Part II · Chapter 6

Targeting and Audiences

Status: reviewed · Prerequisites: What AdTech Is and Why It Exists, AdTech Ecosystem Participants, Core Advertising Metrics, AdTech Economics and Business Models, How an Ad Campaign Is Structured

Targeting and Audiences

The Core Idea: First the Eligible Set, Then the Ranking

Targeting1 does not answer the question “who is certain to buy,” but an earlier question: which advertising opportunities the system must reject, admit, or consider preferable. From the entire stream of ad opportunities2, the advertiser3 first forms an eligible set, then the delivery system ranks it for the chosen outcome.

business intent
  → hard eligibility constraints
  → audience/context signals
  → eligible opportunity set
  → optimization ranks eligible opportunities
  → delivery outcomes become new signals

This chain is the chapter’s main mental model. It separates eligibility rules from relevance hypotheses and from optimization. Without this distinction, it is easy to mistake the Audience field in a vendor UI for a strict filter, even though a particular product may use it only as a model hint.

Why Targeting Matters: From the Whole Set of Opportunities to the Eligible Set

A publisher4 creates a set of opportunities to show ads: in different countries, apps, browsers, contexts, and moments. Buying the entire set is irrational. A subscription app that is available only in Germany and only on supported mobile OSs should not pay for a desktop opportunity in France. But within Germany, it is impossible to know in advance which eligible user will subscribe. The configuration therefore has three semantic categories:

Semantics Meaning Example
Control / constraint The opportunity must pass the rule; otherwise delivery is prohibited geo = DE, device = mobile, supported OS
Signal / hint The attribute helps the model prioritize, but does not always restrict the set Interest fitness, a seed5 (a sample of valuable subscribers from which the model looks for similar people)
Audience exclusion / suppression6 A negative rule filters out the matching entity Do not show ads to current paid subscribers

The word targeting in a product can cover all three rows. The semantics define the setting’s contract, not its position on the screen. It is especially important to determine whether the product can go beyond the positive input and which exclusions remain strict.

A set model for learning, not a production bidder formula:

eligible(o) = geo(o)
           ∧ device(o)
           ∧ context_or_placement(o)
           ∧ audience_control(o)
           ∧ ¬suppression(o)

delivery_score(o) = model(features(o), targeting_signals, optimization_event)

eligible(o) is a boolean predicate for opportunity o. delivery_score(o) is meaningful only after the opportunity passes the mandatory rules: it sets relative priority within the eligible set. The optimization target7 indicates which outcome guides the ranking, for example paid_subscription; targeting itself does not guarantee this outcome.

Where exactly the product stores the rule—at the campaign8, ad group, or ad set9 level—depends on the platform and campaign type. To carry the mental model across vendors, it is more useful to classify each field as a constraint, signal, or exclusion than to memorize a screen.

The data flow looks like this:

advertiser rules and customer events ─────┐
publisher/platform opportunity + context ├─→ normalize/match
optional external data provider ──────────┘       ↓
                                           eligibility
                                               ↓
                                             ranking
                                               ↓
                                      delivery event → outcome
                                               └─→ feedback

The publisher or platform describes the opportunity; the advertiser supplies controls, seeds, and exclusions; an external data provider10 may add an audience segment11; and the DSP12 or closed platform performs matching, eligibility checks, and ranking. Outcomes return later and become inputs to subsequent decisions.

Targeting does not change the pricing model that governs the money flow. It changes the composition of available inventory13 and the distribution of media spend14, but it does not determine the bid15, budget16, or pacing17. External provider data may carry a separate fee; bidding and spending mechanics are covered in Chapter 7.

Available Attributes: Geography, Device, OS/Browser, and Demographics

The simplest targeting dimensions describe the opportunity’s environment or probable properties of an entity. Their presence in a request does not make them error-free facts.

Dimension What it describes Possible sources Typical uncertainty Possible semantics
Geography Likely presence, regular presence, or interest in a location IP, device settings, account data, search/content behavior VPN, travel, stale or inferred location Control or a product-specific broad option
Device Device class and sometimes family/model Request/device metadata, app/platform Coarsening, spoofing, unknown Often a criterion/control
OS/browser Software environment: family and version when available User agent, app SDK, platform metadata Reduced version, missing field, unsupported criterion Channel- and product-specific
Demographics Profile category Declared, partner-supplied, or inferred data unknown, misclassification, stale inference Control or suggestion

Geography

A geography criterion may specify a country, region, city, postal area, or radius around a point. But the value DE does not yet prove a person’s physical location. A product may interpret it as likely presence, regular presence, or interest in Germany. The source almost always provides a best-effort inference, not a certified coordinate.

