What happens in the auction window
The auction window opens when the SSP sends a bid request and closes when the time limit the request carries has passed. Everything that follows happens inside it: the request travels to each DSP's bidding servers, each DSP evaluates it against the line items it is running, each returns a bid or declines, the SSP collects the responses that arrived in time, applies the publisher's floor priceThe minimum bid a publisher will accept for an impression, set in its supply-side platform or ad server. and any deal rules, ranks what remains and selects a winner. The winner's creative markup is returned to the browser, where it begins to render while the rest of the page is still loading.
Four parties act in the window and none of them is a person. The publisher's ad server decided to offer the slot; the SSP describes and sells it; the DSPs value it; the exchange, which is usually the same company as the SSP, settles it. The advertiser's own ad server appears only after the window closes, when the creative calls it to serve and count. The ecosystem guide explains these roles and how the platforms are integrated, and the fundamentals hub follows one impression through the whole chain; this guide assumes both and concentrates on the auction.
Two properties of the window shape everything a buyer does. It is short, so any computation a DSP cannot finish in time is simply not made, and any bid that arrives late is discarded as if it were never sent. And it is repeated an enormous number of times a day, so a small inefficiency in how a DSP values or shades its bids compounds into a large one across a campaign. Pacing and bidding settings are the buyer's handles on this, and in Google's DSP they are covered by the DV360 bidding and pacing guide.
The bid request and what it carries
The bid request is a structured message defined by the OpenRTB specification that IAB Tech Lab publishes. It
describes the impression in nested objects, and the table lists the ones a buyer most needs to understand.
Every field is optional from the protocol's point of view; the publisher and the SSP decide what to include.
The site and source fields are also what a DSP checks against the publisher's list of authorised sellers, its
ads.txt fileA file on a publisher's site listing the companies authorised to sell its ad inventory..
| Object or field | What it carries | Why a bidder cares |
|---|---|---|
id and tmax | The request's identifier and the maximum time, set by the exchange, within which a response must arrive | A response after tmax is discarded; the DSP must budget its processing to it |
imp | The impression or impressions on offer: format (banner, video, native, audio), sizes, position, whether a secure creative is required, bidfloor and bidfloorcur | Decides whether any creative fits and what the lowest acceptable bid is |
imp.pmp | Deal information: the deal IDs this impression is eligible for and whether the auction is private | A deal-targeted line item bids only when its deal ID appears here |
site or app | The domain, page, content categories and publisher of the inventory, and the ref referrer | Inventory lists, contextual targeting, brand safety and supply chain checks |
device | User agent, IP-derived location, device type, operating system, and an advertising identifier where permitted | Geographic and device targeting; frequency capping on apps |
user | A user identifier the exchange knows, the buyer's own matched identifier, extended identifiers, and the consent string | Audience targeting and frequency capping depend on recognising a value here |
regs | Regulatory flags such as coppa and gdpr | Determines which signals may lawfully be used and whether personalised bidding is possible |
source | The SupplyChain object listing every intermediary, plus the transaction identifier | Supply path checks; matching to ads.txt and sellers.json |
at, cur, bcat, badv | The auction type declared by the exchange, accepted currencies, and the content categories and advertiser domains the publisher blocks | Shapes the bid price strategy and whether the advertiser may bid at all |
The request is smaller than the table suggests for most impressions. Privacy rules, browser restrictions and
publisher choices strip the user object of anything useful on a large share of requests, leaving the
bidder with context and device alone. A plan that assumes every impression arrives with an identifier will
reach fewer people than its forecast; the realistic assumption is that audience-based line items see a
fraction of the traffic that contextual onesChoosing where ads appear from the content of the page, app or video being viewed rather than from data about the viewer. do, and the shortfall should be planned for rather than
discovered. The guide to planning contextual targeting in DV360 shows how to build that plan from the
site and app context every request carries.
How a bidder decides
A DSP processes a request in stages, discarding as early as it can because most requests match nothing.
Eligibility. Does any active line item target this inventory, this geography, this format and this size? Is the deal ID, if any, one a line item holds? Is the advertiser's domain blocked by the publisher? Does the DSP hold a creative of the right size that the exchange has approved? A single failure ends the evaluation.
