First-price auctions quietly changed what a bid means

The argument of this piece is that the industry's move from second-price to first-price auctions did more to shape today's buying tools than any privacy change, and that almost nobody on the buy side noticed because the change happened inside the tools rather than to them. A bid used to be a statement of value. It is now a strategic guess, and the guessing has been outsourced to the same platforms that run the auction.

That outsourcing is not a conspiracy. It is the natural equilibrium of a first-price market. But it explains why buyers now set targets rather than bids, why their visibility into clearing prices has collapsed, and why the automated bidding products that dominate spend are so hard to evaluate against each other.

What a second-price bid used to say

In a second-price auction the winner pays the second-highest bid, not their own. The elegant property of this design, well documented in auction theory, is that the dominant strategy is to bid what the impression is actually worth to you. Bid lower and you risk losing impressions you would have profited from. Bid higher and you risk winning ones you would not. The auction rewards truth.

For a media buyer that meant a bid was a piece of internal information made external. If your model said a user from this placement was worth six units in expected revenue, you bid six, and the auction returned any surplus. Price discovery was a by-product. You could see, across thousands of impressions, what the market's second-highest valuation of your inventory was, and that told you something about your competitors.

The system had a weakness that eventually killed it. Publishers and exchanges could not resist adding floors, and the daisy-chained waterfall of exchanges created situations where a second-price win in one auction became a first-price bid in the next. Buyers found themselves paying prices that did not match either rule. The move to first-price everywhere, which the large web exchanges completed around 2019 and which in-app bidding mediation adopted as its default, was sold as a simplification. Pay what you bid, one rule, no hidden floors.

What first-price does to the bidder

Under first-price, bidding your true value is a mistake. You pay exactly what you bid, so the surplus that a second-price auction returned to you is now yours only if you bid below your value. The optimal bid depends on what you think everyone else will bid, which depends on what they think you will bid. Truth stops being a strategy and becomes a starting point for a prediction problem.

This is where the buy side lost something it did not know it owned. Solving the prediction problem requires a view of the distribution of competing bids on each impression. Individual advertisers do not have that view. Platforms and DSPs do, because they see every auction they participate in across every client. The rational response for an advertiser was to hand the bidding over to whoever had the data, and that is what happened. Target-cost and target-return bidding did not become dominant because they were more convenient. They became dominant because first-price made manual bidding a losing game.

The trade-off that gets missed

The standard account of unified first-price auctions is that they were fairer. Every bidder faced the same rule, floors were transparent, and the arbitrage between waterfall tiers disappeared. All true.

The less discussed effect is the transfer of information. In a second-price world the surplus and the price signal both accrued to the advertiser. In a first-price world the surplus goes to whoever shades the bid most accurately, and the price signal is visible only to whoever sees the whole auction. Both of those are the platform. The advertiser gained a simpler rule and lost both the margin and the market intelligence.

A second consequence follows for competition between platforms. When a network runs its own auction and also bids into it on behalf of advertisers using automated targets, the same entity is setting the rules, predicting the competition and reporting the result. Buyers have no independent way to check whether the bid that was placed was the lowest one that would have won. They can only observe whether the reported target was hit, which is a different question.

A framework for what a target actually controls

Since buyers cannot go back to manual bidding at scale, the useful question is what a target setting still controls. Three things, roughly.

The target sets the ceiling on the average outcome the platform will optimise towards, not the price of any impression. The platform is free to overpay on some impressions and underpay on others as long as the aggregate lands near the number.

The budget sets how hard the platform has to work. A target that is loose relative to budget gives the algorithm room to spend on marginal impressions; a tight target with a large budget forces it to compete in more crowded rooms and it will either underdeliver or drift.

The creative and event choice set what the platform is predicting. This is the lever buyers actually own, and it deserves more of the attention that used to go into bid management.

A practical test for whether a target is binding: raise it by a modest amount, say ten percent, for a week on a stable campaign and watch delivery. If spend rises proportionally and the achieved cost barely moves, the target was not the constraint and the platform was already clearing below it. If achieved cost rises to meet the new target immediately, the platform was bidding to the ceiling and you have just given it margin. Neither outcome is visible from the dashboard alone; both are visible from the experiment.

Living with the guess

The buy side will not recover price discovery. What it can do is stop treating a target as a bid and start treating it as a contract with a counterparty that has better information. That means reading achieved cost against the target rather than against last month, running the ceiling test quarterly, and asking each platform a question they will not enjoy: when your automated bidding wins an impression for my campaign, how far below my implied value did it bid, and who kept the difference.

The answer will usually be a description of the model. The absence of a number is the answer.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Alternative app stores will not lower your CPI
• Installs down 7%, revenue up 1%: Sensor Tower's State of Gaming and the end of volume UA