You start a show. There is a two second pause before the ad plays. In that pause, an invisible auction runs: dozens of advertisers bid on the chance to reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.Voir la définition complète → you, specifically, on this screen, at this moment. The winning bid is decided, the creative is assembled, and the ad loads before you notice the delay. This happens millions of times per minute across streaming and web video, and AI now runs almost every step.
This lesson traces how that works, why it makes money, and what breaks when the privacy signals go dark.
Most digital advertising is sold programmatically: bought and sold by software in real time rather than by salespeople negotiating contracts. The core mechanism is real-time bidding (RTB), an automated auction that completes in roughly 100 to 300 milliseconds.
Here is the chain:
When the ad break loads, the SSP sends a bid request describing the opportunity: content genre, device type, rough location, and whatever audience signals are allowed. DSPs evaluate it and respond with bids. Highest usable bid wins.
The Interactive Advertising Bureau maintains the open standard for this. If you want the real spec, the IAB Tech Lab OpenRTB documentation is free and public.
The auction itself is simple. The intelligence is in the bidding.
A DSP does not bid the same amount on every impressionimpressionThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète →. It predicts, in milliseconds, the probability that this specific viewer will click, subscribe, or buy, then bids accordingly. Two models do most of the work:
These feed into a bid price the advertiser is willing to pay. Getting these predictions slightly better than a competitor is the entire game, because you win the impressionsimpressionsThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète → that are underpriced relative to their true value.
The publisher wants the opposite of the advertiser: maximum revenue per available ad slot. This is yield management, the same discipline airlines use to price seats.
The key metric is CPMCPMCost Per Mille: the cost to deliver 1,000 ad impressions. A pricing and benchmarking metric for awareness campaigns where reach matters more than clicks.Voir la définition complète →, cost per millecost per milleCost Per Mille: the cost to deliver 1,000 ad impressions. A pricing and benchmarking metric for awareness campaigns where reach matters more than clicks.Voir la définition complète →, meaning cost per thousand impressionscost per thousand impressionsCost Per Mille: the cost to deliver 1,000 ad impressions. A pricing and benchmarking metric for awareness campaigns where reach matters more than clicks.Voir la définition complète →. A publisher's job is to raise
AI-driven yield tools handle several levers at once:
Header bidding orchestration. Publishers let multiple exchanges bid simultaneously rather than in a waterfall. Machine learning predicts which demand sources are likely to bid highest for a given impressionimpressionThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète → and prioritizes them, shaving latency.
Floor price optimization. A floor is the minimum a publisher will accept. Set it too high and slots go unsold; too low and you leave money on the table. Models set dynamic floors per impressionimpressionThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète → based on predicted demand. A live sports finale gets a high floor; a rerun at 3 a.m. gets a low one.
Ad load balancing. More ads mean more revenue now but more churn later. Models estimate the point where an extra ad break starts costing you the viewer.
Connected TV (CTV) means streaming delivered to a television: smart TVs, streaming sticks, game consoles. It is where premium ad budgets are moving, because it combines the trust of TV with the targeting of digital.
CTV is attractive for concrete reasons:
But CTV has quirks. Many devices do not support cookies at all, so the industry leans on device IDs and IP-based household matching. And a large share of CTV inventory is sold programmatic guaranteed: automated pipes, but fixed price and fixed audience, negotiated in advance. So "programmatic" does not always mean "auctioned."
Winning the auction is half the battle. The other half is showing the right version of the ad.
Dynamic creative optimization (DCO) builds an ad from modular parts (background, headline, product image, call to actioncall to actionA button, link, or message that prompts users to take a specific action such as sign up, buy, download, or learn more.Voir la définition complète →) and assembles the best combination for the viewer in real time. A travel advertiser might show beach footage to one household and city footage to another, pulled from the same template.
AI extends DCO in two directions:
1. Selection: contextual bandit algorithms test creative variants and shift spend toward winners while still exploring. This is more efficient than classic A/B testingA/B testingA/B testing is a controlled experiment that compares two versions of something (A and B) by splitting traffic randomly to learn which performs better on a chosen metric.Voir la définition complète → because it does not wait for a full experiment to conclude.
2. Generation: generative models now produce creative variants (copy, voiceover, localized versions) at scale. A brand can spin up dozens of market-specific cuts from one master asset. This raises real questions about brand safety and rights, so most large advertisers keep humans reviewing generated output before it airs.
Here is the logic inside a DSP, stripped to its essence:
# predicted value of showing this ad to this impression
def bid_price(pctr, pcvr, value_per_conversion, max_margin):
# expected revenue if we win and show the ad
expected_value = pctr * pcvr * value_per_conversion
# bid a fraction of expected value, leaving margin
return expected_value * (1 - max_margin)
# example: 2% click, 5% conversion, $80 per conversion
bid = bid_price(0.02, 0.05, 80, 0.30)
print(round(bid, 4)) # -> 0.056 (i.e. $0.056, or $56 CPM)The numbers are illustrative, but the structure is real: bids are expected value calculations, and better predictions of pCTR and pCVR translate directly into winning more of the profitable impressionsimpressionsThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète →.
Vérification des acquis
1. What fundamentally distinguishes programmatic advertising from traditional ad buying?
2. In the real-time bidding chain, what is the primary role of a supply-side platform (SSP)?
3. Why does a DSP bid different amounts on different impressions rather than a single flat price?
4. Select ALL correct answers about the information typically contained in a bid request sent by an SSP.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers describing the roles of AI models like pCTR and pCVR in the bidding process.
Sélectionnez toutes les réponses correctes.
For years, targeting relied on the third-party cookie: a small file that let advertisers track a user across many different websites. That era is ending. Browsers restrict cross-site tracking, mobile platforms require explicit opt-in for tracking, and privacy laws like the EU's GDPR (General Data Protection Regulation) and California's CPRA constrain what data can be collected and how.
The result is signal loss: less deterministic data about who a viewer is. This reshapes the whole stack. A few responses now dominate.
First-party data. Data a company collects from its own direct relationship with the user (logged-in accounts, subscriptions, purchase history). A streaming service knows its own subscribers; that data is durable because it does not depend on cross-site tracking.
Contextual targeting, revived. Instead of "who is this person," target "what is this content." AI content classification reads the video, audio, and transcript to place a car ad next to a road trip scene. Contextual is privacy-friendly because it needs no personal identifier at all.
Clean rooms. A data clean room is a secure environment where a publisher and an advertiser can match audiences without either side seeing the other's raw user data. Only aggregated results come out. This lets a retailer and a broadcaster find overlapping customers without exposing individual records.
Modeled and probabilistic audiences. Where deterministic IDs are missing, models estimate audience segments from available signals. Useful, but less precise, and it must be handled carefully to stay within privacy rules.
Every optimization above pushes toward more precise, more frequent, more relentless monetization of attention. That creates a counterforce.
Viewers churn when ad loads feel heavy. Regulators tighten as tracking feels invasive. Advertisers pull back when they cannot verify that a human, not a bot, saw the ad (the ongoing fight against ad fraud). The platforms that win long term optimize not just revenue per impressionimpressionThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète →, but revenue per impressionimpressionThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.Voir la définition complète → that a real, retained, consenting viewer actually sees.
That is the real frontier for 2026: doing yield management inside tightening privacy and attention constraints, using models that are good enough to price attention accurately without the cross-site surveillance that used to make it easy.