Google Ads provides a product-specific example of this distinction: advanced location options distinguish between Presence or Interest and Presence, and the documentation explicitly warns that location is inferred from several signals and is not 100% accurate. This is not a universal industry default; in every product, the exact semantics of positive and negative geo rules must be checked.

Device, OS, and Browser

Mobile describes a device class, Android an OS family, Android 15 an OS version, and Chrome a browser family. These axes are related but not interchangeable. An opportunity from an Android tablet may pass the OS rule and fail the phone only device-class rule. In-app inventory may supply an OS but have no useful browser attribute at all.

Granularity depends on the channel and platform. Some platforms expose the device model, while others expose only mobile/desktop/tablet; explicit browser targeting may be unavailable even though the browser remains a model feature or reporting dimension. Reduced version in the table is a consequence of UA reduction: modern browsers truncate the OS and browser version in the User-Agent, so the exact version is often unavailable. Unknown and coarsened values are normal, so the advertiser must find out whether the product exposes unknown as a separate category and explicitly decide whether to include or exclude it, rather than relying on the platform’s silent behavior.

Demographics

Demographics are categories such as age range, gender, parental status, or household income, if supported by the product and policy. Membership may be:

  • declared — the user provided the value;
  • observed/supplied — the value came from a permitted account or partner source;
  • inferred/modeled — the platform estimated the category from behavior and other features.

Consequently, age=25–34 is a classification signal with a source, confidence, and recency, not a verified passport fact. Unknown cannot be treated as a rare error: by excluding the unknown population, an advertiser may substantially reduce the eligible set. Available categories, the use of sensitive characteristics, and the handling of unknown values depend on the country, channel, and policy; legal and consent mechanisms are covered later in the course.

For a subscription app, hard controls might be DE + mobile + supported OS. Demographics should be added only after answering two questions: is this a strict boundary or a suggestion, and what will happen to Unknown?

A User Hypothesis and the Ad Context Are Different Types of Data

Interests and behavior describe a presumed user or their history. Context, keyword, and placement18 describe the current media opportunity or a way to find it. These attributes may correlate, but they answer different questions.

Input Main question Outdoor store example
Interest What is the entity presumed to have a lasting affinity for? interest: hiking
Behavior What named action or pattern was observed, and when? viewed_boots within 14d
Contextual targeting19 What is the current page, app screen, or video about? An article about winter hiking
Search keyword What query/intent does the product match the ad to? winter boots in search
Context keyword What term or semantic concept describes the content? winter boots in the review text
Placement In which explicitly selected property/channel/site/app/video/slot is the ad eligible to appear? outdoor-news app/article_feed

Interest is usually a relatively persistent hypothesis about a preference or affinity. It may be declared, inferred from content consumption, or supplied by a partner. An interest in hiking is not the same as a current intent to buy boots: source, recency, and confidence matter as much as the label.

Behavior is built from past actions or patterns: viewing a product page, searching, adding to cart, installing, making frequent purchases. A useful behavioral rule always names the event and observation window20: viewed_boots within 14 days is more informative than active shopper. An observed action does not guarantee a future purchase and does not by itself define a re-engagement tactic.

Contextual targeting selects opportunities based on the current media context: topic, semantics, page/app/video category, and related signals. It can work without claiming that the platform knows the user’s persistent cross-site identity: classifying the environment in which the opportunity arose is enough. This distinguishes reading an article about hiking now from the audience hypothesis interested in hiking in general.

Keyword targeting has two common semantics. In Search, a keyword is associated with a user query and presumed intent according to the product’s match rules. In display/video, a keyword may describe content and aid contextual selection. In neither case does a keyword necessarily imply a universal exact string match: stemming, semantic expansion, negative rules, and combinations of criteria depend on the product contract.

Placement is a more explicit answer to “where”: a specific publisher property, app, site, channel, video, section, or slot-level scope. A contextual system may dynamically find thousands of hiking pages; a placement rule may allow only the named outdoor-news app. Formats, ad units, and creative21 compatibility are covered in Chapter 8.

Placement is an overloaded product term. On the publisher side, it is a specific ad slot or a rule for ad appearance; on the buy side, the same word may refer to a selectable media scope—an entire site, app, channel, video, section, or individual slot. When reading a configuration, clarify the level rather than treating them as a single entity.

A minimal opportunity description could look like this; this is illustrative JSON, not a vendor API:

{
  "geo": "DE",
  "device": "mobile",
  "os": "Android",
  "context_topic": "outdoor/hiking",
  "placement": "publisher_app:article_feed"
}

Boolean semantics cannot be inferred from fields being adjacent. One product will apply topic OR placement, another topic AND placement, and a third will use topic as a signal within selected placements. The product contract must explicitly define composition and precedence.

