Every monetization vendor now stamps “AI” on the box, and the word hides more than it says. A floor pricing model, a bidding algorithm, and a reporting chatbot are all “AI,” and each does something different to your revenue.
This is the AI companion to our full AdTech glossary. That one defines the terms. This one covers what AI is doing to each of them: the ones it sharpens, the ones it automates, and the ones it is quietly making obsolete. Entries run A to Z, each with a short definition and then the part that matters, how AI changes it.
How to read it depending on who you are. If you run a website, weight the entries on header bidding, floors, contextual signals, and first-party data. If you run a mobile app or a game, start with mediation, in-app bidding, and the format entries. Both share one thread: the market moves faster than a human team can react to it, and closing that gap is what most of the AI below is actually for. The applied version is in how to increase ad revenue with AI.
A
Ad exchange
The marketplace where impressions are auctioned to buyers in real time. AI now sits on top of the exchange rather than inside the definition: models decide which path and floor to send each impression through, so the exchange clears the auction while a layer above it chooses how the auction is set up. For a publisher, the exchange stopped being the smart part of the stack.
Ad network
An intermediary that aggregates inventory and resells it. AI changed what a network is: platforms like AppLovin run the ranking and auction through their own models, so choosing and ordering networks by hand is work a mediation model now does per impression. Picking a network on last quarter’s average is the habit AI retires, which is part of why publishers keep a shortlist of AdMob alternatives rather than one default.
Ad server
The technology that decides which ad to serve and reports on it. The server still serves, but the decisions that move money, floors, tag priority, demand allocation, are shifting to models that read live signals and act. An AI layer plugs into the server through an API and operates those levers continuously instead of a human editing them on a schedule.
Agentic ad operations
Ad ops run by AI agents that perceive a situation, decide, and execute the change, past surfacing a dashboard for a human to act on. The action step is the whole point: an agentic system switches the tag itself. UndrAds has written on how AI agents are changing ad operations.
AI agent
Software that watches signals, decides toward a goal, and acts with limited human input. In monetization it monitors eCPM and fill, catches a drop, and reallocates demand or moves a floor without waiting for someone to notice.
AI floor pricing
Using machine learning to reprice floors continuously against live demand. This is the clearest case of AI making an old task obsolete: a manual floor is a snapshot that ages, and demand shifts by geo, hour, and format faster than anyone resets it by hand. The mechanism and the revenue math are in the AI floor price optimization guide and static floors versus AI floor pricing.
Anomaly detection
An ML technique that flags abnormal patterns automatically, like an eCPM collapse in one geo or a fill rate outside its normal band. It is the detection half of catching drops early. Flagging without a fast fix still loses the money, which is why detection pairs with ad refresh and floor optimization.
ARPDAU
Average revenue per daily active user. AI turns it from a number you read the next morning into one you forecast: models project ARPDAU by cohort and match ad pacing and format to predicted user value, so the metric drives decisions before the day is over rather than after.
App Tracking Transparency (ATT)
Apple’s rule, live since 2021, requiring apps to ask before accessing the cross-app tracking identifier. Opt-in stayed low, and the signal ATT removed is exactly what pushed spend toward AI: aggregated modeling and on-device inference now stand in for the user-level data that went dark.
Attention metrics
Measurement of whether an ad was actually attended to and for how long. AI made attention usable: models predict likely attention from engagement signals before the impression serves, so buyers can price on predicted attention instead of raw viewability.
Automated tag switching
Changing or reordering ad tags automatically in response to performance. This is the execution layer that closes the gap between spotting a problem and fixing it, and it is a task AI removes from human hands entirely. UndrAds runs it through ad tag automation, and the broader case is in cutting ad ops workload with automation.
Autonomous ad operations
Monitoring and acting on monetization in real time through automation, so the delay between a problem and the fix drops to minutes. It amplifies an existing team rather than replacing one, and it assumes there is a setup to optimize. The full explanation is in what autonomous ad operations means and the AI ad operations page.
