Almost every advertisement on a commercial web page was bought in an auction that ran while the page was loading. The mechanics of that auction determine what gets shown far more than the size of any single bid.
The auction is triggered by the page, not the advertiser
When a page loads, the publisher's ad server assembles a description of the available slot. That description covers the site, the position on the page, the format and whatever is known about the visitor.
This bid request is sent out to buying platforms, each of which decides within a few dozen milliseconds whether any of its campaigns wants the impression. Silence is treated as a refusal.
The advertiser is not making a decision in that moment. It set rules earlier, and the buying platform is applying them at a speed no human process could match.
Bids are ranked by expected value, not by price
The number a buyer submits is a maximum, and it is rarely what determines the ranking. Platforms multiply the bid by an estimated probability that the ad produces the outcome the advertiser wants.
A lower bid attached to an ad the system expects to perform can outrank a higher bid attached to one it expects to be ignored. The seller earns more from the first over time.
This is why doubling a bid often changes very little. The bid is one term in a product, and the other term is a prediction the advertiser does not control.
Quality estimates come from history
Those predictions are built from past behaviour, including how often similar users clicked, how quickly they left afterwards, and whether the destination page loaded properly.
A new advertiser has no such history, so the system substitutes an average and updates it rapidly. Early performance therefore carries disproportionate weight in what the campaign can buy later.
The practical effect is that campaigns build or lose momentum. A weak start raises the price of every subsequent impression, which makes recovery harder than the original launch.
Floors and direct deals sit above the auction
Publishers set reserve prices below which they refuse to sell, and they also reserve inventory for deals agreed in advance at fixed rates.
The open auction only sees what is left after those commitments are satisfied. A bidder can win the auction outright and still never appear, because the slot was already promised.
This layering explains why identical bids produce wildly different results across sites. The auction is the last step, not the whole market.
Latency decides who gets to compete
Every participant is working against a timeout measured in milliseconds. A buying platform that responds too slowly is simply excluded, however much it was prepared to pay.
Publishers set that timeout as a trade between revenue and page speed, since a longer wait means more bidders but a slower experience for the reader.
Some of the most valuable optimisation in the industry is therefore engineering rather than marketing, spent on shaving response times rather than adjusting creative.