Somewhere between a page starting to load and an ad appearing in it, an auction happened. Dozens of companies were asked whether they wanted to show something to one specific person, they each computed a price, and the highest bidder won. The whole thing finished faster than you can blink.
Programmatic advertising is often described in language that obscures rather than explains. This guide takes the mechanism apart: what an SSP and a DSP each actually optimise for, what happens in those 120 milliseconds, why the industry switched auction types and what that did to bidding strategy, and where the money goes between an advertiser's budget and a publisher's revenue.
A DSP (demand-side platform) is the buyer's console — advertisers use it to bid on individual impressions across thousands of publishers, optimising for the lowest cost per outcome. An SSP (supply-side platform) is the seller's console — publishers use it to expose inventory to every willing buyer, optimising for the highest yield per impression. They meet inside a real-time auction that resolves in roughly 100 milliseconds. Same impression, opposite incentives.
Executive Performance Asset
Download Deeptanshu Sharma's Multi-Touch GTM Attribution & Server-Side CAPI Playbook
Get immediate access to pre-built GTM server containers, first-party cookie extenders, and value attribution matrix sheets built for Series A to E companies.
1. Two Platforms, Opposite Incentives
The cleanest way to hold this in your head is that programmatic is a market, and every market has a buy side and a sell side with genuinely opposed objectives. Neither platform is neutral, and understanding what each is optimising explains most of the behaviour you will observe.
The DSP works for the advertiser
Its job is to buy the right impression as cheaply as possible. It holds your audience data, campaign rules, frequency caps and budget pacing, and it decides in real time whether a given impression is worth bidding on and how much.
Optimises for: cost per acquisition, reach against a target audience, budget delivered on pace. Examples: DV360, The Trade Desk, Amazon DSP.
The SSP works for the publisher
Its job is to sell each impression for as much as possible. It manages floor prices, decides which buyers may bid, packages inventory into deals, and enforces the publisher's rules about which advertisers are acceptable.
Optimises for: revenue per thousand impressions, fill rate, yield across demand sources. Examples: Magnite, PubMatic, Google Ad Manager.
This opposition is why so much of ad tech is about information asymmetry. The SSP would like the DSP to believe the impression is scarce and valuable. The DSP would like to know exactly how many other routes exist to the same impression and what the real floor is. Most of the innovation of the last decade — header bidding, supply path optimisation, log-level data — is one side trying to close the information gap the other side benefits from.
2. The Real-Time Bidding Lifecycle
Here is what happens between a page beginning to render and an ad appearing, in order. The entire sequence must complete inside the SSP's timeout, typically 100 to 150 milliseconds.
- Impression becomes available. A user loads a page containing an ad slot. The publisher's ad server or header bidding wrapper recognises there is inventory to sell.
- The SSP builds a bid request. It assembles an OpenRTB payload describing the slot size, page URL, device, geography, and whatever user identifiers are permitted — then broadcasts it to every connected DSP.
- Each DSP evaluates independently. It looks up whether it recognises the user, checks which campaigns are eligible, applies frequency caps and budget pacing, predicts the likelihood of the outcome it is optimising for, and computes a price.
- Bids return, or do not. Any DSP that fails to respond within the timeout is excluded. Slow infrastructure loses auctions it would have won on price.
- The SSP resolves the auction. It discards bids below the floor, applies publisher blocklists, and picks the winner.
- The creative is served and measured. The winning ad markup is returned and rendered, and impression, viewability and conversion tracking begin.
Step three is where the real engineering lives. A large DSP evaluates millions of these per second, and its bidding decision must be computed in single-digit milliseconds. This latency budget is a genuine constraint on model sophistication: a marginally better prediction that takes 40 milliseconds longer to compute is worthless if it arrives after the auction has closed.
Tired of Rising CAC & Attribution Leakage?
Work directly with Deeptanshu Sharma to audit your media strategy, funnel bottlenecks, and server-side tracking.
3. What a Bid Request Actually Contains
OpenRTB is the IAB standard that lets any DSP transact with any SSP without bespoke integration work. A simplified bid request looks like this — and reading one carefully tells you a great deal about what buyers can and cannot target.