B
Bid density
How many bids compete per impression. AI does not create demand, but it reads density and reacts, loosening a floor when competition thins at 3 AM in a region, and forecasting the thin windows so the reaction is set up before revenue leaks through them.
Bid shading
A buy-side algorithm that lowers a bid in a first-price auction to just above the expected clearing price. Shading is itself a model, so the publisher’s answer is also a model: AI floors that adapt to the shaded clearing price keep buyers from quietly walking your inventory down.
C
Consent Management Platform (CMP)
The tool that collects and passes user consent for data use. AI’s role here is narrow but real: models can optimize prompt timing and wording to lift opt-in rates, and consent rate feeds directly into how much signal, and therefore demand, your impressions carry.
Contextual signals
Information about the page or app itself, used to target without personal identifiers. This is where AI changed the most. Natural language models read content semantically instead of matching keyword lists, so contextual targeting went from crude to competitive with ID-based targeting as tracking eroded. UndrAds runs it through contextual targeting. For web publishers, page context is now a first-class signal.
CPM
What an advertiser pays per 1,000 impressions. AI does not change the definition, it changes how you move the number: models lift CPM by timing demand, tuning floors, and enriching signals automatically, so the lever is continuous rather than a weekly manual pass.
Curation
Packaging inventory and data into ready-made deals on the sell side. AI increasingly assembles these, matching inventory to buyer intent and pricing the package, which is turning curation from a manual sales motion into a modeled one.
D
Demand-Side Platform (DSP)
The software advertisers use to buy programmatically. The DSP is now mostly an AI: predictive bidding, lookalike modeling, and ROAS forecasting decide what to pay for each impression. When buyers talk about their DSP, they are usually talking about its models.
Dynamic Creative Optimization (DCO)
Assembling or selecting creative in real time to fit the user or context. Generative AI moved this a step further: models now produce the variants, not only pick among pre-built ones, so creative can be generated per audience rather than designed in advance.
E
eCPM
Your actual earnings per 1,000 impressions across formats and deals. AI defends it and forecasts it: it catches the overnight drops that quietly erode blended eCPM, and projects it by cohort so pricing gets ahead of demand. Lifting it is the goal behind increasing CPMs and fill rate.
F
Fill rate
The share of ad requests that return an ad. Fill and eCPM trade against each other, and AI manages that trade in real time, moving floors to protect blended revenue instead of leaving a static setting to over- or under-price your inventory all day.
First-party data
Data you collect directly from your own audience. AI is what makes it valuable at scale: models turn raw first-party signals into predictions like LTV, churn, and lookalike seeds. Without a model on top, first-party data is just a table. For web publishers, authentication rate is the lever that grows this asset.
Floor price
The minimum CPM you accept before an impression goes unsold. The floor is the most direct lever a publisher controls, and it is the one AI most clearly takes over: repriced per impression against live demand, a hand-set floor becomes the thing you no longer maintain. For app and game studios, this governs how much rewarded and interstitial inventory clears well during demand swings.
G
Generative AI
Models that produce content, including ad copy, images, and creative variants, and that classify page content at scale. Its two practical adtech jobs are creative generation on the buy side and content and brand-safety classification on the sell side. The wider set of shifts it is driving is in demystifying adtech trends.
Google Ad Manager (GAM)
Google’s combined ad server and SSP, the backbone of most publisher stacks. It is where floors, tag priorities, and demand allocation live, which makes it the natural place an AI layer plugs in through the API to operate those levers in real time. GAM stays the server, the model becomes the operator. Publishers weighing options can review the GAM alternatives.
Google AdX
Google’s premium exchange, reached through GAM or an MCM partner. AI’s role is allocation: deciding when to lean on AdX demand versus other paths for a given impression. UndrAds connects publishers to Google AdX.