Frequency and recency. If the request carries an identifier the DSP recognises, it checks how many times this person has seen the line item's ads in the capping window and whether a remarketing audience applies. Without an identifier, caps cannot be enforced for that impression, which is why frequency reports from a DSP understate true exposure.
Pacing. Each line item has a budget and a flight, and the DSP decides whether to participate in this auction at all given how the budget is tracking against the schedule. A line item that is ahead of pace declines auctions it would otherwise enter; one that is behind enters more and may bid higher.
Brand safety and verification. Pre-bid segmentsVerification applied before the bid: a vendor's segments, checked by the DSP, decline impressions classed as unsafe, unsuitable, unviewable or invalid. from a verification vendor, inclusion and exclusion lists and sensitive category settings are applied here, before any price is considered.
Valuation. What is the impression worth? For a performance line item the DSP estimates the probability of the outcome it is optimising towards and multiplies by the value of that outcome; for a reach line item it asks whether this impression adds a new person or a useful repeat at an acceptable cost per thousand (CPMCost per 1,000 impressions.). Custom bidding, where the advertiser supplies its own valuation logic, replaces the DSP's default model at this step.
Bid price. The valuation gives a maximum. The auction type then decides how much of that maximum to bid, which is the subject of the next section. The response, carrying the price, the creative reference and the deal ID where one applies, is sent back within the time limit.
The same logic, with search queries in place of impressions, is what automated bidding in paid search does; the Google Ads bidding hub describes that system, and the parallel is a useful way to understand both. The comparison of search, feed and programmatic auctions takes the parallel further, for buyers who run all three.
First-price auctions and bid shading
For most of programmatic's history, exchanges ran second-price auctions: the highest bidder won but paid the
second-highest bid, so the rational bid was the buyer's true valuation. Over time the major exchanges moved to
first-price auctionsAn auction in which the winner pays the amount it bid., where the winner pays what it bid, and Google documents that Ad Manager now runs a
unified first-price auction for the demand that competes in it. The bid request declares which type applies
in its at field.
First price changes the problem. Bidding the full valuation now means paying the full valuation every time, even when the next bid was far lower. The response is bid shading: the DSP estimates, from its own history of wins and losses on similar impressions, the lowest price likely to win, and bids a shaded amount between that estimate and the valuation. Most DSPs apply shading automatically in first-price auctions; whether it can be switched off or tuned varies by platform and is worth checking in each DSP's documentation rather than assuming.
Shading has consequences a buyer should understand. Win rate falls, because some shaded bids lose auctions the full bid would have won; the saving is meant to outweigh the lost impressions, and on reach-driven line items that trade-off should be checked rather than trusted. Floor prices become more decisive, since a floor sets the minimum regardless of how little competition exists, and publishers set floors with that in mind. And the win price returned in the notice becomes the DSP's main source of learning, so a line item with few wins also has a poorly informed shading model, which is one reason new line items often pay more per thousand in their first days than later.
Three bidders, one impression
An impression on a recipe site is offered with a floor of £1.80 CPM and a response deadline of 120 milliseconds. Three DSPs hold matching line items. A kitchenware brand values the impression at £4.20 and its DSP shades the bid to £3.25, having won similar impressions at around £3.00 over the past week. A meal-kit service values it at £3.10 and bids £3.10 unshaded. A supermarket values it at £2.60, but its bid arrives after 130 milliseconds and is discarded. In the first-price auction the kitchenware brand wins and pays £3.25. Had the exchange run a second-price auction, the brand would have bid £4.20 and paid £3.11, a penny above the meal-kit bid. Over a campaign of 20 million such impressions, the difference between a shading model that lands at £3.25 and one that lands at £3.60 is £7,000, which is why DSPs compete on the quality of their shading as much as on their targeting.
The simulator below runs the same mechanics with figures you choose: set bids and floors for several bidders, switch between first-price and second-price rules, and watch who wins, what they pay and what shading does to the result.