How Data Becomes an Audience Segment

An audience segment is not necessarily a list of known people. It is a set of addressable22 entities—ones the platform can recognize by an available identifier and deliver an impression to—for which the following are defined:

  1. Entity: user, account, device, browser, household, or modeled profile.
  2. Membership rule: what condition includes and excludes the entity.
  3. Provenance: who obtained the source data and how.
  4. Observation window: the interval over which events are considered.
  5. Activation contract23: how the platform matches membership to its addressable entities.

Here, the observation window defines which events, and over what period, determine membership, for example, “trial started at least 7 days ago, and no payment since.” This is not an attribution window: the segment determines whom to consider for delivery, while attribution24 later allocates credit for a conversion25.

First-Party Audiences

A first-party audience26 is built from data obtained by an advertiser or publisher through a direct relationship: CRM records, customer lists, site/app events, subscriptions, and purchases. For a subscription app, the pipeline might be:

CRM + app events + subscription status
  → normalize + deduplicate
  → rule: trial_started ≥ 7d ago AND no paid_subscription since trial_started
  → platform match (onboarding)
  → addressable audience

A CDP27 or an in-house data pipeline may calculate membership, but an uploaded record is not yet the same as an activatable entity. The platform must perform a match—compare the advertiser’s identifiers (hashed email, device ID, its own account ID) with the identifiers the platform itself knows; vendors often call the upload and match step onboarding. Some records will remain unmatched, may be updated after a delay, or may be unavailable in a particular channel/region. First-party describes provenance, not automatic consent, a legal basis, exclusive ownership, or a 100% match.

It is useful here to distinguish first-party data from first-party identity. First-party data is data obtained through a direct relationship with the user. First-party identity is an identifier that the party itself issues and controls: a login, its own account ID. The advertiser has both, but the impression happens on a platform with its own identity, so the advertiser’s first-party audience is activated only after a match with the identity of the platform where the ad is shown.

Google Customer Match is a vendor example: an advertiser supplies its own online/offline customer data, and Google matches it for activation; certain activation options on partner inventory are restricted in the EEA, UK, and Switzerland, while owned-and-operated properties are supported under their own terms. The basic concept of a first-party audience has not changed, but availability changed due to product policy.

Third-Party Audiences

A third-party audience28 is supplied by an external provider that is not a party to the advertiser’s direct relationship with the entity. The classic container for such segments is a DMP29: the provider or buyer stores and segments the data there and then activates it in a DSP; the buyer usually pays for third-party segments as a data fee on top of the media cost. A typical flow:

data originators
  → provider aggregation / classification / modeling
  → labeled segment
  → DSP/platform match
  → activation

The label likely subscription buyers does not disclose the segment’s composition. At a minimum, the following are needed before use:

  • provenance and collection method;
  • observation period and recency;
  • membership criteria;
  • the proportions of observed and modeled members;
  • permissions and permitted use cases;
  • activation coverage and match boundary.

The IAB Tech Lab Audience Taxonomy provides a common nomenclature for labels such as demographic, interest-based, and purchase-intent because vendor taxonomies have historically been incompatible. The Data Transparency Standard adds disclosures about provenance, recency, segmentation, and modeling. Neither the taxonomy nor a Data Label is a quality score: identical labels do not guarantee the same set, accuracy, or performance.

For any segment—first-party or third-party—activation runs into addressability: the platform delivers an impression only to entities it can recognize by an identifier available to it. Which identifiers are available depends on the environment: in browsers and in mobile OSs availability is restricted in different ways and has changed in recent years, so the same segment may be activatable on one platform and nearly empty on another. The mechanics of identifiers and matching are covered in Chapter 24; the consequences of addressability limits for targeting and measurement30 are covered in Chapters 40–42.

Criterion First-party audience Third-party audience
Who collects the data The advertiser or publisher itself An external provider, often aggregating several sources
Relationship with the entity Direct: customer, user, subscriber Indirect or absent
Visibility of composition Membership rule and events are fully known The label and the provider’s disclosure are visible; the composition only partially
Typical quality questions Completeness of events, instrumentation, freshness Share of modeled membership, recency, permissions
Activation limits Match with the platform’s identity Match, availability of the segment on a given platform/region, a separate data fee

The party number does not determine quality: sparse or poorly instrumented first-party events produce a weak segment, while a provider may have a relevant source that the advertiser lacks. The comparison should focus on the rule, provenance, freshness, and match.