H
Header bidding
A technique where a publisher offers inventory to multiple demand sources at once, before the ad server call, so buyers compete in one auction. Moving from a waterfall to header bidding typically lifts revenue 20 to 40 percent. AI does not replace it, it decides how much of that lift you keep: header bidding creates the competition, and the models tuning your floors and reactions determine how much of it clears at the right price. The mechanics are in what header bidding is, UndrAds runs it as managed header bidding, and where it is heading is in the future of header bidding.
I
Identity resolution
Linking scattered signals to a persistent user or household ID. As deterministic IDs eroded, this became a modeling problem: AI builds probabilistic graphs and infers likely identity or intent from behavior, which is how addressability survives in patches without cookies.
In-app bidding
The mobile equivalent of header bidding: all demand competes in one real-time auction per impression. AI sits on top the same way it does on the web, catching the swings the auction alone smooths over, and the mediation ranking underneath is ML-driven. The comparison is in in-app bidding versus waterfall, and it runs through in-app mediation. For studios, moving your highest-volume placements to bidding is usually the biggest single yield gain available.
Inference
Running a trained model on live data to get a prediction, as opposed to the training that produced it. A floor model doing inference is pricing the impression in front of it right now, using what it learned earlier. It is the moment an “AI” claim turns into an actual decision.
Interstitial ads
Full-screen ads at natural breaks. AI’s contribution is frequency: models pace interstitials per user against a predicted churn threshold, so you push revenue up to the point of tolerance without crossing it. UndrAds supports interstitial ads as part of the format mix.
L
Large Language Model (LLM)
A model trained on large volumes of text that classifies, summarizes, and generates language. In adtech, LLMs drive content classification for contextual targeting and brand safety, and they compress the reporting side of ad ops by turning raw data into readable analysis.
Latency
The delay ad calls add to load, and the time an auction takes to resolve. AI trims it by tuning timeouts and demand paths and by predicting which demand is worth waiting for, so depth and speed stop being a fixed tradeoff.
Lookalike modeling
An ML method that finds new users resembling a seed audience, mostly a buy-side tool. It matters to publishers indirectly: sharper buyer targeting means stronger bids on inventory that fits a model, which is one reason signal-rich impressions clear higher.
M
Machine learning
Algorithms that learn patterns from data to predict or decide without hand-written rules. It splits into supervised (labeled data), unsupervised (structure in unlabeled data), and reinforcement (learning from reward). Almost every adtech “AI” claim reduces to one of these doing a specific job, so the useful question is which job, not whether the label fits. The applied version for revenue teams is in how to increase ad revenue with AI.
Mediation
The layer that manages multiple networks for an app and routes each impression to the best-paying source. AI replaced the old static waterfall order here: a model ranks sources per impression in real time, which is why mediation quality now sets the ceiling on app revenue. UndrAds sits on top of your existing stack through in-app mediation. For studios, MAX, LevelPlay, or AdMob form the base, and the question is what sits above to catch the swings mediation smooths over.
Model drift and retraining
Drift is the decay of a model’s accuracy as the market moves away from its training conditions. Retraining refreshes it on recent data. A floor or bidding model that is never retrained falls out of step with the market it reads, which is why AI is a maintenance commitment and not a one-time install. Judging a vendor means asking how often their models retrain.
N
Native ads
Ads that match the form of the content around them. AI helps on the margin: generative and contextual models match native creative to the surrounding content, which holds engagement better than a mismatched unit. UndrAds supports native and other in-app ad formats.
Natural Language Processing (NLP)
The branch of machine learning focused on understanding and generating language. In adtech it powers contextual classification, sentiment and brand-safety analysis of pages, and the semantic reading behind modern contextual targeting.
O
Offerwall
A placement listing reward offers users complete for in-app currency. AI personalizes it: models decide which offers to show which users to lift completion, so the wall adapts to intent instead of showing everyone the same list. UndrAds runs offerwall monetization for reward-heavy apps.
