Bid shading: the margin you never see in automated buying
Bid shading is sold as a saving. It is also a margin pool, measured against a counterfactual the shader controls. Here is how to audit it.
Bid shading is the most profitable feature in ad tech that nobody on the buy side has a line item for. Every DSP and most networks now shade bids as standard: the algorithm predicts the lowest price that would still win a first-price auction and bids there rather than at the advertiser's full value. The pitch is that it saves the advertiser money. The claim in this piece is that it also creates a pool of margin whose size gets measured against a counterfactual only the shader can see, and that the buy side has quietly accepted this as the price of automation.
The 2026 edition of the AppsFlyer State of Gaming for Marketers report, published yesterday with Unity and Newzoo, put paid installs at 14.1 billion out of 24.8 billion tracked across the apps it measured, with paid share up ten percent year on year and ad impressions up twenty percent. More of that spend than ever runs through automated bidding. Where the margin sits inside those bids is no longer a specialist concern.
How the saving is calculated
A shading algorithm works from the win-rate curve: for a given impression type, what is the probability of winning at each bid level. If the advertiser's implied value is ten and the curve says a bid of six wins ninety percent of the time, the shader bids somewhere near six and reports a saving of roughly four against the unshaded bid.
Three things in that sentence deserve suspicion. The platform derives the implied value from the advertiser's target, so the baseline is the platform's own construction. The win-rate curve comes out of auctions the platform took part in, which makes it a picture of that platform's competitive footprint rather than of the market. And the saving sits against a bid the advertiser never asked to place. An advertiser who set a target return was never going to pay ten; the platform chose ten as the reference point.
None of this means shading is fake. Shading against a hard floor is real and valuable. The question is who captures the difference between the shaded bid and the clearing price, and who gets to report it.
Where the margin pools
There are three places the surplus can go, and most contracts allow all three.
Take rate is the explicit one. A percentage of media spend goes to the DSP regardless of shading, and this is the visible margin, the one buyers negotiate.
Next comes the spread inside a managed price. When a network sells on a CPI or a target-cost basis and buys impressions on a CPM basis, the gap between what it charges and what it pays is its margin, and shading widens that gap without touching the advertiser's invoice. The advertiser sees a CPI at or below target and concludes the network did well. The network sees the same CPI and a better gross margin. Both are correct.
The softest of the three is a shading model tuned to hold win rates high rather than prices low. Winning more auctions at slightly higher prices grows the platform's spend under management, which is what most platform incentives reward. The advertiser's target still gets hit on average, because the target was a ceiling with slack in it. Nobody lied. The slack simply went to the platform rather than back to the buyer.
The second-order effect: floors chase shading
The consequence most coverage misses is that shading is not a one-sided game. Publishers and their mediation layers see win rates and clearing prices too. When shading pushes clearing prices down, the rational response is to raise floors, either as fixed reserves or as dynamic floors that learn from the same bid distributions the shaders are learning from.
The equilibrium is an arms race that costs both sides compute and leaves clearing prices roughly where they were, with the difference that the reported saving from shading stays large because the unshaded reference bid is still the baseline. So the advertiser hears they saved four on every impression while paying approximately what a truthful second-price rule would have charged them. The saving is real relative to a bid they never placed and imaginary relative to the market they actually operate in.
Dynamic floors also mean the win-rate curve the shader learned last quarter is wrong this quarter. Shading models degrade, and a degraded shader either overpays or loses impressions. Neither shows up as a shading problem in reports; they surface as CPI drift or delivery softness and get blamed on creative.
Auditing the margin you cannot see
A buyer who wants to know what shading is worth to them has two routes, one expensive and one cheap.
The expensive route is log-level data. Where a DSP will hand over impression-level records carrying both the bid placed and the price cleared, the buyer can compute the actual surplus and hold it against the reported saving. Most DSPs do this for large accounts on request and for smaller ones with reluctance. The reluctance is information.
The cheap route is a controlled experiment using the target itself. Take a stable campaign, run it for two weeks with the target lowered by a fixed increment, say fifteen percent, and leave the budget alone. Three outcomes are possible, and each one is diagnostic.
If delivery holds and cost falls to the new target, the platform had slack and was banking it, which means the shading saving it reported to you was flowing to the platform. If delivery falls sharply and cost lands near the new target, the target was binding and the shading was working roughly as described. If delivery holds and cost does not fall, the platform is not honouring the target as a ceiling in the way you assumed, and a harder conversation is due.
An illustrative case. A mid-core title spends 50,000 a week on one DSP at a reported CPI of 5 against a target of 5.20, with reported shading saving at fifteen percent of media. The buyer drops the target to 4.50 for a fortnight. Delivery falls by ten percent and CPI lands at 4.60. So the campaign was leaving roughly ten percent on the table, not because the shader was bad but because the target was generous and the platform had no incentive to return the difference. On a 2.6 million annual budget, that ten percent is real money no dashboard was ever going to surface.
What to ask for in the next contract
Buyers negotiate take rate because it is visible. The margin that matters more is the spread and the slack, and reporting rights govern both of those rather than percentages. The clause worth fighting for is access to bid and clear prices at the impression level, or at minimum a monthly reconciliation of average bid against average clearing price by placement type.
A platform that will not hand over either is asking you to trust that the saving it reports uses a baseline you would have chosen. The fifteen percent target experiment costs a fortnight and answers that question without the platform's help.