Retargeting and Lookalikes: Past Contact Versus New Prospects

Retargeting31 addresses entities again after they performed a named interaction within a specified window; some vendors call the same tactic remarketing. It is re-engagement—bringing back users who are already known, not finding new ones. For a subscription app, this might be the segment trial_started ≥ 7d ago AND no paid_subscription since trial_started; those entities are shown a message about premium features. Recency and the event definition are part of the tactic: a visitor to the pricing page yesterday and a visitor from a year ago are not the same audience.

past interaction + window
  → retargeting membership
  → match at new opportunity
  → re-engagement

Retargeting determines whom to show an ad to again. It neither assigns attribution credit nor proves incrementality: a high conversion rate32 may reflect pre-existing intent.

A lookalike audience33 solves the opposite problem—prospecting34: it finds new prospects, who have had no interaction with the advertiser yet, from a seed audience. The platform extracts or uses features of seed entities and ranks other addressable entities by similarity or predicted relevance.

seed: paid subscribers with 90-day value > €50
  → platform model/features
  → new similar prospects
  → prospecting delivery

The quality of the seed defines the model’s task: a mixed seed of random trials does not teach the same thing as stable high-value subscribers. There is usually a trade-off between similarity and size: a narrower threshold yields a smaller set, while a broader one provides more reach35 but weaker similarity. A lookalike audience does not have to include the seed, does not guarantee a conversion, and does not prove a causal effect.

Inclusion semantics also change by product mode. At the time of writing, Google is transitioning Lookalike segments in Demand Gen (one of the Google Ads campaign types), in phases, from a similarity-threshold constraint to suggestion mode: the seed and reach level guide the model, but delivery may go beyond the threshold; a separate opt-out path preserves the previous constraint semantics. This is not a rule for all campaign types or vendors.

Meta Lookalike retains a seed-based model, but in Advantage+ audience, an included custom/lookalike audience may be a suggestion rather than a boundary. Thus, the word Lookalike describes how the prospect signal is built, but not, by itself, the strictness of delivery.

Exclusions and Overlap: Negative Rules and Set Logic

Suppression is needed when “do not show” matters more than a positive preference. An acquisition campaign usually excludes current customers; a re-engagement campaign excludes users who have already performed the desired action; a population subject to regulation or incompatible with the product may have its own hard control. A negative predicate should have clear precedence over a positive signal:

eligible(entity) = positive_rules(entity)
                AND NOT excluded(entity)

But the actual completeness of an exclusion is limited by observability. If a subscription event arrives late, membership is stale, identity is not matched, or the platform interprets the field as a suggestion, the already-paying customer may still receive an impression. Therefore, an exclusion expresses an intention and is a contract rule, not a promise of zero leakage.

Audience overlap36 is the intersection of memberships in multiple segments within the same set of a platform’s addressable entities. Before launch, a subscription app might inspect:

A = trial_no_purchase
B = existing_paid
C = high_value_seed_lookalike

inspect: A ∩ B, A ∩ C, B ∩ C

A ∩ B often indicates stale events, inconsistent windows, or an error in the rules: the same account is simultaneously considered an unpaid trial and a paid subscriber. B ∩ C may be valid if the lookalike product includes the seed; for prospecting, it is still sensible to remove it through an exclusion if the goal is new customers.

An illustrative Jaccard-like ratio can be used to compare two sets:

overlap(A, B) = |A ∩ B| / |A ∪ B|

The numerator and denominator must refer to the same entity definition, the same set of addressable entities, and the same snapshot time. Platform estimates may differ due to matching and identity coverage. Overlap means that the sum of segment sizes overstates distinct reach; it fragments the setup and complicates segment-level interpretation, but by itself it does not prove duplicate billing or that campaigns “bid against themselves”—auction37 and deduplication semantics depend on the platform.

Broad and Algorithmic Delivery: The Boundary Between Targeting and Optimization

Broad targeting is a setup with few positive audience restrictions. This does not mean an absence of rules: geo, policy, technical compatibility, placements, brand-safety controls (Chapter 37), and exclusions remain. Algorithmic targeting38 uses the optimization goal, seeds, context, creative/landing signals, and outcome feedback to find likely useful opportunities; it goes beyond the selected audience only where the input is declared a suggestion rather than a hard control.

