Between a reader opening a page and an advertisement appearing, a description of that opportunity travels through several independent companies. Following the route explains most of programmatic's oddities.
The publisher assembles the offer
Code on the page collects what is known about the slot and the visitor, including page address, format, device type and any consent signals present.
That package is handed to the publisher's selling platform, which decides which exchanges and buyers to approach and under what commercial terms.
What is included at this first step determines everything afterwards, because no downstream system can add information the publisher chose not to send.
Exchanges broadcast and filter
The exchange distributes the request to connected buying platforms, applying its own rules about which buyers may see which inventory.
Buyers that have expressed no interest in the publisher, the format or the geography are not contacted at all, which keeps the request volume manageable.
The filtering is invisible to the advertiser, who simply observes that certain inventory never becomes available regardless of what is bid for it.
The buying platform decides in milliseconds
On receipt, the buying platform matches the request against every active campaign, checks targeting, budget pacing and frequency, and estimates a value.
All of that must complete inside a timeout typically measured in tens of milliseconds, which constrains how much computation any decision can involve.
Complex targeting rules therefore have a cost in response time, and a platform that misses the deadline loses the auction irrespective of its bid.
Winning is not the same as rendering
The winning response returns a pointer to creative rather than the creative itself, and that pointer may lead through several further servers.
Each hop can fail. A slow creative server, a blocked domain or a reader who scrolls away all produce a won auction with nothing displayed.
This gap between auction wins and measured impressions is normal, and a large gap is usually a symptom of chain length rather than of fraud.
Reporting is reconstructed after the fact
Every party in the route logs its own version of events, and those logs are compiled at different moments with different definitions.
Discrepancies between buyer and seller counts are therefore routine, and a modest gap is expected rather than evidence of a problem.
Reconciliation focuses on whether the discrepancy is stable. A steady difference is a definitional artefact, while a widening one points at something breaking in the path.