how much it decides. A campaign with a $10,000 daily budget that spends it all in the first three hours has effectively opted out of every opportunity in the other twenty-one — including opportunities that might have been cheaper or better-matched to the goal. Pacing algorithms monitor spend velocity against the remaining budget and remaining time, holding bids back or loosening them to keep delivery smooth across the full period rather than front-loaded or exhausted early. This is a mechanical control, not an optimization one — it can produce a technically "paced" campaign that's still bidding on mediocre inventory late in the day simply because the budget curve says it should be spending.

Even-pacing and asap-pacing are the two common settings worth knowing by name. Even pacing spreads spend as smoothly as possible across the whole flight, the safer default for most campaigns. ASAP pacing spends as fast as the auction allows, useful for a time-sensitive promotion or a launch where reaching the audience quickly matters more than smooth delivery — but it also means the campaign has no budget left to bid on better opportunities that might show up later in the day, which is the trade-off worth weighing before choosing it.

A worked example: what happens to the whole process during a demand spike

Every one of these steps assumes normal competitive conditions, and they all get stress-tested during a high-demand period. Prime week roughly doubles DSP click costs across our own book — a 116% agency-wide increase, measured directly, not estimated. That isn't a targeting failure and it isn't something a bid-shading model can algorithm its way around, because every advertiser on the platform is bidding into the exact same competitive week at the same time — the clearing price itself has moved, and a shading model correctly reflects that by bidding higher, not by finding a clever workaround. The practical consequence for planning: a campaign launched a fortnight before a major tentpole event will be judged, if measured naively, on its single most expensive and most competitive stretch.

The mistake, and what to check when a campaign stops winning

The common mistake is treating a drop in win rate as a targeting or creative problem before checking whether the market itself moved. If win rate falls and CPM rises together, check the calendar first — a seasonal or tentpole demand spike explains far more sudden shifts than most teams assume. If win rate falls while CPM stays flat, that's more likely a pacing or bid-shading issue — the algorithm may be holding bids back because of an early-period spend rate, not because the opportunities got worse. Separating a market-driven shift from an internal pacing decision before making any changes avoids solving the wrong problem.

Where Dr. DSP fits

Dr. DSP is Amazon DSP run as a managed product by Full Circle, which has managed more than $500M in revenue across 100+ brands. Either excluding tentpole weeks from a performance read, or launching well ahead of them so a campaign has already built pacing history before the spike hits, is a planning decision that has nothing to do with creative or targeting quality. A reader who never buys anything from us should still leave this page able to tell a market-driven cost spike apart from an actual account problem.