Setup Hard controls Positive audience input Can it go beyond the input? What the optimizer does
Manual / narrow Geo, device, context, and audience boundaries Selected segment as a constraint Usually not, if the contract is truly hard Ranks within a narrow eligible set
Broad Basic geo/policy/inventory guardrails and exclusions Few or no positive restrictions The set is broad from the outset Finds likely outcomes in a broad set
Algorithmic with signals Hard controls and exclusions Seed, interest, keyword, creative/landing signals Yes, if the input is a suggestion Expands or reallocates delivery toward the goal

Two vendor examples demonstrate the portable distinction control ≠ signal:

  • In Google Ads, optimized targeting for supported campaign types may seek conversions beyond manually selected inputs. The specific signals differ: for Display, the documentation names audience segments, keywords, and topics; for Video and Demand Gen, it names audience, custom, and customer-data segments, while selected Video placements remain boundaries. Customer-data exclusions and separate brand-safety controls have different semantics.
  • In Meta Advantage+ audience, Locations, minimum age, languages, and excluded custom audiences may serve as controls, while suggested age, gender, detailed targeting, and included custom audiences guide AI that may deliver to other users within those controls. Availability depends on the campaign and policy category.

This is not a performance comparison or a universal recommendation to enable automation. The conclusion is narrower: audience inputs that look the same may have different contracts.

Subscription App: End-to-End Flow

An advertiser wants new paid subscriptions in Germany. Configuration and data pass through the following stages:

  1. Hard eligibility: DE, mobile, supported OS, permitted placements, and policy rules. The geo option has been checked to determine whether it means presence or interest.
  2. Suppression: current paid subscribers are excluded from acquisition. A match delay is acknowledged as a possible source of leakage.
  3. Retargeting branch: the first-party rule trial_started ≥ 7d ago AND no paid_subscription since trial_started creates a separate audience for re-engagement.
  4. Prospecting branch: high-value paid subscribers form the seed; the lookalike is used as a constraint or hint according to the selected product mode.
  5. Broad/algorithmic branch: positive audience restrictions are minimal, but hard controls and exclusions remain. The model ranks eligible opportunities to optimize for the paid_subscription event.
  6. Feedback: impressions39, clicks40, installs, trials, and subscription events update reporting, segment membership, and model inputs. A significant targeting change at this step may send the optimizer back into the learning phase41 (Chapter 5).

In this flow, targeting establishes eligibility and preferences; optimization directs delivery toward the event. Bid calculation, budget allocation, and pacing are not needed to draw this boundary and are covered in the next chapter.

Six questions are enough when evaluating any targeting product:

  1. What entity belongs to the audience, and what constitutes an opportunity?
  2. Which criteria are hard, which are suggested, and which are negative?
  3. Who created the signal, and what are its provenance, window, and freshness?
  4. Can the model go beyond the selected audience or seed?
  5. What optimization event42 guides the ranking?
  6. What limits remain around unknown values, matching, permissions, and policy?

Common Confusions

  • Targeting ≠ optimization. The former sets constraints and preferences; the latter ranks eligible opportunities toward an outcome.
  • Audience ≠ context. An audience describes a matched entity or history; context is the current media environment.
  • Interest ≠ intent ≠ behavior. Affinity, a current query, and a past action are different kinds of evidence.
  • Placement ≠ contextual selection. A placement explicitly names “where”; a contextual system selects environments based on content signals.
  • First-party data ≠ first-party identity. Data from a direct relationship and an identifier the party itself controls are different things; the advertiser’s audience is activated only after a match with the identity of the platform where the ad is shown, and data provenance replaces neither permission nor that match.
  • Retargeting ≠ attribution. A delivery tactic does not determine who receives conversion credit.
  • Lookalike ≠ a copy of the seed. It is a modeled prospect set; inclusion and expansion semantics are product-specific.
  • Overlap ≠ double charge. It is a set intersection; billing consequences require separate contract evidence.

What to Remember

  1. Targeting narrows or guides the set of opportunities, but does not promise an outcome; optimization begins after eligibility and uses outcomes as feedback.
  2. Every setting should be classified as a hard control, signal/hint, or exclusion; algorithmic expansion goes beyond the input only where the input is declared a suggestion.
  3. Geo, demographics, interests, and behavior are often inferred and may be unknown or stale; device, OS, and browser are different axes with different availability.
  4. A segment is defined not by its label, but by its entity, membership rule, provenance, observation window, and activation contract.
  5. First-party and third-party describe the origin of data, not quality; activation of any segment is limited by the platform’s addressability.
  6. Retargeting again targets entities in a known interaction set; a lookalike finds new prospects from a seed.

Check Yourself

  1. Why might geo = DE mean something other than “the user is physically in Germany”?
  2. How do interest: hiking, viewed_boots in 14d, and an article about hiking differ as targeting inputs?
  3. What five characteristics are needed to meaningfully describe an audience segment?
  4. Why is an included audience in an algorithmic product not always a strict delivery boundary?