{
"id": "8a1f2c9e-4b77-11ef-9c3d", // unique auction identifier
"imp": [{ // the impression(s) for sale
"id": "1",
"banner": { "w": 300, "h": 250 },
"bidfloor": 1.85, // minimum acceptable CPM
"bidfloorcur": "USD"
}],
"site": {
"domain": "example-publisher.com",
"page": "https://example-publisher.com/markets/article",
"cat": ["IAB13"], // content category: personal finance
"publisher": { "id": "pub-4417" }
},
"device": {
"ua": "Mozilla/5.0 ...",
"geo": { "country": "IND", "region": "MH" },
"devicetype": 1, // 1 = mobile
"os": "iOS"
},
"user": {
"id": "ssp-side-user-id",
"eids": [ ... ] // alternative identifiers, if permitted
},
"tmax": 120, // hard timeout in milliseconds
"at": 1 // auction type: 1 = first price
}
Two fields deserve attention. tmax is the timeout — the DSP's entire decision must fit inside it. And at declares the auction type, which as we will see completely changes what a rational bid looks like.
Notice also what is absent. There is no reliable cross-site identifier in a growing share of requests. As third-party cookies became unusable and mobile identifiers became opt-in, a large portion of open-internet inventory now arrives with no durable user ID at all — which is why contextual signals such as site.cat and page content have regained the importance they had lost.
4. DSP vs SSP Comparison Matrix
| Dimension | DSP (buy side) | SSP (sell side) |
|---|---|---|
| Represents | Advertisers and agencies | Publishers and app developers |
| Core objective | Lowest cost per outcome | Highest revenue per impression |
| Key controls | Audience targeting, frequency caps, pacing, bid strategy | Floor prices, buyer allowlists, deal packaging, ad quality rules |
| Primary metric | CPA, ROAS, effective reach | RPM, fill rate, yield |
| Revenue model | Percentage of media spend | Percentage of publisher revenue |
| Wants to know | The true floor and how many duplicate paths exist | The buyer's true valuation of the impression |
5. Header Bidding and the End of the Waterfall
Before header bidding, publishers sold inventory through a waterfall. Exchanges were called in a fixed sequence based on their historical average price. The first one willing to meet the floor won, and everyone below it never saw the impression.
The flaw is obvious once stated. A buyer sitting in exchange number five might have paid triple for that specific impression — perhaps it was their exact retargeting audience — but the auction was decided before they were asked. Publishers were systematically leaving money on the table, and had no visibility into how much.
Header bidding runs the auction simultaneously across all connected SSPs, in the page header, before the ad server is consulted. Every demand source bids on the same impression at the same moment, and the ad server receives the winning price as a competing line item.
What publishers traded for the yield increase
- Page latency. The auction happens before the ad renders, and the wrapper must wait for slow bidders up to the timeout. Aggressive timeouts protect page speed but exclude demand.
- Operational complexity. A wrapper configuration with a dozen bidders, each with its own timeout and floor logic, becomes a system somebody has to own.
- Cost duplication. The same impression is now offered through many paths at once, which is exactly the condition that made supply path optimisation necessary on the buy side.
Server-side header bidding moved the auction off the user's device to reduce latency, at the cost of losing some cookie-matching accuracy. Most large publishers now run a hybrid: a handful of high-value bidders client-side where match rates matter, the long tail server-side.
6. First-Price Auctions and Why Bid Shading Exists
For most of programmatic's history, auctions were second-price: the highest bidder won but paid one cent above the second-highest bid. This has an elegant property — bidding your true valuation is optimal, because what you bid influences whether you win but not what you pay.
Header bidding broke that model. With multiple simultaneous auctions feeding a final decision, the pricing logic became opaque and publishers suspected they were losing revenue in the translation. The industry moved to first-price auctions, where the winner pays exactly what they bid.
Why this changed buying strategy overnight
Under second-price rules, bidding $5.00 when the next bid was $2.00 cost you $2.01. Under first-price rules, the identical bid costs $5.00. Bidding your true valuation now means systematically overpaying by the gap between you and the runner-up. Bid shading — using historical clearing prices to bid just above where the auction is likely to clear rather than at your ceiling — stopped being clever and became mandatory.
This is why identical campaigns can produce different costs on different DSPs even when buying the same inventory: much of the difference is the quality of each platform's shading model, not the quality of its targeting.
7. Supply Path Optimisation and the Ad Tech Tax
Because header bidding exposes the same impression through many SSPs simultaneously, a DSP frequently receives five or six bid requests for one impression, each through a different intermediary charging a different fee. Bid on all of them and you compete against yourself, inflating the clearing price you eventually pay.