Open Bidding
Google’s server-side unified auction inside Ad Manager, where partners bid on the server rather than the browser. It cuts client-side latency at some cost to transparency. AI’s role is upstream of it, in the floors and allocation that decide what the auction competes over.
P
Prebid
The open-source framework most header bidding runs on, across web, server, and app. Being open-source is why config quality varies so much, and it is where AI earns its keep: models tune dynamic floors and timeouts on the wrapper so the setup performs closer to its ceiling.
Predictive bidding
Using historical auction data to forecast which sources win, when, and at what price, and acting ahead of the shift. It is the difference between reacting to a drop and pricing for one before it lands, and it is the same predictive-analytics idea applied to eCPM, fill, and churn.
Private Marketplace (PMP)
An invite-only programmatic arrangement offering select buyers specific inventory at negotiated terms. AI supports it by matching inventory to the right buyers and helping price the deal, giving publishers more control than the open auction with less manual negotiation.
Privacy Sandbox
Google’s initiative to replace third-party cookies with privacy APIs like Topics and Protected Audience. After low adoption, Google shut down most of the Privacy Sandbox APIs in October 2025. What filled the gap was not those APIs but AI-driven contextual and modeling approaches, so treat the Sandbox as wound down and the AI path as the durable one.
Programmatic advertising
The automated buying and selling of inventory through auctions, spanning the open auction, private marketplaces, and guaranteed deals. AI is now the decision layer running across the whole pipeline, pricing, bidding, and allocation, which is why nearly every term in this glossary describes a place a model touches.
R
Real-Time Bidding (RTB)
Per-impression auctions that resolve in milliseconds as a page or screen loads. RTB opened the window, AI is what acts inside it: the bid, price, and allocation decisions all come from models fast enough to fit the millisecond budget.
Real-time optimization
Adjusting floors, tag priority, and demand allocation continuously as signals change, instead of in periodic manual reviews. The value scales with how fast demand moves in your inventory, which for high-session apps and global sites is very fast, and it is only feasible because a model, not a person, is making the adjustments.
Reaction gap
The stretch between a performance drop starting and someone fixing it. eCPM can fall for hours overnight, in a region where your team is asleep, before anyone notices, and that revenue does not come back. Closing this gap is the specific problem AI ad ops exists to solve, and the human-versus-automated math is in human ad ops versus AI ad ops.
Reinforcement learning
Machine learning where an agent learns by trying actions and receiving rewards, suited to sequential decisions like bidding and allocation where each choice shapes the next. It is the technique behind systems that keep improving their allocation rather than following fixed rules.
Rewarded ads
Opt-in ads a user watches for a reward, with high completion and strong eCPM. AI tunes delivery: it paces placement density and timing per user and personalizes the reward economy to lift completion without inflation. UndrAds supports rewarded ads, with more in the role of rewarded ads and the benchmarks in how much games make per ad.
RPM
Revenue per 1,000 pageviews or sessions, rolling up every ad on a page. AI lifts it by optimizing the whole page or session at once, floors, format mix, and demand allocation together, rather than tuning one placement in isolation. Web publishers manage to RPM while apps lean on eCPM and ARPDAU.
S
Signal loss
The shrinking availability of tracking identifiers as privacy rules and browser defaults tighten. AI is the industry’s answer to it: contextual modeling, aggregated attribution, and predictive segments replace the deterministic signals that went away, which is why the modeling entries in this glossary keep gaining weight.
SKAdNetwork
Apple’s privacy-preserving install attribution framework, now evolving into AdAttributionKit. It reports conversions in aggregate and with deliberate delay, and that sparsity pushed UA optimization toward ML models that can work with less data, since the old user-level feedback loop is gone.
Supply-Side Platform (SSP)
The software publishers use to sell inventory programmatically. SSPs now run ML for traffic shaping, floor optimization, and supply path decisions, so the optimization that used to be a manual account-management job increasingly happens inside the platform’s models.