Sources and Further Reading

  1. IAB Tech Lab — Audience Taxonomy — a common nomenclature for audience categories; updated 11 December 2024, accessed 2 September 2026.
  2. IAB Tech Lab — Data Transparency Standard — disclosure of provenance, recency, segmentation, and modeling without a quality grade; updated 23 April 2024, accessed 2 September 2026.
  3. Google Ads Help — About advanced location options — Presence or Interest, Presence, and best-effort geo inference; accessed 2 September 2026.
  4. Google Ads API — Targeting criteria — product-specific criteria and scopes; accessed 2 September 2026.
  5. Google Ads Help — About Customer Match — first-party customer-data matching and current availability limits; accessed 2 September 2026.
  6. Google Ads Help — About optimized targeting — manually selected segments as signals and separate controls; accessed 2 September 2026.
  7. Google Ads Help — Use Lookalike segments to grow your audience — phased Demand Gen transition to suggestion mode during 2026; accessed 2 September 2026.
  8. Meta Business Help Center — Audience controls and Audience suggestions in Advantage+ audience — product-specific distinction between controls and suggestions; accessed 2 September 2026.

Footnotes

  1. Targeting — the campaign’s rules about to whom and where an ad may be shown: hard controls (eligibility rules — geo, device, context, audience/keyword criteria, exclusions) that restrict the set of permitted opportunities, and signals/hints that in algorithmic products only steer delivery without narrowing the eligible set. Targeting answers the question “for which opportunities/users/contexts is showing possible,” as opposed to placement (“where exactly”) and optimization (“which of the permitted impressions to choose”); even if the API stores them in one object, these are different axes. ↩

  2. Ad opportunity — a specific opportunity to select and show an ad that arises when a user reaches a placement. A short-lived event: in a digital environment it can vanish faster than a person can make a decision, so systems process it in milliseconds, and in programmatic it is put up for bidding as a separate unit. ↩

  3. Advertiser — a market participant that funds advertising for a business result: installs, subscription purchases, growth in awareness. It buys not a person but the opportunity to address an audience in a specific context; it creates demand and sets goals, budget, and creative. ↩

  4. Publisher — the owner or operator of a media environment where advertising opportunities arise: a website, a mobile app, a video service, a game. It monetizes the environment with advertising, creates supply, and decides which placements, formats, and buyers to admit, balancing revenue against user experience. ↩

  5. Seed audience — a set of known entities (for example, high-value paying subscribers) that the advertiser hands to a platform as a sample for building a lookalike or as a signal for algorithmic delivery. The quality and size of the seed determine the quality of the model: too small a seed yields noise, too broad a seed dilutes the features. The seed itself is not the target audience for delivery; many platforms exclude it from the lookalike by default. ↩

  6. Audience exclusion / suppression — a negative targeting rule designed to exclude entities with a named membership or status; the actual completeness of the exclusion is limited by identity match, freshness, propagation latency, and the semantics of the specific platform. ↩

  7. Optimization target — the outcome toward which a bidding system tunes its future decisions: most often an optimization event (install, paid subscription) plus a goal for it. Not to be confused with a strategy parameter: target CPA $4 is a bid-strategy constraint, not the event the system predicts. It does not turn the contract into CPA billing: the payment model establishes the contract, the realized metrics describe what happened post-fact, and the optimization target influences only decisions before the event. ↩

  8. Campaign — an organized set of an advertiser’s advertising activity with a goal, a budget, delivery rules, and creatives. Before delivery, its conditions are communicated to the buying systems; objects and events are linked by identifiers such as campaign_id. ↩

  9. Ad group / ad set — the intermediate level of the campaign hierarchy between the campaign and the ad, grouping ads with shared execution rules. The names and the set of fields are vendor-specific, and there is no universal schema: in the Google-like model, the ad group groups closely related ads and triggering criteria (for example, related keywords); in the Meta-like model, the ad set holds the audience, placements, optimization, bid, budget, and schedule. ↩

  10. Data provider — a provider that sells or supplies datasets, attributes, or ready-made audience segments for targeting, measurement, and enrichment. It works in parallel to the buying chain, not “between” two mandatory neighbors; third-party segments are usually paid for by the buyer — often as a data CPM charged through the buying platform on top of the media cost. Not to be confused with an identity provider, which does not supply data but establishes correspondence between identifiers. ↩

  11. Audience segment — an addressable set of entities whose membership is defined by a named rule, data source/provenance, observation window, and platform matching/activation contract; an entity can be a user, account, device, browser, household, or modeled profile, so a segment is not necessarily a list of known people. ↩