Real-time bidding auction simulator
Win notice, creative serve and the impression
Winning is not the same as serving, and serving is not the same as being counted. The OpenRTB specification
defines a win notice, the nurl, that the exchange calls when a bid wins, with a macro that substitutes the
clearing price so the DSP learns what it paid. It defines a separate billing notice, the burl, that fires
when the impression becomes billable, which exchanges tie to the creative beginning to render rather than to
the auction result. The two notices exist because a won auction can still fail to produce an impression: the
page may be abandoned, the creative may fail to load, or a later ad server decision may replace the ad.
The creative itself reaches the browser either as markup in the bid response, the adm field, or by being
fetched from the win notice. For most advertisers the markup is not the creative but a tag from the
advertiser's ad server, which serves the actual asset, counts the impression under its own rules and attaches
any verification script. The industry's measurement guidelines count an impression at the point the creative
begins to render in the browser, and ad servers that follow them count later and lower than DSPs that count
wins.
Where the exchange supports it, a loss notice, the lurl, tells the DSP why a bid lost: below floor, outbid,
creative blocked, timed out. Not every exchange sends it and not every DSP exposes it, but where it is
available it is the single most useful diagnostic for a line item that bids and does not win.
The practical consequences are three. Reconcile the DSP's won or served count with the ad server's rendered count and treat the ad server as the record. Expect the gap between them to be larger on slow pages, heavy creatives and app inventory. And when the gap widens suddenly, look for a creative or tag problem before a market one.
Latency, timeouts and lost auctions
The request's tmax is the exchange's deadline for responses, and a bid that misses it is lost in a way that
never appears in a win-rate report as a loss on price. Several things eat the time. The request has to travel
from the exchange's data centre to the DSP's and back, so distance matters: a DSP bidding from a region far
from the exchange's servers starts with less of the window than a nearby one. The DSP's own processing has to
fit inside what remains: eligibility checks, audience lookups, pre-bid verification calls, custom bidding
logic and shading. Under heavy load, a DSP may throttle or sample requests to protect its response times,
which means some auctions are never entered at all.
Lost auctions come in several kinds, and distinguishing them is the beginning of fixing under-delivery. Bids that never happen because the request was filtered or sampled before evaluation; bids that happen and are rejected before competing, because the price was below the floor or the creative was not approved; bids that compete and lose to a higher price; and bids that arrive too late. Only the third is a pricing problem, and raising bids to fix the others wastes money.
What a buyer can do depends on the platform. Most DSPs report bid requests received, bids made and wins by exchange, which shows where in the funnel the losses occur, and the better ones show rejection reasons where the exchange supplies them. Pre-bid verification segments and custom bidding scripts add processing time; when a line item that uses them under-delivers against a sibling that does not, latency is a candidate cause. Creative weight affects rendering rather than the auction, but a creative that fails to render turns wins into unbilled impressions, so it belongs in the same investigation. The deal types guide covers the other way to avoid losing auctions, which is not to be in one: a preferred or guaranteed deal removes the competition and the timing risk at the cost of a fixed price.
The CPM calculator turns a line item's spend and impression count into the effective price per thousand that these mechanics produce, which is the figure to compare across exchanges and against the valuation the plan assumed.
For the buy side and sell side behind each bid request, the comparison of demand-side and supply-side platforms explains who each platform works for and what each optimises.
Common mistakes
- Raising bids to fix under-delivery before checking whether the losses are on price at all. Filtered requests, floors, unapproved creatives and timeouts do not respond to higher bids.
- Reading the DSP's win count as impressions delivered. The ad server's rendered count is lower and is the one to report.
- Assuming every request carries an identifier. Audience line items see a fraction of the traffic that contextual ones do; plan reach accordingly.
- Ignoring floors when setting bids. In a first-price auction the floor is the price on uncontested impressions; a bid just under a common floor wins nothing.
- Treating bid shading as free. It lowers the price paid and the win rate together; on reach-driven line items check that the trade is worth it.
- Judging a new line item in its first days. The shading model has not learned yet, and early prices run high.
- Adding pre-bid segments and custom logic without watching latency. Each adds processing inside a fixed window, and the auctions lost to time are invisible in the price reports.