Supply Path Optimization (SPO)
Buyers cutting redundant routes to the same inventory in favor of cleaner, direct paths. AI scores those paths on both sides, helping buyers choose and helping publishers present the efficient path, so being on the short route to demand is increasingly a modeled decision rather than a relationship one.
T
Third-party cookies
Cross-site tracking cookies, historically the backbone of addressability. After Google’s reversal, they remain on by default in Chrome in 2026, while Safari, Firefox, and Brave still block them, leaving roughly 17 to 20 percent of traffic cookieless regardless. AI-driven contextual and first-party modeling is the hedge that works across all of them, which is why it stayed the plan even after cookies survived.
Training data
The historical dataset a model learns from, and the ceiling on what it can do. A floor model trained on three weeks of one app’s auctions knows that app and little else. When a vendor claims a model works, the first question is what it was trained on, since coverage and quality decide the output more than the algorithm does.
V
Viewability
Whether an ad met a measurable “seen” standard, commonly half its pixels in view for one second on display and two on video. AI moved the useful version pre-bid: models predict viewability before the impression serves, so buyers can price on predicted view time instead of finding out after the fact.
W
Waterfall
The older model that calls demand sources one at a time, ranked by historical eCPM, until one fills. It is the term AI most clearly makes obsolete: because sources bid in sequence rather than together, a buyer willing to pay more gets skipped for sitting lower in the order, and ML-ranked bidding recovers exactly that gap. The comparison is in in-app bidding versus waterfall.
Yield optimization
Maximizing revenue per impression or session across floors, format mix, demand allocation, and reaction speed at once. AI is now the engine that runs it, working several levers in real time, which is how reaction speed itself became a yield factor rather than a support function.
Z
Zero-party data
Data a user intentionally shares, like stated preferences or survey answers, distinct from behavior you observe. AI turns that declared input into predictions and segments, so the cleanest consent signal you have becomes usable at scale rather than sitting as static profile fields.
Frequently asked questions
How is this different from your main AdTech glossary?
The main glossary defines the terms in plain language with a publisher takeaway for each. This one is the AI layer on top: for every term, what AI is doing to it, whether it sharpens the term, automates it, or makes the old way of doing it obsolete. Read that one to learn what header bidding is, read this one for how AI changes how you run it.
Is “AI” in an adtech product just a marketing label?
Sometimes, which is why the test is mechanism, not the word. Ask what the model does, floors, bidding, creative, classification, what it was trained on, and whether it acts or only suggests. A system that flags a drop is doing less than one that fixes it, even when both are sold as AI.
Which ad ops jobs is AI actually making redundant?
The repetitive, time-sensitive ones: resetting static floors, maintaining a manual waterfall order, switching tags by hand, and watching dashboards every few hours for a drop. It does not remove the team. It removes the delay between a problem and the fix, and frees the people for the judgment calls a model should not make, as the human versus AI comparison lays out.
With third-party cookies still alive, does AI-driven contextual still matter?
Yes. Chrome kept cookies, but Safari, Firefox, and Brave block them, so a real share of traffic is cookieless no matter what Chrome does. Language models made contextual targeting semantic rather than keyword-crude, which is why it holds up as a durable signal alongside first-party data, and consent duties under GDPR and CPRA did not change when Google reversed course.
Where does an AI optimization layer actually plug in?
For most publishers, at Google Ad Manager, through API access rather than a new SDK or app change. That is where floors, tag priority, and demand allocation are set, so it is where an autonomous layer can read signals and act, with a lightweight integration that keeps your release cycle clean.
See where the reaction gap is costing you
Most of the revenue lost in a programmatic setup does not come from a bad floor or a weak network. It comes from the hours between a drop and the fix, the one thing no amount of manual effort fully closes. UndrAds runs a revenue-leak audit on one of your properties, shows you where that delay is costing money, and lets you test the fix for ten days against your own numbers, with no SDK and no app changes. Book an audit and start with one site or app.