  12. DSP (Demand-Side Platform) — a platform that automates the purchase of advertising opportunities for an advertiser: it aggregates access to several supply sources, evaluates each opportunity, and decides whether to bid and how much to pay. It belongs to the demand side not because it “creates demand” but because it acts in the buyer’s interest. ↩

  13. Advertising inventory — the set of available or forecast opportunities to show advertising, bounded by the environment, the slot, the format, the time, the publisher’s rules, and the permissible audience context. Inventory is not a list of users and not a warehouse of pre-produced impressions: an opportunity arises and disappears at the moment the product is used. ↩

  14. Media spend — the money an advertiser allocates and spends on buying advertising. It passes through the media sellers and ultimately forms publisher revenue, shrinking along the way by fees; it is not a synonym for publisher revenue, nor for intermediary fees — these are different amounts in one chain. ↩

  15. Bid — a buyer’s offer to pay a certain price for a specific ad opportunity, usually expressed as a CPM. A bid is a pre-event decision: it is neither the final price (clearing price) nor the billable amount; in programmatic, the DSP forms a bid for each bid request separately. ↩

  16. Budget — a constraint on the available spend over a given scope (a campaign, an ad set, a shared pool of several campaigns) and time basis. Common are the daily budget (a hard calendar cap, an average daily target, or another form — the product defines this) and the lifetime/flight budget for the whole interval. A budget constrains spending but does not guarantee inventory, conversions, even pacing, or an exact invoice amount. ↩

  17. Pacing — the policy of distributing spend over time within budget constraints: whether to spend evenly, accelerate on cheap or high-quality opportunities, or hold back the budget. Pacing allocates the resource but does not promise even spending: the available opportunities and the allocation change. ↩

  18. Placement (ad slot or selectable media scope) — an overloaded product term. On the publisher side, it is a predefined ad slot or rule for ad appearance that generates ad opportunities as the product is used. On the buy side, placement can mean an explicitly selected delivery environment — a site, app, channel, video, section, or individual slot. The entity level must be clarified from the product contract. ↩

  19. Contextual targeting — selecting ad opportunities based on the current content or media context: topic, semantics, page/app/video category, and related signals. It describes the delivery environment and does not require asserting that a persistent cross-site identity is known for the user. ↩

  20. Observation window — the period over which an event counts toward forming an audience segment: “trial started in the last 7 days,” “visited the pricing page within 30 days.” It determines the freshness of membership and the size of the segment. Not to be confused with the attribution/lookback window (how far before a conversion a touchpoint counts) or the conversion window (how long to wait for a conversion after a click): these are windows of different mechanisms. ↩

  21. Creative — the advertising material that is delivered and shown to the user: a banner, a video, a native format, a playable. Along with targeting rules it is an object of delivery and must match the format and rules of the publisher’s environment; for a campaign, several creatives are usually prepared. ↩

  22. Addressable — a user, device, or household that a platform can recognize by an available identifier and deliver a specific impression to; addressability is the share of inventory or audience for which this is possible. It depends on the environment (browser, iOS, Android, CTV), consent, and available identifiers, so the same segment can be activatable on one platform and empty on another. ↩

  23. Activation — handing data or an audience segment to a system that can use it for targeting, suppression, personalization, or measurement: from a CDP/DMP to a DSP, from the advertiser’s CRM to a walled garden, from an identity graph to an SSP. Before activation, data is a record in storage; after it, a rule that affects real impressions. Activation requires matching identifiers between the source and the receiving platform and a permitted legal basis. ↩

  24. Attribution — the rules by which the available advertising touchpoints (impression, click) are matched to outcomes (install, registration, purchase) and receive credit for the conversion. Attribution answers the question “which source was credited with the action under the chosen rules,” not “whether the ad was the true cause”: credit does not prove causality. Attribution providers are a separate cross-cutting category; in mobile, the MMP. ↩

  25. Conversion — a valuable action or outcome that the advertiser has defined as significant (install, trial, first payment, renewal, purchase) and that a system has counted under its measurement and attribution rules. Conversions cannot be summed without an explicit counting rule: one click can lead to several conversions. An observed conversion is a recorded event; an attributed conversion is an event to which an attribution rule assigned credit; neither proves causality. ↩

  26. First-party audience — an audience segment built from data collected by an advertiser or publisher in a direct relationship: CRM/customer records, site/app events, subscriptions, and purchases. First-party describes provenance but does not guarantee consent, legal basis, ownership, accuracy, or completeness of the platform match. ↩