Supply path optimisation is the discipline of choosing one route per publisher and ignoring the rest. Buyers evaluate paths on fee transparency, win rate, latency, and whether the path is authorised in the publisher's ads.txt file — the public declaration of which sellers may legitimately resell that inventory.
| Where an advertiser's dollar goes | Typical share |
|---|---|
| DSP fee | ~10–15% |
| SSP / exchange fee | ~10–20% |
| Data, verification and brand safety vendors | ~5–10% |
| Unattributable — the "unknown delta" | varies, sometimes substantial |
| Reaches the publisher | roughly half |
Treat those figures as orders of magnitude rather than precise accounting — the split varies enormously by path and contract, and the unattributable portion is exactly the part nobody can pin down. The reason large buyers invest in log-level data and SPO is that this is one of the few levers that improves media efficiency without touching creative or targeting at all. For how these platforms compare in practice, see our breakdown of DV360 vs The Trade Desk vs PubMatic vs Taboola.
8. Deal Types: Not Everything Is an Open Auction
The open auction is only one of four ways inventory changes hands programmatically, and for premium publishers it is usually the least valuable. Knowing the four explains why two buyers can see completely different prices for the same site.
Open auction (RTB)
Anyone with a seat may bid. Maximum reach, lowest prices, least control over adjacency. This is where remnant inventory clears and where most fraud risk concentrates.
Private auction (PMP)
An invitation-only auction over a defined inventory package, usually with a higher floor. The publisher controls who competes; the buyer gets known, verified supply.
Preferred deal
A fixed price negotiated in advance, offered to one buyer before the impression reaches any auction. No guaranteed volume — the buyer gets first refusal, not a commitment.
Programmatic guaranteed
Fixed price and fixed volume, contractually committed — a traditional direct buy executed through programmatic pipes. Highest cost, complete certainty about where ads run.
The practical implication: if brand safety or adjacency genuinely matters to you, the answer is rarely a longer blocklist on the open auction. It is moving budget up this list. Blocklists are a filter applied after the fact; deal types decide what enters the pool in the first place.
9. Invalid Traffic and Where It Enters
Because the open auction accepts bids on inventory the buyer has never seen, it creates an obvious incentive to manufacture inventory that does not exist. The main patterns are worth recognising by name.
- Domain spoofing. Low-value inventory declared as a premium publisher in the bid request. The
ads.txtandsellers.jsonstandards exist specifically to make this detectable, by letting publishers publish who is authorised to sell them. - Ad stacking and pixel stuffing. Multiple ads layered in one slot, or rendered in a one-pixel frame. All bill as impressions; none is visible to a human.
- Bot traffic. Automated agents generating page views and clicks. Sophisticated variants mimic human timing closely enough to pass naive filters.
- Made-for-advertising sites. Not fraud in a strict sense, but sites built purely to arbitrage cheap traffic into ad revenue, with dense ad layouts and thin content. They absorb a meaningful share of open-auction budget.
The defences are unglamorous and effective: buy against an ads.txt-authorised seller list, use a verification vendor for pre-bid filtering, keep an inclusion list of publishers you actually want rather than an ever-growing exclusion list, and review your top delivering domains monthly. If a domain you have never heard of is consuming a large share of impressions, that is the signal.
10. What This Means If You Are Buying Media
- Ask which paths you are bidding through. If your DSP cannot tell you, you are almost certainly competing against yourself on some inventory.
- Check the publisher's ads.txt against your seller list. Buying through an unauthorised reseller is the most common way budget reaches inventory that is not what it claims to be.
- Do not read a low win rate as failure. On first-price with shading, a low win rate at a good clearing price is usually healthier than a high win rate achieved by overbidding.
- Separate viewability from delivery. An impression served is not an impression seen; measure them as different things or you will optimise toward cheap inventory nobody looks at.
- Expect DSP and publisher numbers to disagree. Different counting points, different timezones, different filtering. Reconcile once, document the expected gap, then monitor for change.
- Watch your frequency across paths, not within one. Frequency caps are enforced per DSP, so the same user bought through two platforms can see your ad far more often than either report suggests. This is one of the most common causes of unexplained creative fatigue.
- Treat CPM as an output, not a target. Chasing a low CPM reliably buys cheap, unviewable, made-for-advertising inventory. Optimise toward cost per viewable engagement or cost per outcome and let the CPM land where it lands.
- Revisit floor prices if you sell inventory. Floors set once and forgotten are the publisher-side equivalent of never testing bids — dynamic floors that respond to demand typically recover meaningful yield that static floors leave on the table.
An SSP and a DSP are the two ends of the same 100-millisecond auction, built by people with directly opposed commercial interests. Once you hold that frame, the rest of programmatic stops being jargon: header bidding is publishers refusing to be asked in sequence, first-price auctions are publishers refusing an opaque discount, bid shading is buyers adapting to that change, and supply path optimisation is buyers refusing to pay several intermediaries for one impression. If you are spending on the open internet, the highest-leverage question is not which DSP you use — it is how many paths your budget is taking to reach the same publisher, and what each of them is charging you for the privilege.