What is Ad auction?

Three advertisers compete for one impression. A bids A$4.00 with a predicted click rate of 1%; B bids A$2.50 at 2%; C bids A$3.00 at 1.2%. Ranked on bid alone A wins. Ranked on expected value the order is B at 0.050, A at 0.040 and C at 0.036 — B wins on the lowest bid of the three, because the platform expects to earn more from serving it.

Every ad impression on a major platform is sold by an auction, and there is one auction per impression. A page loads, a feed slot comes into view, a video reaches a break, and a process resolves in milliseconds to decide which advertiser fills it and what they pay.

The volume is the first thing to grasp: billions of these resolve daily, each conditioned on who the person is, what they are doing, what device they are using and what the advertiser has said it wants. No human process can participate at that granularity, which is why every part of the buying stack is automated and why the interesting decisions moved from the individual bid to the instructions given to a bidder.

The defining feature of modern ad auctions is that the highest bid does not automatically win.

The platform ranks competitors by what it expects to earn from serving each one, which combines the bid with the probability of the event the platform is paid on. In a cost-per-click marketplace, that is roughly:

Rank ≈ bid × predicted click-through rate × quality signals

An advertiser bidding A$2.50 with a predicted click rate of 2% has an expected value of five cents per impression. One bidding A$4.00 at 1% has four cents. The lower bid wins.

This is not a reward for virtue. The platform is maximising its own revenue per impression, and an ad more likely to be clicked is worth more to it at any given bid. The alignment between good advertising and cheap advertising is a happy consequence of that arithmetic rather than a policy.

The quality signals vary by platform and typically include the relevance of the ad to the query or audience and the quality of the landing experience. On social platforms, estimated action rates and user-feedback signals do similar work.

Search and social auctions are second-price or a close variant: the winner pays approximately what was required to beat the runner-up rather than their own bid. Programmatic display largely moved to first-price, where the winner pays their bid, and the distinction changes what a bid means on each.

Two consequences follow.

A bid is a ceiling, not a price. Setting it higher does not automatically mean paying more; it means being willing to pay more in the cases where competition demands it.

An average cost well below the bid cap is a sign of headroom rather than a sign of misconfiguration. It means the auctions being won are clearing comfortably, and it is one of the more commonly misread diagnostics in an account.

In auctions where rank includes quality, the clearing price is adjusted for it as well: an advertiser with a strong quality estimate needs a lower bid to beat the same competitor, so the effect appears both in whether they win and in what they pay.

The practical consequence of expected-value ranking is worth stating directly, because it reframes work that is usually filed under creative as work that belongs in a budget discussion.

Improving the predicted action rate does two things at once. It wins auctions that were being lost, and it lowers the clearing price of the auctions already being won. No bid change does both; raising a bid wins more and pays more.

This is why the sequence for an uncompetitive account is copy and landing page first, structure second, bids last. The first two improve the terms of every future auction; the third only improves participation.

An account with a ceiling on what a conversion can cost should lose every auction that clears above it.

This sounds obvious and is routinely forgotten, because the reporting presents lost impressions as a deficiency. A campaign losing two thirds of its eligible auctions may be behaving perfectly, declining traffic it cannot afford. Chasing those impressions by bidding above the ceiling converts a profitable account into a busy one.

The diagnostic question is not how many auctions were lost but which. Losses concentrated in segments the business wants, at prices within the ceiling, are a real competitiveness problem. Losses spread across marginal traffic at prices above the ceiling are the strategy working.

Competitors' bids are never visible. Platforms report some comparative signals — how often another advertiser appeared in the same auctions, how often they appeared above you, a share-of-voice figure — which indicate who competes for the same inventory without revealing what they paid.

Tools that claim to show a competitor's bid are inferring it from observed positions and public signals, and the inference is loose. Treating those estimates as facts leads to bidding against a number nobody actually set.

What is visible and useful is the account's own impression share and its two lost columns, which together describe the size of the opportunity and whether money or competitiveness is the constraint.

Three conclusions follow from how the auction works, and they shape most of what an account actually does.

The bid is one input among several, and it is the only one that raises costs while raising wins.

Relevance, expected response and experience are pricing levers, and they compound: improving them today makes tomorrow's auctions cheaper as well as more winnable.

And the ceiling belongs to the business rather than the auction. The auction decides what an impression costs; only the contribution a conversion produces decides what it is worth.