  27. CDP (Customer Data Platform) — a system that collects first-party events and customer data from the product, CRM, and other systems into a persistent unified customer record and makes profiles, events, and segments available to downstream systems. It does not buy media and does not assign attribution credit; it belongs to customer lifecycle management, which makes it part of MarTech. ↩

  28. Third-party audience — an audience segment supplied by an external data provider independently of an advertiser’s direct relationship with the entity; evaluating the segment requires provenance, collection method, recency, membership criteria, modeling metadata, permissions, and activation coverage, because the label alone does not prove its composition or accuracy. ↩

  29. DMP (Data Management Platform) — an advertising-oriented system for managing, segmenting, and activating audience data: it collects campaign signals and external datasets, builds audience profiles and segments, and activates them through advertising integrations. It does not have to be the system of record for the full customer lifecycle. The historical rule “a DMP is only cookies and anonymous data” is unreliable; the working boundary with the CDP runs along the primary task — advertising audience activation versus a durable customer record. ↩

  30. Measurement — the function of recording and aggregating advertising events (impressions, clicks, outcomes) for reporting and optimization. Measurement providers give an independent point of observation; different participants have different timestamps and counting rules, so reports do not have to match perfectly. In mobile, the dedicated category is the MMP. ↩

  31. Retargeting — targeting entities that previously performed a named interaction or event within a specified observation window in order to address them again; it is a delivery tactic, not an attribution rule or proof of causal effect. ↩

  32. CVR (Conversion Rate) — the share of qualifying conversions among the declared eligible precursor events: CVR = qualifying conversions / declared eligible precursor events × 100%. A bare CVR hides the transition: only named transitions such as click-to-install or install-to-trial are comparable; a value above 100% is possible with several conversion actions on one interaction. ↩

  33. Lookalike audience — a modeled audience of new prospects found by similarity or predicted relevance to a seed audience; the result depends on seed quality and the platform model, and the use of a lookalike in delivery can be a hard constraint or a suggestion according to the product contract. ↩

  34. Prospecting — directing a campaign at users who have not yet interacted with the advertiser: lookalike, interest, contextual, or broad targeting. The opposite of retargeting and re-engagement, which address already-known users. Prospecting typically shows lower conversion rates and a higher value of the incremental result, since the user would not have come on their own. ↩

  35. Reach — the de-duplicated number of unique entities (persons, devices, households, accounts, or a modeled audience) that received at least one qualifying exposure in a period, or the share of such entities in an explicitly defined population: reach rate = reached unique entities / declared universe × 100%. Comparable only with identical entity definition, exposure rule, population, de-duplication scope, and period. ↩

  36. Audience overlap — the intersection of memberships in multiple audience segments within a single activation universe and at the same snapshot; it affects distinct reach, reporting, and segmentation design, but does not by itself prove duplicate billing or self-competition. ↩

  37. Auction — the mechanism for choosing a winner among several bids for one ad opportunity and for determining the price (clearing price) under the auction’s rules — first-price, second-price, and their variants. An auction win is the event of choosing a winner; it is not an impression and does not by itself make the impression billable. ↩

  38. Algorithmic targeting — a delivery mode in which a model uses an optimization goal, context, audience seeds, and other signals to find and rank opportunities, potentially beyond suggested segments but within the stated hard controls and exclusions. Broad targeting is a setup with minimal positive audience restrictions, not an absence of constraints. ↩

  39. Impression — a delivery or display event for the selected ad, recorded by a system under a specific rule; a counting unit, not a fact of human perception. Served, rendered, and viewable are different delivery states, so the counters of different participants in the same deal may not match. ↩

  40. Click — a qualifying user-initiated interaction with an ad, recorded under a specific system’s rule. Even a single action has several measurement points: initiated, measured (click tracker), received (destination), and resolved (landing page) clicks need not match due to dropped connections, redirect chains, and filters. ↩

  41. Learning phase (learning status) — the state of the optimizer in which the delivery/bidding model gathers evidence after launch or a material configuration change: earlier data describe the new decisions less accurately. Learning is one of the axes of the operational state alongside the configured, effective, and review statuses, not the opposite of active: impressions and spend continue, and vendor-specific significant edits (targeting, creative, optimization event, the set of ads, the bid strategy) can return the optimizer to learning without changing the configured status. ↩

  42. Optimization event — the observable event that the delivery system learns from and optimizes impressions toward: install, trial start, paid subscription, purchase. The advertiser selects it in the campaign settings (different platforms call it a conversion action, conversion goal, or performance goal). It is not the campaign objective (the business intent), not the billable event (what is paid for), and not a strategy parameter such as target CPA. A deeper event yields fewer signals and a longer learning phase. ↩